OPENAI FLAGSHIPGPT-6 FAMILY1.05M CONTEXT2.5x SOL, LINE FOR LINE272K PRICE CLIFF
GPT-6 Astra API Pricing
GPT-6 Astra is OpenAI’s new top of stack, announced September 3, 2026 and described in the docs as our most capable model, built for the hardest end-to-end work. Standard short-context rates are $10/M input, $1/M cached input, $12.50/M cache writes and $50/M output — an effective $6.82/M on this site’s agentic blend. The striking part is the ratio: every one of the thirteen rates OpenAI publishes for Astra is exactly 2.5x the matching rate for GPT-5.6 Sol — standard, long context, cached, cache writes, Batch, Flex and Fast mode alike. Read the multiple with care, though: Sol’s card is itself promotional at least through November 21, 2026, so the clean 2.5x is measured against a discounted baseline. Astra launched gated to OpenAI’s Trusted Access enterprises; that door opened within three days — as of our September 6 check the model page publishes ordinary rate limits for usage tiers 1 through 5, and the docs now tell any developer unsure where to start to reach for Astra first.
Input - per 1M tokens
$10.00/M
Short context up to 272K input 2.5x Sol
Output - per 1M tokens
$50.00/M
Long context output is $75/M 2.5x Sol
Cached input - 90% off
$1.00/M
Cache writes cost $12.50/M -90%
Effective - agentic blend
$6.82/M
92/8 split - 82% cache
§ 01 / TERMINAL
Run the numbers.
Rates shown are standard short-context. Anything over 272K input tokens reprices the whole request, not just the overflow — set the input slider past that mark and the figures step up accordingly.
$ /mo
Workload split
Prompt cache hit rate
Tokens you can process
—
Words equivalent (English)
—
Effective rate
—
§ 02 / SCENARIOS
Real-world presets.
END-TO-END TASK
Repo-wide migration run
$0.868/task
RESEARCH
Long-document research pass
$0.953/pass
COMPUTER USE
Browser task loop
$0.155/loop
DOC CREATION
Report section draft
$0.075/draft
§ 03 / TOKENIZER
Paste text. See tokens. See cost.
Exact · o200k_base Auto-counts as you type
Counted with the o200k_base BPE encoding — the same tokenizer the API uses — entirely in your browser. The encoder loads on your first keystroke (~1 MB, one time); your text never leaves this page.
Characters 1,124
Words 188
Tokens (exact) 266 tokens
Cost as input · uncached $0.00266 USD
Cost as output · uncached $0.0133 USD
Cost as cached input $0.00027 USD
| Model | Input /M | Output /M | Effective blended | Context | Best for |
|---|---|---|---|---|---|
| GPT-6 Astra Current | $10.00 cache $1.00 | $50.00 | $6.82 agentic 92/8 | 1.05M | Hardest end-to-end work, and now generally available |
| GPT-5.6 Sol | $4.00 cache $0.40 | $20.00 | $2.73 the card Astra multiplies by 2.5 | 1.05M | The frontier tier anyone can call today |
| GPT-5.5 | $5.00 cache $0.50 | $30.00 | $3.61 no cache-write rate published | 1M | Broadly available OpenAI flagship |
| GPT-5.6 Terra | $2.00 cache $0.20 | $12.00 | $1.52 same blend | 1.05M | Production agents that do not need the top tier |
| GPT-5.6 Luna | $0.20 cache $0.02 | $1.20 | $0.152 same blend | 1.05M | High-volume routing and extraction |
| Claude Mythos 5.1 | $10.00 cache $0.25 | $50.00 | $6.26 identical sticker, cheaper cache read | 1M | Same $10/$50 sticker, two days older, still gated |
| Claude Fable 5.1 | $10.00 cache $0.25 | $50.00 | $6.26 same card as Mythos, no gate | 1M | The $10/$50 frontier model anyone can call today |
Compare the card
Frequently asked.
What the card actually says, why every line of it is 2.5x Sol, and what changed in the three days between launch day and general access.
Q · 01 How much does GPT-6 Astra cost? +
OpenAI lists
$10/M input, $1/M cached input, $12.50/M cache writes and $50/M output on the standard service tier. Under this site’s 92/8 agentic blend at an 82% cache-hit rate that works out to $6.82/M effective. Batch and Flex are half of standard at $5/$25; Fast mode, the tier renamed from Priority on July 30, 2026, is double at $20/$100.Q · 02 Is it really exactly 2.5x GPT-5.6 Sol? +
Yes, on every published rate. Input
$4 to $10, cached $0.40 to $1, cache writes $5 to $12.50, output $20 to $50, and the same multiple again across long context, Batch, Flex and Fast mode — thirteen rates, one ratio. One caveat worth carrying: Sol’s own card is promotional at least through November 21, 2026. If Sol returns to its pre-August $5/$30, the multiple becomes 2x on input and about 1.7x on output without Astra changing at all.Q · 03 Can I actually use it? +
Yes — though that was not true when Astra shipped. OpenAI launched it on September 3 to enterprises in our Trusted Access Program, promising API and Plus, Pro, Business and Enterprise access in the coming days, and delivered inside three. The clearest single tell is the model page’s own rate-limit table, which now runs from Tier 1 upward: tier 1 is what any account reaches after spending $5, so an approval-only model would have no such ladder. The docs also now open by telling a developer unsure where to start to use Astra. OpenAI published no announcement of the opening, so treat September 6 as the date we verified it rather than the day the gate lifted. One route is still gated, and it is not OpenAI’s: Microsoft began reselling Astra on launch day through the Foundry Limited Access Program, with a rate card matching OpenAI’s cell for cell.
Q · 04 What happens above 272K input tokens? +
The entire request reprices, not the overflow:
2x the input and cache rates and 1.5x the output rate, giving $20/M input, $2/M cached, $25/M cache writes and $75/M output. A 300,000-in, 12,000-out job costs about $2.74 at long-context rates against $1.52 if the same tokens had fitted under the threshold — 1.8x for crossing a line by 28,000 tokens. Trimming a prompt back under 272K is the single highest-leverage optimisation on this model.Q · 05 How does it compare with Claude Mythos 5.1? +
Claude Mythos 5.1 arrived two days earlier with the identical sticker —
$10/M in, $50/M out — and launched behind the same shape of gate, Anthropic’s Glasswing programme in place of OpenAI’s Trusted Access. Two things separate them now. Access: Astra’s door opened within three days, while Mythos 5.1 still has no row in Anthropic’s model catalogue and no published self-serve API alias, checked September 6. Price: the identical sticker still produces different bills, because the cache read differs — Anthropic charges $0.25/M for a cache hit against OpenAI’s $1/M, so on the same blend Mythos lands at $6.26/M and Astra at $6.82/M, about 9% apart. Sticker parity is not price parity once caching is in the mix.Q · 06 Does Astra really score 99.9% on ARC-AGI-3? +
It depends who runs it. ARC Prize published two numbers for the same model on the same ARC-AGI-3 Semi-Private set:
62.7% at $26,098 under its own Standard harness at max reasoning effort, and 99.9% at $18,817 under OpenAI’s Provider Adapter at high effort. OpenAI’s launch copy quotes the 99.9%. Both are real measurements — of different things, because the harness decides what the model sees and how it is allowed to act, which makes it part of the score rather than a detail beneath it. Note which way the money runs: the configuration that scored 37 points higher also cost 28% less, the reverse of the usual pay-more-for-effort trade. Source: ARC Prize, September 3, 2026.Q · 07 Is Astra the model behind the Daybreak aliases? +
Not yet, despite the cyber framing of the rollout. OpenAI’s pricing page still resolves
gpt-daybreak-blue-latest to gpt-5.6-sol and gpt-daybreak-red-latest to gpt-5.6-cyber, and says the aliases will be repointed as new models are released through the Daybreak program. So code that pins a Daybreak alias is not getting Astra today, and will silently start getting it — at 2.5x the rate — whenever OpenAI moves the pointer.Q · 08 What are the specs? +
Text and image in, text out; a 1,050,000-token context window with a 922,000 maximum input and 128,000 maximum output; knowledge cutoff April 30, 2026; reasoning tokens with effort levels
low, medium, high, xhigh and max. Chat Completions, Responses and Batch are supported; Realtime, fine-tuning and the audio endpoints are not. Data-residency endpoints carry the usual 10% uplift, and Fast mode is unavailable for Astra with EU data residency.