Jais 3 or Claude Haiku 5.5 for Arabic?

Short answer: Jais 3 to try it until 5 November, Claude Haiku 5.5 for production

LLMTR is a gateway that gives access to language models from different providers through one OpenAI-compatible API. Jais 3 MoVA 36B A4B (ifm/jais-3-mova-36b-a4b) and Claude Haiku 5.5 (anthropic/claude-haiku-5.5) are called here on the /v1/chat/completions endpoint at https://llmtr.com/v1 with the same API key from your llmtr.com account; moving from one to the other is, for most requests, a change of the model string.

According to IFM's model card, Jais 3 was built for Arabic and English assistant and agent applications. It is free with a daily allowance of 200,000 tokens until 5 November 2026, 23:59 (TRT); from 6 November 2026, 00:00, requests are answered with 410 model_retired. Claude Haiku 5.5 is paid, and Anthropic's multilingual support page lists Arabic among the languages Claude is evaluated in. For Arabic work that carries customer data or has to keep running after 6 November, the option is Claude Haiku 5.5.

• IFM, Jais-3-MoVA-36B-A4B model card · 2026-10-19
• Anthropic, multilingual support · 2026-10-19

Jais 3 or Claude Haiku 5.5 for Arabic text?

Four questions decide it: how long the work will run, how sensitive the data you send is, whether you need images or a very long context, and how you steer tool calling. Jais 3's data notice says prompts and completions may be logged by the provider and used to improve the model; do not send sensitive, personal or customer data to it.

The table shows the two rows with the values in the LLMTR catalog and docs; prices are per 1M tokens, in the order input / cached input / output.

What does the same Arabic task cost on each model?

Take a customer support reply: a 3,000-token prompt and a 500-token answer in Arabic. On Claude Haiku 5.5 this request costs $0.00055, and 10,000 requests of the same size cost $5.50. On Jais 3 the same request takes nothing from your balance until 5 November 2026; the daily allowance of 200,000 tokens counts the sum of input and output and covers 57 requests of this size. The allowance renews every day at 00:00 (TRT); when it runs out the request returns 429, and the Retry-After header tells you when to try again. You can also get 429 before your own allowance runs out, because of a combined total counted across all users; newly opened accounts start with a lower allowance on their first day. On Claude Haiku 5.5, once a prompt passes 100,000 tokens, input, cache and output are all billed at five times the price; if you split long Arabic documents into parts, keep each part under that threshold.

The Python example below sends the same Arabic prompt to both models and prints the token counts and the answers. reasoning_effort: low is accepted by both models: it turns reasoning off on Jais 3 and keeps it short on Claude Haiku 5.5. Read the results side by side with your own Arabic samples.

OpenAI Python SDK example that sends the same Arabic customer support prompt to Jais 3 and Claude Haiku 5.5 through https://llmtr.com/v1/chat/completions.

import os

from openai import OpenAI

client = OpenAI(base_url="https://llmtr.com/v1", api_key=os.environ["LLMTR_API_KEY"])

messages = [
    {"role": "system", "content": "أجب باللغة العربية الفصحى وباختصار."},
    {"role": "user", "content": "اكتب رداً مهذباً على عميل يسأل عن موعد تسليم طلبه المتأخر."},
]
# reasoning_effort "low" is accepted by both: it turns reasoning off on Jais 3
# and keeps it short on Claude Haiku 5.5.
for model in ["ifm/jais-3-mova-36b-a4b", "anthropic/claude-haiku-5.5"]:
    response = client.chat.completions.create(
        model=model, messages=messages, max_tokens=2048, reasoning_effort="low"
    )
    usage = response.usage
    print(model, usage.prompt_tokens, usage.completion_tokens)
    print(response.choices[0].message.content)

Tool calling, JSON and reasoning

On Jais 3, tool_choice works only with auto: required or a named function returns 400, and none returns 400 while tools are attached. If you want a specific tool to be called, use auto and say so in the prompt; to turn tool use off, do not send the tools field at all. Jais 3 accepts json_object and json_schema in response_format; Claude Haiku 5.5's card does not list JSON Schema, so validate the output on your side.

Jais 3 reasons before answering by default; reasoning_effort selects high, or low to turn reasoning off, while medium, none and minimal return 400. Reasoning tokens count toward max_tokens and toward the daily allowance; when you add earlier assistant messages to the history, you do not need to send the reasoning_content field back. Claude Haiku 5.5 accepts six levels from none to max and defaults to medium; it does not support temperature and top_p, and if you send them they are removed and appear in the llmtr_dropped_parameters list of the response. On Haiku 5.5 the message list must end with a user message.

• Anthropic, Claude Haiku 5.5 overview · 2026-10-19

Which one when?

The choices below follow from the differences in duration, data notice, price and parameters; compare the quality of the Arabic text with your own samples.

• Trying a model built for Arabic at no cost until 5 November 2026, with non-sensitive text: Jais 3.
• Extracting schema-following JSON from Arabic text during the trial period: Jais 3, with json_schema.
• Customer data, personal data, or production traffic that has to keep running after 6 November 2026: Claude Haiku 5.5.
• Requests that carry images, or a context beyond 524,288 tokens: Claude Haiku 5.5.
• If you need to force a specific tool with tool_choice, Jais 3 does not fit; it accepts auto only.

Trying both with the same key

Open an account on llmtr.com and create a key under Dashboard > API Keys. Jais 3 runs without charging your balance until 5 November 2026; for Claude Haiku 5.5, top up your balance under Dashboard > Billing. The same key works on both models, and you do not need separate accounts with IFM and Anthropic.

If you start on Jais 3 and plan to move to Claude Haiku 5.5, make the move before 6 November: from that date requests return 410 and are not redirected to another model. When you switch, review the model id, your tool_choice use and your JSON validation in one pass.

Comparing Jais 3 and Claude Haiku 5.5 with an Arabic prompt

Send the same Arabic prompt to both models with one LLMTR key, compare the answers and token counts, and choose by what your work needs.

1. Key. Create an API key on llmtr.com; top up your balance for Claude Haiku 5.5.
2. Same prompt. Send your own Arabic prompts with the two model ids to the https://llmtr.com/v1/chat/completions endpoint.
3. Choice. Compare the answers, token counts and the data notice, then choose the model id.

Frequently asked questions

Until when is Jais 3 free?

Until 5 November 2026, 23:59 (TRT), with 200,000 tokens a day per account, input and output combined. From 6 November 2026, 00:00, the model returns 410 model_retired and requests are not redirected to another model.

Does Claude Haiku 5.5 support Arabic?

Anthropic's multilingual support page lists Arabic among the languages Claude is evaluated in. Compare the two models on the same prompt with your own Arabic samples.

What changes when moving from Jais 3 to Claude Haiku 5.5?

The model id changes and requests are paid: up to 100,000 tokens, $0.10 input and $0.50 output per 1M tokens. You can send images and use a context beyond 524,288 tokens; JSON Schema is not listed on the card, so validate the output on your side.

Related posts

• Jais 3 or K2 Horizon? Choosing between IFM's two models
• Arabic language model API: a first request with Jais 3
• K2 Horizon vs DeepSeek V4 Pro, GLM-5.3, Nemotron 3 Ultra
• Jais ASR vs Grok Voice STT vs Gemini 3.8 Flash for Arabic

• Jais 3 MoVA 36B A4B model
• Claude Haiku 5.5 model
• IFM Jais documentation
• Tool calling documentation