How Much Does GPT-6 Astra Token Usage Cost?

2026-09-20

Key Takeaways

GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens on OpenAI’s standard API tier, for requests at or under 272,000 input tokens.
Cached input drops to $1 per million, a 90 percent discount, while the first write to the cache costs $12.50 per million.
Crossing 272,000 input tokens in one request reprices the entire request at $20 input and $75 output per million — not just the tokens past the line.
Batch and Flex processing run at half the standard rate; Fast mode doubles it to $20 and $100 per million.
Reasoning tokens bill as output at $50 per million, and effort level is the single largest cost lever — independent testing found a fourfold swing between low and maximum effort.
Astra is 2.5 times GPT-5.6 Sol’s promotional rate, but that promotion is guaranteed only through 21 November 2026, after which the multiple falls to 2x on input and about 1.67x on output.
Because Astra completes tasks in fewer tokens, it can use up to 40 percent of Sol’s tokens on the same job and still produce an identical bill.

GPT-6 Astra. Image source: OpenAI

The headline answer is straightforward: GPT-6 Astra bills at $10 per million input tokens and $50 per million output tokens through the OpenAI API, with cached input at $1 and cache writes at $12.50. Those rates apply on the standard tier for any request that stays at or under 272,000 input tokens, and they are the same whether you call the model directly or through Microsoft Azure or AWS Bedrock.

The headline answer is also close to useless for budgeting. Three things move a real bill far more than the sticker rate: how much reasoning effort you request, whether a single request crosses OpenAI’s long-context threshold, and how well you reuse cached input. A long agent session that reuses its prompt efficiently can cost less than half what the same work costs without caching, and one oversized request can cost double what a slightly smaller one does.

The Full GPT-6 Astra Rate Card

Astra was released on 3 September 2026 as OpenAI’s flagship, succeeding GPT-5.6 Sol. One model-tracking site lists the release as 4 September; OpenAI’s own announcement and the bulk of coverage use 3 September, so treat the earlier date as canonical. Rates below were read from OpenAI’s pricing documentation on 16 September 2026.

Token type	Standard (≤272K input)	Long context (>272K input)	Batch / Flex	Fast mode
Input	$10.00	$20.00	$5.00	$20.00
Cached input	$1.00	$2.00	$0.50	$2.00
Cache writes	$12.50	$25.00	$6.25	$25.00
Output	$50.00	$75.00	$25.00	$100.00

All figures are per million tokens. Batch over the 272K threshold runs $10.00 input and $37.50 output. Two details catch people out. The cache write line is real and easy to miss — the first time you send a prefix it bills at 1.25 times the standard input rate, not at the cheap read rate. And the long-context surcharge applies to the whole request: a call at 280,000 tokens costs roughly double a call at 270,000 tokens for about 4 percent more input.

Model Specifications That Set the Bill

Astra carries a 1,050,000-token context window with a maximum input of 922,000 tokens and maximum output of 128,000. The knowledge cutoff is 30 April 2026. Reasoning effort accepts low, medium, high, xhigh and max, with no option to disable reasoning entirely — the floor is higher than on older models. The API model identifier is gpt-6-astra, and OpenAI’s documentation lists exactly one snapshot with no dated form and no “latest” alias.

Why You Will See Different Prices Quoted

Search for this model’s cost and you will find $10/$50, $5/$25, $11/$55 and $20/$100. None of those is wrong, and none is the list price changing.

$5/$25 is the batch tier. This is OpenAI’s own rate, not a reseller discount — batch processing runs at half the standard rate across every line. It is the right answer for overnight evaluation runs and the wrong answer for anything interactive.

$11/$55 is a 10 percent uplift for data residency. OpenAI documents a 10 percent surcharge on residency endpoints for models released on or after 5 March 2026, which includes Astra. Multiply list by 1.1 and the arithmetic matches exactly. Note that Fast mode is not available with EU data residency, so residency and speed cannot be combined.

$20/$100 is Fast mode. This delivers up to twice the speed of standard processing at twice the price. It is a different service level, not a different model.

Rates below list belong to third-party gateways. Some resellers advertise Astra under OpenAI’s rate. That is a commercial decision by the operator and carries whatever data-handling terms that operator applies.

How Astra Compares Across OpenAI’s Own Range

Model	Input per 1M	Cached input	Output per 1M
GPT-6 Astra	$10.00	$1.00	$50.00
GPT-5.6 Sol (promotional)	$4.00	$0.40	$20.00
GPT-5.6 Sol (standard)	$5.00	$0.50	$30.00
GPT-5.6 Terra	$2.00	$0.20	$12.00
GPT-5.6 Luna	$0.20	$0.02	$1.20
Claude Fable 5.1	$10.00	—	$50.00
Claude Opus 5	$5.00	$0.50	$25.00

Astra prices level with Anthropic’s Claude Fable 5.1 on both headline numbers and roughly double Claude Opus 5. Against Sol, the multiple depends entirely on which Sol rate you use. OpenAI’s pricing page describes Sol’s $4/$20 as promotional and available at least through 21 November 2026. Measured against that, Astra is exactly 2.5 times Sol on every line. Measured against Sol’s standard $5/$30, Astra is 2x on input and about 1.67x on output. Both are correct on the same day, and anyone budgeting past November should use the second pair.

For readers comparing across vendors rather than within OpenAI’s range, our breakdowns of Grok 4.5 API and plan pricing and Grok 4.5 against Claude Opus on coding cost use the same per-task method applied below.

A Worked Example: What a Real Agent Run Costs

Token rates mean little without a workload attached. Take a 200-call agent run of the kind used for an overnight migration. Each call carries 10,000 input tokens, of which 8,000 are cache reads and 2,000 are fresh, and produces 500 output tokens. Across the run that is 400,000 fresh input tokens, 1,600,000 cached input tokens and 100,000 output tokens, plus a single 10,000-token cache write. Every request stays under 272K, so no long-context surcharge applies.

Model	Fresh input	Cached input	Cache write	Output	Total
GPT-6 Astra	$4.00	$1.60	$0.13	$5.00	$10.73
GPT-5.6 Sol (promo)	$1.60	$0.64	$0.05	$2.00	$4.29
GPT-5.6 Sol (standard)	$2.00	$0.80	$0.06	$3.00	$5.86
GPT-5.6 Terra	$0.80	$0.32	$0.03	$1.20	$2.35
GPT-5.6 Luna	$0.08	$0.03	$0.00	$0.12	$0.23

Caching does more work here than any other line. If every call had to resend its 8,000-token prefix uncached, Astra’s bill on this run would rise from $10.73 to roughly $25.00.

The Break-Even That Actually Decides It

Because Astra is uniformly 2.5 times Sol’s promotional rate on every line, it can consume up to 40 percent of Sol’s tokens on the same task and land on the same bill. Use a third of Sol’s tokens and Astra is cheaper; use half and it is more expensive. Against Sol’s standard rate the break-even relaxes to about 55 percent on a cache-heavy mix.

This is where vendor claims and independent measurement diverge, and both are worth knowing. OpenAI reports Terminal-Bench 4.0 at 57.9 percent for Astra against 37.3 percent for Sol at approximately 9 percent lower estimated API cost per task, and says Astra uses about 65 percent fewer output tokens than Claude Opus 5 on Agents’ Last Exam. Artificial Analysis, evaluating independently on 9 September 2026, measured Astra using about one third of Sol’s tokens per task on its Coding Agent Index — yet still found Astra at $7.09 per task at maximum effort, roughly 15 percent more than Sol, for a seven-point higher score. On the Intelligence Index the same evaluation put Astra between $0.82 and $3.26 per task across effort levels.

Both results can hold at once. Token savings are real and large, but savings concentrated on output tokens do not shrink the cached-input and fresh-input lines that dominate a long agent session. Whether Astra is cheaper per task than Sol depends on the mix, and the answer flips between workloads.

The Three Levers That Move Your Bill

Reasoning effort. Reasoning tokens bill as output at $50 per million. Independent measurement found a spread of $0.82 to $3.26 per task between low and maximum effort on identical work — a factor of four. Set effort per call rather than globally.

Retries. An agent that fails and reruns pays for its whole prefix again. This is the mechanism behind OpenAI’s fewer-retries argument, and it is why per-token arithmetic misleads in both directions: a model costing 2.5 times per token but finishing in one pass beats a cheaper model needing three. Measure completed tasks per dollar, not tokens per dollar.

Request size. Chunking a long-horizon task into sub-272K requests is usually cheaper than sending it whole, even though the model is capable of taking it whole. Crossing that line is the easiest way to double a bill without noticing.

Consumer Access and the Free Question

There is no free tier and no free API path. Free and Go ChatGPT users do not get Astra at all. Astra usage is included within existing subscription allowances for Plus, Pro, Business and Enterprise, with additional credits purchasable; Pro, Business and Enterprise plans also receive GPT-6 Astra Pro. Enterprise administrators must enable Astra for their workspace, as access is off by default at launch. Reporting indicates Plus subscribers reach Astra through ChatGPT Work and Codex rather than regular chat, though OpenAI’s own announcement describes broader availability, so this is worth checking against your own account rather than taking on trust.

One capacity note for anyone planning around a subscription: on 10 September 2026, OpenAI paused new sign-ups and upgrades to its $200 ChatGPT Pro tier, citing demand for Astra. Plus, Go, the API, and Business and Enterprise accounts stayed open. Consumer tiers got rationed; API and enterprise access did not. For a wider view of what the major assistants charge, see our comparison of Gemini, ChatGPT, Claude and Grok subscription prices.

What to Watch

One date and one number. The date is 21 November 2026, when Sol’s promotional pricing is no longer guaranteed — if it lapses, every cost-per-task comparison currently in circulation becomes a fifth to a third more favourable to Astra without anything about Astra changing. The number is your own break-even: run one representative long-horizon task on both models, count completed tasks and total spend, and you will have a figure more useful than any published benchmark.

Anyone running a mixed stack should also watch how the cheaper tiers price against Astra for work that does not need frontier reasoning — the gap between a flagship and a fast model is now large enough that routing matters, as the arithmetic in the Grok 4.5 versus Claude Opus cost comparison shows, and as Gemini 3.5 Flash’s rates illustrate at the lower end.

Pricing described here was read from OpenAI’s published documentation on 16 September 2026 and is subject to change. Third-party gateway rates and negotiated enterprise terms may differ. Confirm current rates on OpenAI’s own pricing page before committing to a budget.

If you are interested in this topic, we suggest you check our articles:

Sources: OpenAI, OrcaRouter, LLM Stats, CloudZero, DataCamp

Written by Alius Noreika