LLMTR Blog

LLMTR Blog

Guides on Turkey-focused LLM APIs, OpenAI-compatible gateways, model comparison, pricing, and data policy.

All posts

RAG and data · 2026-06-02

How to measure AI search visibility for LLMTR

A practical guide to measuring AI search visibility for technical products through citations, referrals, source quality, freshness, and structured content surfaces.

Trust and compliance · 2026-08-13

BİGR-compliant AI architecture: matching model choice to data classification in Turkish public institutions

A public institution does not make one AI decision; it makes a separate architectural decision for every class of data it processes. This article applies the asset group and criticality rating logic of Turkey's Information and Communication Security Guide to large language model integrations and offers a workable mapping between data confidentiality level and model architecture.

Agent and MCP guides · 2026-08-13

Municipal AI call centers in Turkey: tender analysis and solution-line design

In municipal call center and citizen solution-line projects, artificial intelligence is not a single piece of software but four distinct jobs with different model requirements. This article covers task decomposition, human handover thresholds, KVKK boundaries, how a lowest-price tender changes bid arithmetic, and the metrics worth writing into a technical specification.

Pricing and benchmark · 2026-08-13

Türkiye's AI Action Plan 2026-2030 and Public Budgets: Using the 2 Percent Share

The Türkiye Artificial Intelligence Action Plan 2026-2030 creates both an obligation and a funding channel for public agencies. This article covers the practical budget meaning of the 2 percent commitment, how to assemble an internal funding case, and why choosing a portable architecture is a risk management decision while regulation is still pending.

Integration guides · 2026-08-28

muse-spark-1.2-contributor: LLMTR model ID and endpoint

Calling Contributor through LLMTR requires the model ID, API base URL and credentials to match. This guide starts with the data warning, then builds a first text request and checks the relevant API contract.

Model comparison · 2026-09-30

Ling Tiny alternatives: free Flash Fin versus paid Flash

Flash Fin was a free alternative after Ling Tiny; its free offer ended on 30 September 2026 and the row was retired. Paid Flash remains the successor named in LLMTR's retirement policy for both Tiny and Fin.

Integration guides · 2026-08-29

Adding the LLMTR gateway entry to OpenClaude and how discovery works

OpenClaude defines LLMTR as its own gateway record with a dedicated base URL, a dedicated credential variable and a hybrid catalog fed by seed entries. This guide walks that record field by field and explains how the model list fills up.

Integration guides · 2026-08-29

Cline with LLMTR: provider setup inside the editor panel

Cline picks a provider from a settings panel inside the editor and stores the chosen model separately for Plan and Act. This guide walks the panel flow, the request the handler builds and the SDK registration.

Agent workflows · 2026-08-29

PentestCode LLMTR provider setup for authorised engagements

Working from the actual files on the PentestCode LLMTR branch, this guide covers key resolution, the seeded and live model catalog, and the request transform that normalises reasoning fields per model, all inside an authorised engagement scope.

Agent workflows · 2026-08-29

Strix LLMTR setup: model prefix, dedupe and the reporting chain

Built from the real files on two separate Strix branches, this guide walks through model prefix resolution, why the dedupe model needs its own endpoint, and which Turkey-hosted rows actually support the tool calls an agent loop needs.

Agent workflows · 2026-08-29

HarnessRouter LLMTR provider: one key, three harnesses

Connecting the LLMTR provider in HarnessRouter adds no new name to any model picker; it makes models already listed reachable through Türkiye. Here is the intersection logic and the wire shape each harness uses.

Integration guides · 2026-08-29

Hermes Agent LLMTR plugin: profile fields and install

Hermes Agent keeps third-party providers out of the core tree and in a plugin directory. The LLMTR profile is a declarative ProviderProfile with no custom transport hooks; this guide covers its fields and its lifecycle.

Integration guides · 2026-08-29

DeepSeek Harness LLMTR plugin: the LLM seam and catalogue

In DeepSeek Harness a provider integration is not written into the core; it arrives as a plugin attached to the ctx.llm seam. The model list is not bundled either, it is read from the endpoint.

Integration guides · 2026-08-29

LLMTR provider for the Pi coding agent and its generated catalog

Pi registers a provider through a small factory function and leaves the model list to a generation script rather than maintaining it by hand. This guide explains how those two files relate and what the request body contains.

Agent workflows · 2026-08-29

Setting up the LLMTR provider in OpenCompany workflows

In a workflow engine a model node builds its client from the selected credential and then resolves the model. Because LLMTR is a gateway, that resolution reads two separate sources: wire shape from the vendor, thinking from the gateway.

Integration guides · 2026-08-29

Connecting OpenStory to the LLMTR gateway with a team key

OpenStory is a multi-team production platform, so the key is not a single environment variable. An LLMTR key can be entered per team, and it outranks the other providers whenever the gateway carries the requested model.

Integration guides · 2026-08-29

Pointing the OpenReview code review bot at LLMTR

OpenReview explores the repository on every review, so token spend follows the model you pick. In this branch swapping the gateway is not a code change but a matter of three environment variables.

Integration guides · 2026-08-29

Running OpenBot on LLMTR: one base URL, two agent runtimes

OpenBot is a self-hosted system that can give every agent a computer of its own. The integration is a matter of environment values rather than code, and a single base URL redirects three separate readers at once.

Integration guides · 2026-08-29

Running the Open SEO SAM agent through an LLMTR gateway

Open SEO is an open source SEO analysis platform whose agent features are optional. This guide adds a second path without removing the default one, and shows the precedence rule and where spend visibility moves to.

Model comparison · 2026-08-29

The models.dev LLMTR provider entry and metadata sync

models.dev is not a service but an open metadata database holding model ids and capabilities in TOML files. We compare the current LLMTR entry against the live catalog to see which rows are missing and which one is retired.

GPT-6 Astra guides · 2026-09-05

What Is OpenAI GPT-6 Astra? Features and Access

Explore GPT-6 Astra, OpenAI’s new model for complex work: its context window, reasoning, access rollout and practical considerations for API developers.

GPT-6 Astra guides · 2026-09-05

How to Use GPT-6 Astra API Through LLMTR

Send your first GPT-6 Astra request through LLMTR with Responses API, choose a supported reasoning level and inspect output usage before scaling up.

Model comparison · 2026-10-03

What is Motif 3? The free agentic coding model, measured

What Motif 3 is for, the capabilities measured through LLMTR, the vision claim its own published metadata contradicts, and how it compares with the other free models sharing the same context window.

Integration guides · 2026-09-19

What is EVREN? How to use SAYZEK's LLM API

Explains the EVREN LLM Inference Service with sources: who runs it, that it is free until 1 November, its eleven-model list, its OpenAI compatibility and the setup path from e-Devlet login to a first chat request.

Pricing and benchmark · 2026-09-19

EVREN quota and credit system: free, but metered

Explains, with sources, the meters that run on EVREN even during the free period: a sliding-window token cap, a credit balance, per-request charge state, the cooldown on consecutive requests and how auto mode affects consumption.

Agent workflows · 2026-09-20

Moving from an EVREN row to a global model: one field changes

Moving from an EVREN row to a global model needs no code change: the endpoint, the header and the body schema stay the same and only the model id differs. How an agent verifies capability before calling is read from the same place.

Privacy and compliance · 2026-09-25

Reading the data flow between EVREN and LLMTR

Follows an EVREN request step by step from the application to LLMTR, on to EVREN and back, and tabulates what is seen and what is kept at each stop.

EVREN guides · 2026-09-25

EVREN long documents, context, and output budgets

Explains what goes into the context window, how the output budget is set, the difference between sending a document in one request and splitting it, and how long inputs affect the EVREN quota.

EVREN guides · 2026-09-25

EVREN response waiting and streaming experience

Explains the difference between real streaming and chunked delivery after completion, the factors behind waiting time on EVREN rows, and timeout and retry behaviour.

EVREN guides · 2026-09-25

EVREN API errors: choosing the right remedy

Explains the error types on EVREN rows, the meaning of the three separate 403 errors, the upstreamStatus field, and the cases that return no error but still give an unexpected result.

Agent workflows · 2026-09-25

Choosing Noul, Choice, and Score with Jev

Explains how yes/no, unordered-option and ordered-level decisions map to Jev, the criteria format of each type, and how the three types are used together in one request.

Agent workflows · 2026-09-25

Multiple independent questions over one document with Jev

Calculates the cost difference between many questions in one request and a request per question, and explains the limits of independent questions, the role of question keys and the request limits.

Agent workflows · 2026-09-25

Explaining a Jev decision with a chat model

Explains a two-step flow that takes the decision from Jev and the user-facing text from a chat model, including the context sent to the chat model, data policy, and the balance of cost and latency.

Model comparisons · 2026-09-25

How to evaluate the Ember-1 research preview

Explains what Ember-1 promises, what the preview period means, which capabilities do not carry over to its successor moonshot/kimi-k3, and how to compare the two models on the same tasks.

Model comparisons · 2026-09-25

Ember-1 reasoning tokens, quality, and cost

Explains the on/off settings for reasoning, why reasoning_effort levels are not supported, how reasoning tokens affect cost, and how reasoning content is read in a stream.

Model comparisons · 2026-09-25

Tool calling and JSON Schema with Ember-1

Covers auto, required and named tool_choice behaviour, the difference between json_object and strict json_schema, what a schema does and does not guarantee, and where schema validation moves when switching to Kimi K3.

Model comparisons · 2026-09-25

Screenshot and error analysis with Ember-1

Explains the image input formats, which text to send alongside a screenshot, why the output is read as a list of hypotheses, and how the image format changes when moving to moonshot/kimi-k3.

Model comparisons · 2026-09-25

Ember-1 context, cache, and usage cost

Explains Ember-1's price items, a cache-friendly prompt layout, the example cost of a request with and without cache, and how cost changes when moving to moonshot/kimi-k3.

Model comparison · 2026-09-22

Turkish embedding models compared: choosing for RAG and search

Comparing chat models and comparing embedding models answer different questions. This guide covers embedding selection for a Turkish RAG pipeline through dimension, multilinguality, and measurement on your own dataset.

Public sector and trust · 2026-09-22

AI competency and training guide for public-sector staff

Building the right AI architecture does not stop staff from continuing to do the same work with personal accounts. This guide covers role-based training levels and the risk of shadow usage.

Integration guides · 2026-09-22

Embeddings and rerank on the OpenAI-compatible gateway

A gateway migration is usually framed around chat endpoints alone, but embeddings and rerank endpoints can be called with the same identity and base URL. This guide shows when to use each of the three together.

Model comparison · 2026-09-22

Multi-page document analysis with DeepSeek V4 Flash Vision

The DeepSeek V4 Flash Vision guide covers the basics of sending a single image. This article covers the ordering and context issues that come up when processing a multi-page document (a report, a contract, a scanned form) in one request.

Agent workflows · 2026-09-22

MCP security: permission management for tool access

The MCP explainer article covers the protocol's general framework. This article covers the decision of which tools and data get exposed, and in what scope, to an agent connected to an MCP server.

Model comparison · 2026-09-22

Native LLM models: fine-tuning or prompt engineering?

The native LLM strategy article covers which model fits which task. This article covers how to further adapt the chosen model to a specific task, choosing between fine-tuning and prompt engineering.

Agent workflows · 2026-09-22

Error handling and fallback strategy in AI agents

The AI agent guide covers architecture and tool use. This article covers a part of that architecture: error handling, and what an agent should do when a tool fails.

Trust and security · 2026-09-22

Gemini API key rotation and least-privilege practice

The Google API key stolen article focuses on post-incident response. This article covers the rotation and privilege-scoping practice you set up beforehand so the same incident does less damage next time.

RAG and data · 2026-09-22

Ranking sources in a RAG pipeline with Muse Spark 1.2

The long-context budget article covers how to divide space between input and output. This article covers why ordering matters once you fit multiple source document chunks into that input space.

Integration guides · 2026-09-22

Two-speaker dialogue with Gemini TTS

The long-text narration article covers splitting long content into chunks with a single voice. This article covers the different setup needed when narrating a two-speaker dialogue.

Pricing and budget · 2026-10-06

Gemini TTS: Flash or Pro? Model choice and cost comparison

The Turkish voice generation article covers a single Turkish narration flow with the Pro Preview version. This article covers which criteria to use when choosing between the Flash and Pro versions of the same family.

Gateway basics · 2026-09-22

Avoiding vendor lock-in in public-sector AI architecture

The public-sector AI API architecture article covers data residency and provider independence decisions. This article covers how to preserve that independence at the code level: avoiding vendor lock-in in practice.

Integration guides · 2026-09-30

Financial analysis with reasoning and forced tool calls on Ling 3.0 Flash Fin

Ling 3.0 Flash Fin was a model tuned for multi-step investment research; its free offer ended on 30 September 2026 and the row was retired. This article covers when to keep thinking on, when to turn it off, and when to force a tool call; the same flow works on its metered successor, Ling 3.0 Flash.

Integration guides · 2026-09-22

How reasoning_effort levels change roleplay output on Aion 2.0

Aion 2.0 supports four separate reasoning_effort levels for roleplay and storytelling. This article covers the trade-off between fast character replies and more consistent, long-arc scene-building through those levels.

RAG and data · 2026-10-21

A long-document research flow with 1M context on MiMo V2.5 Pro

MiMo V2.5 Pro's 1M context window carries long documents in a single request. Because the model-side web_search tool is not available on LLMTR right now, this article covers a flow that collects current information on your side and adds it to the context. With the model retired, the flow is built on its successor MiMo V2.6 Pro today.

Model comparison · 2026-09-22

When Recraft V3 Vector still makes sense: a comparison with V4

Recraft V3 Vector and V4 Vector carry the same per-image price; the real difference is in the prompt length limit. This article clarifies that difference and covers the one concrete scenario where V3 still makes sense.

Pricing and budget · 2026-09-22

Low-cost, high-volume API usage with Qwen-Flash

Qwen-Flash is a lightweight model designed for low latency and high volume. This article covers its context-length pricing tiers and how cache reads affect cost.

Trust and security · 2026-09-22

qwen3.6-27b-free retired: managing the daily quota on qwen3.8-27b-free

For anyone looking for this row: the old identifier now returns a 410 model_retired. The successor qwen3.8-27b-free is also free and carries the same core capabilities (tool calling, JSON mode, image input); this article covers the free tier's daily quota behavior.

Model comparison · 2026-09-22

Complex reasoning and video analysis tasks with Qwen3.8-Max

Qwen3.8-Max stands out in the catalog with a 'Latest' label. This article covers how the three features that suit it to complex, multi-step tasks — 1M context, video input, thinking by default — work together.

Pricing and budget · 2026-09-22

When GLM-4.6V-FlashX fits a budget-conscious vision workflow

GLM-4.6V-FlashX is listed in the catalog with a 'Budget' label and offers basic image/document understanding. This article covers which tasks that level is sufficient for, and when to move to a more capable vision model.

RAG and data · 2026-10-06

Calculating two-stage retrieval cost with Voyage Rerank 2.5

Voyage Rerank 2.5 evaluates a query and a candidate document list together. This article covers how rerank cost is calculated in a two-stage retrieval pipeline, and how that cost translates into savings on the language-model side.

Pricing and budget · 2026-09-22

When GPT-5.6 Sol Pro is worth it over Terra Pro

The Terra Pro article covered this structure (that it isn't a separate OpenAI model). This article covers how the same mechanism works on Sol Pro, and when Sol Pro's noticeably higher price over Terra Pro is justified.

Trust and security · 2026-09-22

Sincap has been retired: where to move your free-trial flow

The Sincap model page now carries a 'Sağlayıcıda kapalı' (closed at provider) status label and is marked retired in the catalog. This article covers what a prototype or trial flow depending on this identifier should do.

Model comparison · 2026-09-22

Apertus v1.5 8B Thinking: reasoning or tool calling — not both

In the Apertus v1.5 8B family, reasoning and tool calling don't combine on the same row: the Thinking version reasons but doesn't call tools, and the standard version calls tools but doesn't return a separate reasoning trace. This article covers that binary choice.

Model comparisons · 2026-09-28

MiniMax M3.1 vs M3: which one to use and when

M3.1 is MiniMax's preview release after M3 and is free on LLMTR until 30 September 2026, 23:59 (TRT). M3 is metered and offers video input, prompt caching and switchable thinking.

Pricing and budget · 2026-10-06

API Model Pricing: Input and Output per Million Tokens

Compare LLMTR catalog prices in USD per 1M input and output tokens, convert pricing strings from /v1/models, estimate request costs, review special billing units, and make an authenticated first API call.

Integration guides · 2026-10-06

xAI API Images: aspect_ratio, Grok Imagine and b64_json

Use POST /v1/images/generations with xai/grok-imagine-image-quality, pass aspect_ratio and resolution, choose b64_json, decode data[0].b64_json, and save the image. Quality costs $0.05 per image; the standard model costs $0.02.

Model comparison · 2026-10-06

Muse Spark 1.2 Contributor vs Muse Spark 1.2

Meta/muse-spark-1.2 and its Contributor tier share the same 1.2 checkpoint, capabilities and 1,048,576-token context. Compare 2026-10-06 prices, training-data use, rate limits, reasoning billing and monthly costs.

Integration guides · 2026-10-06

How to Use Muse Spark 1.2 Contributor via API

Connect to meta/muse-spark-1.2-contributor through the LLMTR OpenAI-compatible API. The guide covers curl and Python requests, reasoning_effort, max_tokens, supported inputs, Contributor data use, pricing, rate-limit handling, and account setup.

Model comparison · 2026-10-06

Gemini Flash or Pro: Which Model Fits Your Workload?

Gemini Flash is the economical default for high-volume and tool-using work; Gemini Pro targets harder reasoning. Compare prices, thinking levels, retirement dates, and a 20-50-prompt routing test.

RAG and data · 2026-10-06

Voyage Models With Quantization and Flexible Dimensions

Seven Voyage models on LLMTR combine 256, 512, 1024 or 2048 output dimensions with quantization. Compare catalog prices, storage needs and fixed-dimension alternatives, then configure embeddings through the OpenAI-compatible API.

RAG and data · 2026-10-06

Gemini 2.5 Flash: Embedding Model or Chat Model?

Gemini 2.5 Flash is a chat model for text output; Gemini embedding models create vectors. Compare endpoints, inputs, token limits and prices, then place each model in the correct RAG step. Also review Gemini 2.5 Flash's 2026-10-16 retirement and LLMTR setup.

RAG and data · 2026-10-06

400 Bad Request: embeddinggemma Does Not Support Chat

EmbeddingGemma returns numeric vectors rather than generated text. Route Ollama requests to POST /api/embed and LLMTR requests to POST https://llmtr.com/v1/embeddings with llmtr/embeddingmagibu-200m, the successor of the retired llmtr/embeddinggemma-300m. On LLMTR, a chat endpoint returns HTTP 400 unsupported_operation for this model.

Model comparison · 2026-10-06

Gemini 3.1 Pro vs 3.7 Flash for Coding Guide

Use Gemini 3.7 Flash for coding edits, tests, refactors and agent loops, then route tasks that fail your test set to Gemini 3.1 Pro Preview. Compare context, output limits, supported tools and session costs through one OpenAI-compatible API.

Model comparison · 2026-10-06

GPT-5.6 Luna vs Luna Pro: API and pricing

GPT-5.6 Luna and Luna Pro use the same OpenAI model and per-token prices. Luna Pro defaults to pro reasoning, while Luna accepts it per request. Compare output_tokens on the same prompts, then route selected quality-boost requests and start with one API key.

Integration guides · 2026-10-06

Multi-LLM Gateway Without Data Training Risks

LLMTR provides one OpenAI-compatible API while model pages document data policies for hosted, EU, Meta, Contributor, and free rows. Read each model id, scope API keys to approved models, and keep confidential, personal, or customer data away from training rows.

Integration guides · 2026-10-06

Secure LLM APIs for Integration Without Data Sharing

LLMTR does not store prompt or response bodies. Use one key per environment, apply spending, expiry, model and IP restrictions, and select TR or EU model rows by processing location.

Integration guides · 2026-10-06

Grok Image Editing: What Is the Image ID Format?

Grok image editing has no separate image ID. On LLMTR, POST /v1/images/edits accepts URL inputs or base64 data:image/ values, with image for one image and images for 1-5 references. Model prices are $0.02 and $0.05 per image.

Integration guides · 2026-10-06

OpenAI gpt-image-2 Transparent Background 400 Error

OpenAI opened transparent backgrounds to gpt-image-2 as a preview on 20 August 2026; background transparent works with output_format png or webp. The same request returned 400 before that date. On LLMTR, transparent with jpeg returns 400 invalid_request.

Model comparison · 2026-10-19

TensorX API: EU-Hosted Models Through One API

LLMTR offers 13 TensorX chat models and 1 embedding model as tensorx/<model>-tensorx ids at https://llmtr.com/v1. A call needs the TensorX data-transfer consent on the account; the same key and balance also work for every other model.

Model comparison · 2026-10-10

Can You Use Kimi K3 Through an EU-Hosted API?

Kimi K3 is available on LLMTR as tensorx/kimi-k3-tensorx. Per the provider's statement, requests are processed inside the European Union, and the TensorX data-transfer consent is required. Context is 1,048,576 tokens; price is $3.00 / $0.75 / $15.00 per 1M tokens.

Trust and compliance · 2026-10-31

DeepSeek GDPR: DeepSeek V4 Through an EU-Hosted API

DeepSeek's own privacy policy says it stores personal data in the People's Republic of China. The four TensorX DeepSeek V4 rows on LLMTR are processed inside the European Union per the provider; the GDPR assessment stays with you, and a data-transfer consent is required.

Agent workflows · 2026-10-19

EU-Hosted Models for AI Agents and Coding

Every TensorX chat model on LLMTR supports tool (function) calling and streaming. Nine rows suited to agents and coding are compared, with a session cost example, reasoning token notes, API key scoping and the TensorX data-transfer consent.

RAG and data · 2026-10-10

EU-Hosted Embedding API for a GDPR-Minded RAG

A RAG pipeline can run its embedding and answer steps on TensorX rows through LLMTR. Embedding 10 million tokens costs $0.20, and one data-transfer consent covers both steps. The GDPR assessment stays with you.

Gateway alternatives · 2026-10-10

Should I Use TensorX Directly or Through LLMTR?

On LLMTR the TensorX rows carry TensorX's USD list price; an 8% margin is added once, on the credit top-up. In return you get one key, one balance and the option to switch to a same-family row by changing the model string. If you will only ever use TensorX, going direct can also make sense.

Integration guides · 2026-10-10

EU-Hosted Model Error: A Two-Way Fallback Guide

If an EU-hosted row returns an error, the same API key and balance can call the other row of the same model family; GreenPT and TensorX pairs work in both directions. LLMTR does not switch providers; the client does. No fallback on 400 or 403.

Pricing and budget · 2026-10-19

TensorX Pricing per Million Tokens on LLMTR

LLMTR serves TensorX models at TensorX's own USD list prices, with no markup on the token price. The table lists input, cache-read and output prices for 11 chat rows and 1 embedding row, followed by worked cost examples.

Integration guides · 2026-10-10

Free Claude Haiku 5.5 API on LLMTR until 14 October 2026

On LLMTR, anthropic/claude-haiku-5.5-free costs $0.00 per 1M tokens until 14 October 2026, 23:59 (TRT). Use is limited per account per day; prompts go up to 100,000 tokens and responses up to 32,768. After the end date, requests return HTTP 410 model_retired.

Agent workflows · 2026-10-10

Claude Code gateway: ANTHROPIC_DEFAULT_HAIKU_MODEL for free Haiku 5.5

Claude Code sends background tasks such as session titles to the model in ANTHROPIC_DEFAULT_HAIKU_MODEL. On LLMTR, set it to anthropic/claude-haiku-5.5-free: $0.00 per 1M tokens until 14 October 2026, 23:59 (TRT), a 100,000-token prompt limit, and HTTP 410 after the end date.

Integration guides · 2026-10-10

Step 5 Preview API: OpenAI-compatible access, 1M context and pricing

On LLMTR, StepFun Step 5 Preview is called as stepfun/step-5-preview at https://llmtr.com/v1/chat/completions. The context is 1,024,000 tokens; per 1M tokens, input costs $1.00, cached input $0.10 and output $2.70. reasoning_effort accepts low, medium and high.

Integration guides · 2026-10-10

Video and image input with Step 5 Preview: an API request with base64

stepfun/step-5-preview reads base64 video and images sent as video_url and image_url parts in a https://llmtr.com/v1/chat/completions request and returns a text answer. Base64 video is limited to about 9 MB, a message carries at most 16 parts, and media is billed as input tokens at $1.00 per 1M tokens.

Model comparison · 2026-10-10

Step 5 Preview or Step 3.7 Flash? Context and cost for agent work

stepfun/step-5-preview has a 1,024,000-token context and costs $1.00 / $0.10 / $2.70, while stepfun/step-3.7-flash has a 256,000-token context and costs $0.20 / $0.04 / $1.15 (input / cached input / output per 1M tokens). Both are called with the same key at https://llmtr.com/v1/chat/completions; Step 5 Preview fits long context and json_schema, Step 3.7 Flash short, frequent turns.

RAG and data · 2026-10-10

Image classification API: yes/no, category and quality score with Perplexity Decider

perplexity/pplx-decider-v1.1-27b evaluates a base64 image added to the state array against your questions and returns a yes/no probability, a category or a quality score. On LLMTR it is called at https://llmtr.com/v1/systemone; the image is counted again for each question, at most about 2,050 tokens per question.

Integration guides · 2026-10-10

Gemini Nano Banana 2.1 API: 1K, 2K and 4K image pricing and how to call it

On LLMTR, google/gemini-nano-banana-2.1 is called through https://llmtr.com/v1/images/generations. Image output costs $30 per 1M tokens; Google's per-image equivalents are $0.0336 for 1K, $0.0504 for 2K and $0.113 for 4K. Input is billed at $1.50 and thinking at $7.50; n is not applied.

Integration guides · 2026-10-10

Editing with reference images, multi-reference and try-on in Nano Banana 2.1

To edit with Nano Banana 2.1, upload the reference image to https://llmtr.com/v1/files and add the returned id to a /v1/images/generations request through file_ids. One request takes up to 8 references; multi-reference and try-on use the same flow. /v1/images/edits is not open for Google models.

Model comparison · 2026-10-10

Nano Banana 2.1 or Nano Banana 2? Per-image pricing and differences

Both models are called on LLMTR through https://llmtr.com/v1/images/generations with the same key. Nano Banana 2.1 has lower per-image equivalents (1K $0.0336, 2K $0.0504, 4K $0.113; Nano Banana 2: $0.067, $0.101, $0.151), while its input and thinking tokens cost more. The 512 size is published only for Nano Banana 2.

Integration guides · 2026-10-10

Mistral Large 4 API: OpenAI SDK usage and pricing

On LLMTR, mistral/mistral-large-4 is called with the OpenAI SDK at https://llmtr.com/v1/chat/completions. It offers a 524,288-token context and up to 262,144 output tokens per response. Per 1M tokens it costs $0.68 input / $2.09 output until 20 October 2026, 00:00 UTC, and $1.36 / $4.18 from then on.

Agent workflows · 2026-10-10

A tool-calling agent with Mistral Large 4: tool_choice and reasoning

Mistral Large 4 calls tools on LLMTR through /v1/chat/completions. For tool_choice, auto, none and a named function work; required returns 400. In a multi-turn loop, assistant messages go back with reasoning_content, and that trace is billed as input tokens.

Model comparison · 2026-10-10

Mistral Large 4 or Large 3? Context, reasoning and price

On LLMTR both models are called through /v1/chat/completions with the same API key. Large 4 offers a 524,288-token context and Large 3 256,000; Large 4 reasons by default and accepts the none and high levels. The mistral/mistral-large-latest id is pinned to Large 3 and does not move to Large 4 on its own.

Integration guides · 2026-10-12

Alice AI LLM API: OpenAI SDK usage, pricing and streaming

On LLMTR, yandex/alice-ai-llm is called with the OpenAI SDK at https://llmtr.com/v1/chat/completions. It offers a 131,072-token context; per 1M tokens it costs $4.09836 for input and $9.836064 for output. The model works with text only and supports tool calling, JSON output and streaming.

Integration guides · 2026-10-12

Prompt caching on Alice AI LLM Flash: cached_tokens and cost

Alice AI LLM Flash is called on LLMTR through /v1/chat/completions and can read a repeated prompt prefix from a cache. Cached tokens bill at $0.204918 per 1M, a quarter of the input rate. A hit is not guaranteed: keep the fixed part at the start and byte-identical, and put varying content at the end.

Integration guides · 2026-10-12

YandexGPT 5.1 Pro API: OpenAI SDK usage and pricing

On LLMTR, yandex/yandexgpt-5.1-pro is called with the OpenAI SDK at https://llmtr.com/v1/chat/completions. The model offers a 32,768-token context, and input and output both bill at $6.557376 per 1M tokens. It takes and returns text only; tool calling, JSON output and streaming are supported.

Integration guides · 2026-10-12

YandexGPT 5 Lite API: high-volume short tasks and pricing

On LLMTR, yandex/yandexgpt-5-lite is called with the OpenAI SDK at https://llmtr.com/v1/chat/completions. Its context window is 32,768 tokens, and input and output both bill at $1.639344 per 1M tokens. 100,000 messages of 250 input and 10 output tokens cost $42.622944.

Model comparison · 2026-10-12

Yandex models compared: Alice AI LLM, Flash, YandexGPT Pro and Lite

On LLMTR the five Yandex models are called through /v1/chat/completions with the same API key. Context ranges from 32,768 to 131,072 tokens; tool calling, JSON output and streaming are available on all five, and the cache discount only on Alice AI LLM Flash. A request with 10,000 input and 1,000 output tokens costs between $0.009836 and $0.108197 depending on the model.

Model comparison · 2026-10-12

Alice AI LLM vs Alice AI LLM Flash: context, pricing and choice

Alice AI LLM (yandex/alice-ai-llm) offers a 131,072-token context at $4.09836 input / $9.836064 output per 1M tokens. Alice AI LLM Flash (yandex/alice-ai-llm-flash) offers a 65,536-token context at $0.819672 / $1.639344 and a cache price of $0.204918. A request with 30,000 input and 2,000 output tokens costs $0.142623 on Alice AI LLM and $0.027869 on Flash.

Model comparison · 2026-10-12

YandexGPT 5.1 Pro or 5 Pro? Price difference and switching

On LLMTR both models are called through /v1/chat/completions with the same API key. Both offer a 32,768-token context, tool calling, JSON output and streaming. YandexGPT 5.1 Pro is the newer release of the YandexGPT Pro family; at $6.557376 per 1M tokens it costs one third less than 5 Pro at $9.836064.

Agent workflows · 2026-10-12

A tool-calling agent with Alice AI LLM Flash: tools, tool_calls and the loop

Alice AI LLM Flash calls tools on LLMTR through /v1/chat/completions with the OpenAI-style tools array. The loop reads tool_calls from the response, runs the tool, appends the result as a role tool message with its tool_call_id and repeats until no tool call is left. The tools field does not mean the model will call a tool; treat a response without a call as the final answer.

RAG and data · 2026-10-12

YandexGPT JSON output: extracting structured data from text with 5.1 Pro

On LLMTR, yandex/yandexgpt-5.1-pro returns a JSON object through /v1/chat/completions with response_format {"type": "json_object"}. State in the prompt that the answer must be JSON and list the keys, then validate the answer with json.loads and your own key and type checks. 1,000 documents of 2,000 input and 300 output tokens each cost $15.081965.

RAG and data · 2026-10-12

LLM API for Russian text: summarizing and classifying

On LLMTR, a Russian document is summarized with yandex/alice-ai-llm (131,072-token context) and many short messages are classified with yandex/yandexgpt-5-lite (32,768-token context); both are called at https://llmtr.com/v1/chat/completions. An 800-token summary of a 60,000-token document costs $0.25377, and labeling 10,000 short messages costs $2.62295.

Integration guides · 2026-10-16

Free Qwen API: Qwen3.8 27B on LLMTR until 2 November 2026

On LLMTR, heabsy/qwen3.8-27b-heabsy-free costs $0.00 per 1M tokens until 2 November 2026, 00:00 (TRT). Grant the Heabsy data-transfer consent in Settings before the first request. Use is limited per account per day; the context window is 262,144 tokens and responses go up to 32,768. After the end date, requests return HTTP 410 model_retired.

Trust and compliance · 2026-10-16

Free Qwen in Europe: where requests go and what is stored

On LLMTR, heabsy/qwen3.8-27b-heabsy-free costs $0.00 until 2 November 2026, 00:00 (TRT). According to Heabsy, requests are processed in EU/EEA data centres, and prompts and responses are neither stored nor used for training. Calling the row needs a Heabsy data-transfer consent on the account. The article compares the five Qwen3.8 27B rows in the catalog by processing location and data note.

Agent workflows · 2026-10-16

Free Qwen for AI agents: tool calling with Qwen3.8 27B

heabsy/qwen3.8-27b-heabsy-free costs $0.00 per 1M tokens until 2 November 2026, 00:00 (TRT), and its model card lists tool/function calling; max_tokens goes up to 32,768 and reasoning is on by default. Use is limited per account per day; after the end date, requests return HTTP 410 model_retired.

Integration guides · 2026-10-17

Glimpse API: a reasoning agent for images and video

On LLMTR, elorian/glimpse is called with the OpenAI SDK at https://llmtr.com/v1/chat/completions. It takes text, up to 20 images and one video by https address per request, and returns text. The context is 220,000 tokens; per 1M tokens it costs $0.60 for input, $0.06 for input read from cache and $3.00 for output.

Integration guides · 2026-10-17

Video analysis API: asking Glimpse about a video

On LLMTR you send a video to elorian/glimpse at https://llmtr.com/v1/chat/completions as an https address in a video_url part. One video per request is accepted; a video sent as base64 gets a 400. The video's frames are billed as input tokens at $0.60 per 1M tokens.

Pricing and budget · 2026-10-17

Glimpse reasoning_effort and cost: low, medium, xhigh

On LLMTR, reasoning_effort for elorian/glimpse takes low, medium and xhigh; when you send none, medium applies. The level does not change the token price: reasoning tokens and the agent's steps count in completion_tokens and are billed at the output price of $3.00 per 1M tokens.

Integration guides · 2026-10-17

Muse Spark Open API: OpenAI SDK usage and pricing

On LLMTR, meta/muse-spark-open is called with the OpenAI SDK at https://llmtr.com/v1/chat/completions. The context window is 1,048,576 tokens, and per 1M tokens it costs $1.25 input, $0.15 input read from cache and $4.25 output. The model reasons before every answer; the level is chosen with reasoning_effort as low, medium, high or xhigh.

Integration guides · 2026-10-17

Send images, video and audio to Muse Spark Open

On LLMTR, media reaches Muse Spark Open (meta/muse-spark-open) as a content part added to the message's content array in a request to https://llmtr.com/v1/chat/completions: image_url for images, video_url for video, input_audio for audio. Media tokens are billed inside prompt_tokens at the $1.25 per 1M token input price; the answer is text.

Pricing and budget · 2026-10-17

Muse Spark Open reasoning_effort: levels and cost

On LLMTR, Muse Spark Open (meta/muse-spark-open) reasons before every answer; you choose the level in a /v1/chat/completions request with reasoning_effort set to low, medium, high or xhigh. Without the field the model reasons at a high effort. Reasoning tokens are billed inside completion_tokens at the output price of $4.25 per 1M tokens; every 1,000 reasoning tokens cost $0.00425.

Integration guides · 2026-10-17

Free Nemotron 3.5 Super VL API: image and video input

On LLMTR, nvidia/nemotron-3.5-super-vl-free is called with the OpenAI SDK at https://llmtr.com/v1/chat/completions and is free within a daily usage quota, at $0.00 per 1M input tokens and $0.00 per 1M output tokens. Images and short videos go in as base64 data URLs, the context window is 1,000,000 tokens and reasoning is on by default.

Agent workflows · 2026-10-17

Tool-calling agent on the free Nemotron 3.5 Super VL

On the nvidia/nemotron-3.5-super-vl-free row the model card lists tool calling, and the price is $0.00 per 1M input tokens and $0.00 per 1M output tokens. The context window is 1,000,000 tokens, reasoning is on by default and use is limited by a daily quota; the model card also lists JSON Schema output and image and video input.

Model comparison · 2026-10-17

Nemotron 3.5 Super VL: free vs paid row

On LLMTR, Nemotron 3.5 Super VL comes under two ids. nvidia/nemotron-3.5-super-vl-free is free with a daily usage quota and offers a 1,000,000-token context; nvidia/nemotron-3.5-super-vl is deducted from credit at $0.15 input, $0.02 input read from cache and $0.50 output per 1M tokens, with a 262,144-token context. Their data notices differ.

Integration guides · 2026-10-17

Send images to the free Qwen3.8 27B: API guide

On LLMTR, qwen/qwen3.8-27b-free is called with the OpenAI SDK at https://llmtr.com/v1/chat/completions, and the model card lists image input. The image travels inside the message as an image_url part. The row is free with a daily usage quota; the context is 131,042 tokens and one response can carry up to 16,384 output tokens.

RAG and data · 2026-10-17

JSON output from the free Qwen3.8 27B: json_object guide

On LLMTR the model card for qwen/qwen3.8-27b-free lists JSON mode. You add response_format {"type": "json_object"} to a /v1/chat/completions request, describe the keys in the prompt, then parse the answer with json.loads and validate it in code. The card does not list JSON Schema (structured outputs). The row is free with a daily usage quota.

Integration guides · 2026-10-17

Moondream 3.1 API: base64 images, reasoning and pricing

On LLMTR, moondream/moondream3.1-9b-a2b is called with the OpenAI SDK at https://llmtr.com/v1/chat/completions. Images are accepted only as base64 data URLs. The context is 32,000 tokens, and 1M tokens cost $0.30 for input and $2.50 for output. Reasoning is off by default and is turned on with reasoning: true or :think.

Model comparison · 2026-10-17

Moondream 3.1 or Moondream 3 Preview? Differences and price

On LLMTR both models are called through /v1/chat/completions with the same API key and the same request format. In the catalog the context (32,000 tokens), the price ($0.30 input / $2.50 output per 1M tokens) and the capabilities are identical; the difference is the release. Moondream 3.1 is the family's newer release, and Moondream 3 Preview is the preview release of Moondream 3.

Integration guides · 2026-10-17

Turkish embedding API: a guide to EmbeddingMagibu 200M

On LLMTR, llmtr/embeddingmagibu-200m is called with the OpenAI SDK's embeddings.create at https://llmtr.com/v1/embeddings. Trained for Turkish text and hosted in Turkey, it accepts up to 8,192 tokens per text and returns a normalized 768-dimensional vector. It costs $0 from 15 October 2026, 00:00 to 20 October 2026, 23:59 (Turkey time) and $0.01 per 1M input tokens from 21 October 2026, 00:00 (Turkey time).

RAG and data · 2026-10-17

Turkish semantic search with EmbeddingMagibu, step by step

Semantic search over Turkish documents takes four steps: split the text into chunks, embed the chunks with prompt_name document, embed the question with query, and pick the chunks with the highest cosine similarity. On LLMTR the vectors come from llmtr/embeddingmagibu-200m at https://llmtr.com/v1/embeddings; the example runs as a single file with only the openai package.

Integration guides · 2026-10-17

Liquid d1-ultra API: probability decisions on /v1/systemone

liquid/d1-ultra is a decision model that evaluates typed questions against a state you send, with optional images, and generates no text. On LLMTR it is called at https://llmtr.com/v1/systemone; 1M input tokens cost $0.042, output costs $0, and every question is billed along with the state.

RAG and data · 2026-10-17

Categorizing a product catalog with d1-ultra, with review

liquid/d1-ultra is a decision model that returns a probability for every category question you ask about a product description, and generates no text. This article categorizes a catalog through https://llmtr.com/v1/systemone with a two-level tree; products whose confidence stays below the example threshold, or whose request fails, go to a review queue.

Model comparison · 2026-10-17

Jev 1.14 or Jev 1.13? Differences, pricing and migration

On LLMTR the two versions of TypeSafe's Jev sit under two ids: typesafe/jev-1.14.0 (Jev 1.14) and typesafe/jev, pinned to Jev 1.13. Both run on /v1/systemone at $0.042 per 1M input tokens and $0 for output; Box questions exist only on Jev 1.14, and the two versions can return slightly different probabilities for the same question.

Integration guides · 2026-10-17

Jev 1.14 Box: finding an object's location in an image

typesafe/jev-1.14.0 answers a Box question with the box around an object in an image, as four edges given as fractions between 0 and 1 of the image's width and height. The request goes to https://llmtr.com/v1/systemone with the TypeSafe-Danger-Zone-Images: true header. Box has no not-found answer; ask whether the object is there with a noul question in the same request.

Model comparison · 2026-10-17

Models added to LLMTR in October 2026: facts and guide

Eight of the models LLMTR added in October 2026 in one table: Glimpse, Muse Spark Open, Nemotron 3.5 Super VL (paid and free rows), Moondream 3.1, EmbeddingMagibu 200M, the two decision models on /v1/systemone and Beam 1.0, free until 22 October 2026, 23:59 (Turkey time). All are called at https://llmtr.com/v1 with the same key; the free Qwen3.8 27B row, added earlier, is listed next to them.

Integration guides · 2026-10-17

Arabic language model API: a first request with Jais 3

On LLMTR, Jais 3 is called with the OpenAI SDK at https://llmtr.com/v1/chat/completions under the id ifm/jais-3-mova-36b-a4b. The model is built for Arabic and English, carries a 524,288-token context and accepts text only. It is free until 5 November 2026, 23:59 (TRT), with a daily token allowance.

Agent workflows · 2026-10-17

An Arabic agent that calls tools with Jais 3: the OpenAI tools loop

On LLMTR, Jais 3 calls tools in the OpenAI format. An agent loop uses tool_choice auto: the model calls the tool, gets the result back in a tool message and answers in Arabic. required, a named function and none with tools attached return 400 on this model. The model is free with a daily allowance until 5 November 2026, 23:59 (TRT).

RAG and data · 2026-10-17

JSON output from Arabic text: Jais 3 and json_schema

On LLMTR, ifm/jais-3-mova-36b-a4b accepts json_object and json_schema through response_format. To turn an Arabic customer message into a JSON object with a category, an order number and an Arabic summary, you send the schema as strict and validate the result in code. The model is free with a daily allowance until 5 November 2026, 23:59 (TRT).

Integration guides · 2026-10-17

reasoning_effort on Jais 3: the difference between low and high

On LLMTR, Jais 3 accepts two reasoning levels: high is the default and the model reasons before answering, while low turns reasoning off. medium, none and minimal return 400; reasoning enabled false is the same as low. Reasoning tokens count toward max_tokens and toward the daily allowance, which runs until 5 November 2026.

Integration guides · 2026-10-17

The K2 Horizon API: a 375B-parameter open-weight model

On LLMTR, K2 Horizon is called at https://llmtr.com/v1/chat/completions under the id ifm/k2-horizon-375b-a23b. It is IFM's Apache 2.0 model with 375 billion total and 23 billion active parameters; it reasons before every answer and carries a 524,288-token context. It is free until 5 November 2026, 23:59 (TRT), with a daily token allowance.

Integration guides · 2026-10-17

reasoning_effort and max_tokens on K2 Horizon

On LLMTR, K2 Horizon does not turn reasoning off; reasoning_effort selects medium or high, and low, none and minimal return 400. Because reasoning tokens count toward max_tokens, a short limit can leave the answer empty; when finish_reason is length, retry with a higher limit. The model is free with a daily allowance until 5 November 2026.

Model comparison · 2026-10-17

Jais 3 or K2 Horizon? Choosing between IFM's two models

Jais 3 and K2 Horizon are models from the same maker, MBZUAI's IFM, and on LLMTR both come with a 524,288-token context, free with a daily allowance until 5 November 2026. Jais 3 is for Arabic and English, 36B (4B active); K2 Horizon is 375B (23B active) and reasons before every answer. Forcing a tool works only on K2 Horizon.

Integration guides · 2026-10-17

Transcribing Arabic recordings through an API: Jais ASR

On LLMTR, Jais ASR is called at https://llmtr.com/v1/audio/transcriptions under the id ifm/jais-asr. The audio travels base64-encoded in a JSON body; the response carries the transcript, the recording's language or dialect, its duration and whether it contains speech. The model is free until 5 November 2026, 23:59 (TRT), with an allowance of 60 minutes of audio a day.

RAG and data · 2026-10-17

Arabic dialect identification API: labelling recordings with Jais ASR

On LLMTR, Jais ASR's verbose_json response carries the recognised language or dialect in the language field next to the transcript: msa for Modern Standard Arabic, and country codes such as MA, EG or SA for dialects. For identification, leave the language field out or send ar. The model is free with 60 minutes of audio a day until 5 November 2026.

Model comparison · 2026-10-17

Arabic models on LLMTR: Jais 3, Jais ASR and K2 Horizon

In October 2026 three models from MBZUAI's IFM were added to LLMTR: ifm/jais-3-mova-36b-a4b for Arabic and English chat, ifm/jais-asr for Arabic recordings, and ifm/k2-horizon-375b-a23b, which reasons before every answer. All three are free with a daily allowance until 5 November 2026, 23:59 (TRT), and are retired on 6 November.

Integration guides · 2026-10-19

MiniMax M3.1 Flash API: first request, reasoning_effort, price

On LLMTR, MiniMax M3.1 Flash is called as minimax/minimax-m3.1-flash on /v1/chat/completions. It has a 1,000,000-token context, returns at most 524,288 tokens per response and takes text, image and video input; it costs $0.15 input, $0.015 cached input and $0.60 output per 1M tokens.

Model comparison · 2026-10-19

MiniMax M3.1 Flash or M3? Price and feature differences

On LLMTR both models are called with the same API key on /v1/chat/completions, offer a 1,000,000-token context and 524,288 tokens per response, and take image and video. M3.1 Flash costs $0.15 / $0.015 / $0.60 at a single tier; M3 starts at $0.30 / $0.06 / $1.20 and doubles for prompts above 512,000 tokens.

Model comparison · 2026-10-19

1M-context Flash models: MiniMax, DeepSeek, Qwen and GLM

On LLMTR all four models have a 1,000,000-token context and are called with the same API key on /v1/chat/completions. Their prices are close; the differences are in the output limit per response, video input, reasoning control, JSON support and DeepSeek's weekday peak-hour price.

Model comparison · 2026-10-19

MiniMax M3.1 Flash vs Haiku 5.5, GPT-6 Luna, Gemini 3.8 Flash

On LLMTR all four models are called with the same API key on /v1/chat/completions. For prompts up to 100,000 tokens Claude Haiku 5.5 and GPT-6 Luna cost less; once a prompt passes 272,000 tokens MiniMax M3.1 Flash moves ahead, because its price does not change up to 1,000,000 tokens. Video input is on MiniMax M3.1 Flash and Gemini 3.8 Flash.

Model comparison · 2026-10-19

Which MiniMax model? M3.1 Flash, M3, M2.7 and M2.5

On LLMTR six MiniMax text models are called with the same API key on /v1/chat/completions. M3.1 Flash and M3 offer a 1,000,000-token context and image and video input; M2.7 and M2.5 are text models with a 204,800-token context. The Highspeed versions give faster output at twice the input and output price.

Model comparison · 2026-10-19

Qwen-Image 2.1 vs Grok Imagine vs Nano Banana 2.1

On LLMTR all five models generate and edit images with the same API key. In a single-reference edit Qwen-Image 2.1 Turbo costs $0.016 per generated image and the reference you send is not billed. On Grok Imagine the image you send is counted too, and 4K output is only on Nano Banana 2.1.

Model comparison · 2026-10-19

DeepSeek V4.1 Flash 1017: which model id to choose

DeepSeek V4.1 Flash 1017 is called on LLMTR with three ids. The row that, per the provider's statement, processes requests inside the European Union is TensorX's. If cost comes first, deepseek/deepseek-v4.1-flash-1017 serves the same release at a lower price that does not change by hour; deepseek/deepseek-flash follows DeepSeek's current release.

Model comparison · 2026-10-19

Jais 3 or Claude Haiku 5.5 for Arabic?

Both models are called on LLMTR with the same API key through /v1/chat/completions. Jais 3 was built for Arabic and English and is free with a daily allowance until 5 November 2026, 23:59 (TRT), after which it returns 410. Claude Haiku 5.5 is paid and Anthropic lists Arabic; it is the option for customer data and for production after 6 November.

Model comparison · 2026-10-19

K2 Horizon vs DeepSeek V4 Pro, GLM-5.3, Nemotron 3 Ultra

All four models are called on LLMTR with the same API key through /v1/chat/completions. K2 Horizon is free until 5 November 2026, 23:59 (TRT) and reasons before every answer. The free option with no end date is Nemotron 3 Ultra. On the paid side, DeepSeek V4 Pro costs less than GLM-5.3 outside peak hours.

Model comparison · 2026-10-19

Jais ASR vs Grok Voice STT vs Gemini 3.8 Flash for Arabic

All three models run on LLMTR with the same API key. Jais ASR transcribes the Arabic dialects of 19 countries, names the dialect and is free for 60 minutes a day until 5 November 2026, 23:59 (TRT). Grok Voice STT costs $0.10 per hour of audio; Gemini 3.8 Flash can translate or summarize the transcript in the same request.

Model comparison · 2026-10-29

Mercury 3 vs Mercury 2.5 Preview vs Mercury 2

All three Mercury models are called on LLMTR with the same API key on /v1/chat/completions. Mercury 3 has a 1,000,000-token window and a 131,072-token output limit; its per-token price is above Mercury 2.5 Preview's and below Mercury 2's. Mercury 2.5 Preview and Mercury 2 retire on 15 November 2026, and inception/mercury-3 is the replacement for both.

Model comparison · 2026-10-29

Mercury 3 vs Claude Haiku 5.5, GPT-6 Luna, Gemini 3.8 Flash

On LLMTR all four models are called with the same API key on /v1/chat/completions. Mercury 3 accepts text only; at its introductory price through 2 November 2026 it matches Haiku 5.5 and GPT-6 Luna on input and is below them on output, and from 3 November Haiku and Luna cost less on short prompts. For images, PDF, audio and video the other three models come first.

Model comparison · 2026-10-22

Shanghai AI Lab models: Intern-S2 to Sage-A1.5 compared

Shanghai AI Lab's five models are called on LLMTR with the same API key through /v1/chat/completions and are free until 31 October 2026, 23:59 (TRT). Intern-S2 is for scientific questions with images, Atria-Dawn for long text-based agent work, Atria-Dawn Preview for trying the open release, Agents-A1 for tool agents that read images, and Sage-A1.5 for flows with subagents. From 1 November 2026 all five return 410.

Integration guides · 2026-10-22

Reading scientific charts and tables with the Intern-S2 API

On LLMTR, Intern-S2 is called with the model id internlm/intern-s2-free at https://llmtr.com/v1/chat/completions with the OpenAI SDK, accepts images alongside text and is free until 31 October 2026, 23:59 (TRT). Send the image of a scientific figure or table as a base64 data URL and ask for the rows as JSON in the prompt; response_format is not supported on this row. The allowance is 500,000 tokens a day per model.

Model comparison · 2026-10-22

Atria-Dawn vs Atria-Dawn Preview: stable release or open preview

Both rows are called on LLMTR with the same API key through /v1/chat/completions; they offer a 524,288-token context, text-only input, json_schema and tool calling, and they are free until 31 October 2026, 23:59 (TRT). Atria-Dawn is the stable release; Atria-Dawn Preview is the preview whose weights are open under the MIT license.

Model comparison · 2026-10-22

Agents-A1 vs Sage-A1.5: which agent model for which job

Both models are called on LLMTR through /v1/chat/completions with the same API key and are free until 31 October 2026, 23:59 (TRT). Agents-A1 is a general agent model; Sage-A1.5 is built on Agents-A1 and Qwen3.6, and its model card describes subagent orchestration through a spawn_subagent tool you define. From 1 November 2026 both return 410.

Agent workflows · 2026-10-22

Building a tool-calling agent loop with Agents-A1

Agents-A1 runs a multi-turn tool loop on LLMTR through /v1/chat/completions and is free until 31 October 2026, 23:59 (TRT). Define the tools in tools, send auto as tool_choice, run every returned tool call and send the result back in a role: "tool" message, and cap the number of steps. Sage-A1.5 runs the same loop.

Pricing and budget · 2026-10-22

Reasoning and the daily token allowance on Shanghai AI Lab models

All five rows reason before answering by default, and reasoning tokens count toward both max_tokens and the daily token allowance. To turn reasoning off, send "reasoning": false or add :fast to the model id; reasoning_effort returns 400. The rows are free until 31 October 2026, 23:59 (TRT).

RAG and data · 2026-10-22

Long-document extraction with Atria-Dawn and JSON Schema

Atria-Dawn can read a long document in one request and return fields that follow a strict JSON Schema, and it is free until 31 October 2026, 23:59 (TRT). Input, output and reasoning share the same context window and the same daily allowance; turning reasoning off for extraction saves tokens.

Model comparison · 2026-10-22

Intern-S2 vs Qwen3.8 27B, Nemotron 3.5 Super VL, Agnes 2.5 Flash

All four models read images on LLMTR with the same API key through /v1/chat/completions. Intern-S2 is free until 31 October 2026, 23:59 (TRT) and returns 410 from 1 November 2026; Qwen3.8 27B, Nemotron 3.5 Super VL and Agnes 2.5 Flash run on quotas with no end date set and can be used free of charge after 31 October 2026 as well. Video is open only on Nemotron 3.5 Super VL and Agnes 2.5 Flash; Agnes takes images only as base64 data URLs.

Model comparison · 2026-10-22

Agents-A1, Sage-A1.5 and Qwen3.6 35B-A3B compared

All three models are called on LLMTR through /v1/chat/completions with the same API key, and all three are MoE models with 35 billion total and 3 billion active parameters. Agents-A1 and Sage-A1.5 are free until 31 October 2026, 23:59 (TRT) and return 410 from 1 November 2026. Qwen3.6 35B-A3B is paid: $0.07 input and $0.70 output per 1M tokens.

Model comparison · 2026-10-22

Shanghai AI Lab models after 1 November 2026: migration guide

The five Shanghai AI Lab models are free until 31 October 2026, 23:59 (TRT); from 1 November 2026, 00:00, they return 410 model_retired and are not redirected. Scientific questions with images and image-reading agents move to Nemotron 3.5 Super VL or Qwen3.6 35B-A3B; long text-based agent work moves to Nemotron 3 Ultra, Agnes 2.5 Flash, DeepSeek V4.1 Flash or MiniMax M3.1 Flash.

Model comparison · 2026-10-26

Nanbeige4.2 3B vs Ministral 3 3B: two small models

Both models are called on LLMTR through /v1/chat/completions with the same API key. Nanbeige4.2 3B is hosted in Turkey and costs $1.00 input and $3.00 output per 1M tokens. Ministral 3 3B costs $0.10 input and $0.10 output, reads images and does not carry LLMTR's Turkey hosting mark.

Model comparison · 2026-10-26

Nanbeige4.2 3B vs Qwen 3.8 27B: two Turkey-hosted models

Nanbeige4.2 3B and Qwen 3.8 27B are both hosted in Turkey on LLMTR and are called through /v1/chat/completions with the same API key. Nanbeige4.2 3B costs $1.00 input and $3.00 output per 1M tokens. Qwen 3.8 27B costs $5.00 input and $7.00 output and offers a 262,144-token context and JSON schema output.

Model comparison · 2026-10-26

Turkish text classification: Nanbeige4.2 3B or Leyla?

Nanbeige4.2 3B and Leyla are both hosted in Turkey on LLMTR and are called with the same API key. Leyla answers the questions you define on /v1/systemone with a choice, a score or a yes-probability, at $0.03 per 1M input tokens. Nanbeige4.2 3B, on /v1/chat/completions, also writes a reason or a summary next to the label, at $1.00 input and $3.00 output.

Model comparison · 2026-10-26

Nanbeige4.2 3B vs GPT-5 nano: choosing a small model

Both models are called on LLMTR through /v1/chat/completions with the same API key. Nanbeige4.2 3B is hosted in Turkey and costs $1.00 input and $3.00 output per 1M tokens. GPT-5 nano costs $0.05 input and $0.40 output, reads images and does not carry LLMTR's Turkey hosting mark. OpenAI shuts GPT-5 nano down on 11 December 2026; after that date GPT-6 Luna can be used on LLMTR.

Model comparisons · 2026-10-28

What is Giotto Alto? API pricing, per-request cost and reasoning levels

giotto/giotto-alto is Giotto.ai's model that accepts text and image input. Input costs about $2.59 and output about $12.93 per 1M tokens; the context window is 120,000 tokens. Every request carries the model's own instructions, about 8,000 input tokens.

Integration guides · 2026-10-30

Qwen-SEA-LION-v4.5-27B-IT API: first request with LLMTR

Qwen-SEA-LION-v4.5-27B-IT is called on LLMTR as marscompute/qwen-sea-lion-v4.5-27b through https://llmtr.com/v1/chat/completions. Prompt and reply together can be 8,192 tokens. The price is $0.60 input and $3.60 output per 1M tokens; a 1,500-token prompt with a 500-token reply costs $0.0027.

Integration guides · 2026-10-30

A language-model API for Thai, Indonesian, Vietnamese and Malay

AI Singapore's Qwen-SEA-LION-v4.5-27B-IT is trained for English and seven Southeast Asian languages and is called on LLMTR as marscompute/qwen-sea-lion-v4.5-27b. Turkish is not among its supported languages. The price is $0.60 input and $3.60 output per 1M tokens.

Model comparison · 2026-10-30

SEA-LION v4.5 vs Gemma 4 31B vs Qwen3.8 27B for Southeast Asia

All three models are called on LLMTR through /v1/chat/completions with the same API key. Qwen-SEA-LION-v4.5-27B-IT costs $0.60 input and $3.60 output per 1M tokens and offers an 8,192-token window. Gemma 4 31B costs $0.14 and $0.40, Qwen3.8 27B $0.40 and $2.40; both offer a 262,144-token context, tool calling and image input.

Integration guides · 2026-10-30

SEA-LION's 8,192 tokens: long text, max_tokens and cost

On Qwen-SEA-LION-v4.5-27B-IT, prompt and reply together can be at most 8,192 tokens, and the prompt text cannot exceed about 32,000 characters. Split long text into pieces and set max_tokens so that it and the prompt stay within 8,192. A 5,000-token prompt with a 1,000-token reply costs $0.0066.

Integration guides · 2026-11-03

GLM-5.3-FlashX-Legal API: first request with LLMTR

GLM-5.3-FlashX-Legal is called on LLMTR as zai/glm-5.3-flashx-legal through https://llmtr.com/v1/chat/completions. The context window is 1,000,000 tokens and one response can carry at most 131,072 output tokens. The price is $0.333 input and $1.125 output per 1M tokens; a 20,000-token prompt with a 2,000-token reply costs $0.00891.

Integration guides · 2026-11-03

AI contract review: the API and a JSON Schema

Contract review runs on LLMTR with the zai/glm-5.3-flashx-legal model through https://llmtr.com/v1/chat/completions. When the response_format field carries a JSON Schema, the response is an object that matches the schema. A request that compares two 15,000-token versions and produces a 3,000-token reply costs $0.013365.

RAG and data · 2026-11-03

AI for legal research: legislation search and Voyage Law 2

In legal research two models work together through LLMTR: voyageai/voyage-law-2 indexes your legislation and case texts through /v1/embeddings, and zai/glm-5.3-flashx-legal makes a tool call through /v1/chat/completions and answers from the returned passages, citing them. An index of 1,000,000 tokens costs $0.12 once.

Integration guides · 2026-11-03

Hydrogen-1.0 API: first request with LLMTR

Hydrogen-1.0 is called on LLMTR as deep-principle/hydrogen-1.0 through https://llmtr.com/v1/chat/completions. The context window is 1,000,000 tokens and input is text only. The price is $1.40 input and $3.80 output per 1M tokens; a 2,000-token prompt with a 4,000-token reply costs $0.018.

Integration guides · 2026-11-03

AI for chemistry: experiment planning with Hydrogen-1.0

Hydrogen-1.0 is called on LLMTR as deep-principle/hydrogen-1.0 through https://llmtr.com/v1/chat/completions and returns an experiment plan that matches the JSON Schema given in response_format. A 3,000-token prompt with an 8,000-token reply costs $0.0346.

Integration guides · 2026-11-03

Drex Max Finance API: first request with LLMTR

Drex Max Finance is called on LLMTR as nace/drex-max-finance through https://llmtr.com/v1/systemone. One request takes up to 512 questions, and the 262,144-token limit covers the whole request, state, images and questions included. Input costs $0.03 per 1M tokens and output tokens are not billed.

Integration guides · 2026-11-03

AI invoice review: pay, hold or reject an invoice

For AI invoice review, the invoice image goes in the images array and the purchase order in the state field, sent to nace/drex-max-finance through https://llmtr.com/v1/systemone. The model returns a probability or one of the options you list for each question. A 3,000-token check costs $0.00009.

Integration guides · 2026-11-04

Hume Audio Decision Model API: your first request with LLMTR

Hume Audio Decision Model is called on LLMTR as hume/audio-decision-model through https://llmtr.com/v1/systemone. It takes one MP3 or WAV recording (5 seconds to 5 minutes) and typed questions, and answers each question with probabilities instead of text. Input costs $0.47 per 1M tokens and output tokens are not billed.

Integration guides · 2026-11-04

Testing voice agent and TTS quality with AI

For TTS quality testing, the test recording goes in the audio field and the script with a style note in the state field, sent to hume/audio-decision-model through https://llmtr.com/v1/systemone. Naturalness and style come back as scores, the script's words and talking over the caller as probabilities and glitches as time spans; two voices are compared with the same questions in the same order.

Integration guides · 2026-11-04

Support call analysis with AI: tone, impatience and cancellation

For call recording analysis, the recording goes in the audio field to hume/audio-decision-model through https://llmtr.com/v1/systemone, and the agent's tone, the caller's impatience, background noise and the cancel moment are asked as four separate questions. Your code combines the answers into one decision with thresholds.

Integration guides · 2026-11-04

Cohere Embed 5 API: first request for text and image search

Cohere Embed 5 is called on LLMTR as cohere/embed-v5.0-pro and cohere/embed-v5.0-fast through https://llmtr.com/v1/embeddings. It places text, images and inputs that combine both in one vector space, so you can index a collection with Pro and send queries with Fast. 1M tokens cost $0.12 on Pro and $0.08 on Fast.

Integration guides · 2026-11-04

Cohere Rerank 5 and Context Focus: fewer tokens in RAG

Cohere Rerank 5 is called on LLMTR as cohere/rerank-v5.0-pro and cohere/rerank-v5.0-fast through https://llmtr.com/v1/rerank. It scores documents against a query between 0 and 1, and with Context Focus (relevant_spans) enabled it also returns the relevant passages in each document. 1M tokens cost $0.05 on Pro and $0.02 on Fast.

Integration guides · 2026-11-04

Morph Reflex API: first request with LLMTR

Morph Reflex is called on LLMTR as morph/reflex through https://llmtr.com/v1/systemone. It answers typed questions about text, JSON and images with probabilities, without generating text. Input costs $0.042 per 1M tokens and cached input $0.008; output is not billed.

Integration guides · 2026-11-04

AI image classification: probability thresholds and caching

For AI image classification, the photo goes in the images array and the questions in the questions field, sent to morph/reflex through https://llmtr.com/v1/systemone. The model returns an answer with probabilities for each question, and a person looks at a photo on the wrong side of the threshold. When the questions are asked in one request, repeated input can be read from the cache.

Integration guides · 2026-11-05

GPT Image 2.6 API: Flare or Sunburst, and your first request

GPT Image 2.6 is called on LLMTR as openai/gpt-image-2.6-flare and openai/gpt-image-2.6-sunburst through https://llmtr.com/v1/images/generations. The two models share fields and prices: per 1M tokens, $5 text input, $8 image input and $30 image output. Pick the model by the kind of work, quality by the detail you need and complexity by how much planning the scene takes.

Integration guides · 2026-11-05

Animated GIFs and markup-guided image edits through an API

To generate a GIF through an API, add output_format gif to a request to https://llmtr.com/v1/images/generations; frame_count is 2 to 16 and frame_duration_ms 20 to 2000. To edit an image, send images and a markup field carrying a transparent PNG to https://llmtr.com/v1/images/edits; there is no separate fee for markup.

Integration guides · 2026-11-05

OpenRouter Auto Router: how model routing works

Auto Router is OpenRouter's model router for apps, chat and general agents. You can try the id openrouter/auto on LLMTR at https://llmtr.com/v1/chat/completions. This article explains what a router is, how OpenRouter says it decides, which fields to log, and when a pinned model is the better choice.

Integration guides · 2026-11-05

OpenRouter Auto Code: a router for coding agents

Auto Code is OpenRouter's model router for coding agents and harnesses. You can try the id openrouter/auto-code on LLMTR at https://llmtr.com/v1/responses. This article explains how Auto Code differs from Auto Router, how OpenRouter says it decides, and what to log in an agent loop.

Integration guides · 2026-11-07

Alpie Core API: first request with LLMTR

Alpie Core is called on LLMTR as 169pi/alpie-core through https://llmtr.com/v1/chat/completions. The context window is 65,000 tokens and input is text only. The price is $0.15 input and $0.40 output per 1M tokens; a 200-token prompt with 500 output tokens costs $0.00023.

Integration guides · 2026-11-07

What is a 4-bit quantized model? Alpie Core memory and cost

A 4-bit quantized model stores each weight in 4 bits; against 16 bits the weight memory falls to a quarter, and accuracy has to be evaluated per task. Alpie Core is called on LLMTR as 169pi/alpie-core through https://llmtr.com/v1/chat/completions; 169Pi has opened the weights under Apache 2.0.