GPT-6 Astra guides · 2026-09-05
GPT-6 Astra API Pricing: Cache Writes and Real Costs
Calculate GPT-6 Astra API costs with standard and long-context rates, cache write and read charges, and a worked example that separates input from output.
Four prices, not one input price
This article prices the paid openai/gpt-6-astra model, not the separately quota-limited free card. An Astra estimate needs separate quantities for ordinary input, cache writes, cache reads and output. A repeated prompt is not enough to assume a cache hit: use the usage returned for the request. The table shows standard provider prices per one million tokens, checked on 5 September 2026.
| Metric | Up to 272,000 input tokens | Above 272,000 input tokens |
|---|---|---|
| Ordinary input | $10.00 | $20.00 |
| Cache read | $1.00 | $2.00 |
| Cache write | $12.50 | $25.00 |
| Output | $50.00 | $75.00 |
Do not count cached tokens twice
In LLMTR’s Astra accounting, cache reads and writes are subsets of input. Subtract them from total input before charging the remaining ordinary input. Adding a cache charge to the full input charge would count the same tokens twice.
The calculation is: ordinary input × input rate + cache writes × write rate + cache reads × read rate + output × output rate, all divided by one million. Output quantity must include billed reasoning tokens; the text you see is not necessarily the full billed output.
Worked example: cold and warm requests
Assume 10,000 input tokens: 8,000 are written to cache, 2,000 remain ordinary input, and the response has 1,000 total billed output tokens. This is an illustrative calculation, not a measured performance result. The cold request costs 0.02 + 0.10 + 0.05 = $0.17.
If a later request reads the same 8,000 cached tokens, with unchanged ordinary input and output quantities, it costs 0.02 + 0.008 + 0.05 = $0.078. The total falls by about 54.1%, not 90%. The lower cache-read rate applies only to the eligible input portion; output still costs the same in this example.
The long-context threshold applies to the whole request
When input exceeds 272,000 tokens, OpenAI lists double input and cache rates and 1.5 times the output rate for the full request. Do not apply the higher price only to tokens beyond the threshold. Exactly 272,000 and more than 272,000 are different cases.
Before sending a large document collection, remove irrelevant material and decide whether every document is necessary for the question. Splitting an inseparable analysis merely to avoid the threshold can lose important context; evaluate answer quality alongside cost.
Build a budget from measured usage
Track total billed output, cache-hit quantity and retries over representative tasks. A reasoning level changes the amount of work the model may perform, so budget for the complete response rather than the length of visible prose. Keep an explicit output budget and inspect usage before increasing volume.
These are model token charges, not a quotation for every possible service or tool. LLMTR keeps provider model rates separate from its credit top-up margin. Consult the current pricing page for payment terms and the linked integration guide for your first request.
Frequently asked questions
Are cache writes free on Astra?
No. The published standard cache-write rate is $12.50 per million tokens; cache reads are $1.00. Long-context rates differ.
Does a cache hit reduce the whole bill by 90%?
No. The input discount applies to cache-read tokens. Ordinary input and output remain separate charges, as the worked example demonstrates.