OpenAI launched GPT-6 Astra, its new flagship model, on September 3, 2026. The headline price is $10 per million input tokens and $50 per million output, 2.5x GPT-5.6 Sol's current promotional rate. No introductory discount, no scheduled expiration. This is the standard price.
That headline number is only part of the picture. Astra ships with four different rate tiers depending on context length and processing mode, and which one applies changes your real cost substantially, sometimes by compounding two tiers together.
01The Full Rate Card
| Tier | Input | Cached input | Cache write | Output |
|---|---|---|---|---|
| Standard (≤272K input) | $10.00 | $1.00 | $12.50 | $50.00 |
| Long-context (>272K input) | $20.00 | $2.00 | $25.00 | $75.00 |
| Batch / Flex (standard) | $5.00 | — | — | $25.00 |
| Fast mode (standard) | $20.00 | — | — | $100.00 |
All figures per 1 million tokens. Cross a request past 272K input tokens and the entire request reprices at the long-context tier, not just the tokens above the threshold.
Cached input is the biggest lever most people will overlook: 90% off standard input, at $1 per million versus $10. Any workload that repeats a large system prompt or document context across calls should be routing through caching by default.
See what Astra costs for your actual usage pattern
Model standard, long-context, and cached scenarios side by side02The Compounding Gotcha
Two tiers can stackBatch/Flex and Fast mode aren't fixed discounts off the $10/$50 headline rate, they're multipliers applied to whichever base tier is active. Run Fast mode on a request over 272K input tokens, and you're not paying the standard Fast rate ($20/$100), you're paying 2x the long-context rate: $40 input / $150 output. That's 4x the headline number most coverage quotes, on a request that also crossed the context threshold. Batch on long-context works the same way in the other direction: $10/$37.50, half of the long-context rate rather than half of standard.
Regional data residency adds one more layer: endpoints with data residency requirements carry a flat 10% uplift on any model released after March 5, 2026, which includes Astra. Fast mode specifically isn't available at all on EU data residency endpoints. Neither of these show up in the headline $10/$50 figure most launch coverage led with.
03Why It's Gated
Astra isn't fully open at launch. OpenAI rated it "Critical" for cyber risk under its Preparedness Framework, the first model to hit that tier, and is rolling it out through a staged access program: a small set of trusted organizations first, then ChatGPT Plus, Pro, Business, and Enterprise users over the following days, alongside general API and AWS availability. The public-facing version declines advanced cyber tasks like proof-of-concept exploits; the loosened version goes only to vetted organizations.
Practically: if you're not seeing Astra in your API access yet, that's expected, not a bug.
04On the Benchmark Claims
OpenAI's launch materials lead with a 99.9% score on ARC-AGI-3. Worth knowing before repeating that number: it came from a provider-specific test harness that preserved reasoning state between actions. The standard, neutral harness produced 62.7%, at meaningfully higher compute cost per task. Both numbers are real, they're just not the same test. Treat the 99.9% figure as OpenAI's best case, not an independent result.
05Where This Leaves Sol
OpenAI now runs a two-track lineup within the 5.6 generation: Sol at $4/$20 (promotional, guaranteed at least through November 21) for general frontier work, and Astra at $10/$50 for long-horizon agentic tasks, computer use, and work that specifically benefits from the larger context window and higher reasoning ceiling. The 2.5x gap is deliberate segmentation, not Astra simply replacing Sol. If your workload doesn't need Astra's specific strengths, Sol remains the cheaper frontier option.
Update, Sep 23: the GPT-6 family Astra launched into is now complete. OpenAI has added two lower tiers beneath it, GPT-6 Sol ($2.00/$10.00) and GPT-6 Luna ($0.10/$0.50), confirmed directly against OpenAI's official pricing page. Both scale down from Astra by an exact 1/5th at each step (Astra → Sol → Luna), across input, output, batch, and cache-read alike — a cleaner, more mechanical tiering than the old GPT-5.6 lineup used. Astra itself is unaffected: it remains the top of the stack at $10/$50, and this pricing article's rate card is unchanged. GPT-6 Sol and Luna are covered in their own article.
Update, Sep 30: GPT-6.1 Sol launched September 29 as a new sibling to GPT-6 Sol, sitting in the same tier below Astra — Astra's own rate card is still unaffected. A pricing report circulating the same week also claimed Astra's rate had been cut to $5.00/$25.00 with a $0.50 cache-read; that claim was checked directly against OpenAI's official pricing page and found to be false. Astra remains at $10.00/$50.00, unchanged since launch.
Compare Astra against Sol, Claude, and Gemini
Run your real token counts across every flagship model