GPT-6 Astra
Last verified September 4, 2026 · OpenAI pricing ↗$10.00/1M input · $50.00/1M output · 1.1M context · OpenAI
Limited access
Not generally available
GPT-6 Astra is not on OpenAI’s open price list. Access is gated to vetted partners or a waitlist, so the rate below is what those customers pay rather than something you can sign up for. It is excluded from the calculator’s default model list and from every ranking.
OpenAI’s own page ↗Count Tokens
—
Tokens
—
Input Cost
—
Output Cost
Estimate Monthly Cost
Monthly Cost Estimator
Quick:
<$0.0001/mo
Pick a preset above or enter custom usage
Alternatives to GPT-6 Astra
Pricing Details
OpenAIOpenAI flagship, announced September 3, 2026. Card: $10.00 input, $1.00 cached input, $12.50 cache write (1.25x uncached input), $50.00 output. Batch and Flex 50% ($5.00/$0.50/$25.00). Fast mode 2x ($20.00/$2.00/$100.00), and Fast is UNAVAILABLE for this model under EU data residency, which OpenAI states on the pricing page. LONG CONTEXT: prompts over 272,000 input tokens are billed at 2x input AND 2x cache rates and 1.5x output ACROSS THE WHOLE REQUEST, not the excess, giving $20.00/$2.00/$75.00 Standard, $10.00/$1.00/$37.50 Batch and Flex, $40.00/$4.00/$150.00 Fast. Same 272,000 threshold and same 2x/1.5x rule as GPT-5.6 Sol. Regional processing (data residency) endpoints carry a 10% uplift, matching Anthropic's 1.1x inference_geo and Google's 10% non-global Vertex premium. EXACTLY 2.500x GPT-5.6 Sol on all ten published lines (noting Sol's $4.00/$20.00 is itself promotional, available 'at least through November 21, 2026', so the ratio would fall to 2.00x input / 1.67x output if Sol reverted to $5.00/$30.00) (input, cached, cache write, output, both Batch/Flex legs, both Fast legs, both long-context legs), so no token mix, cache hit rate or tier changes the ratio: Astra is cheaper per task than Sol only if it spends under 40.0% of Sol's tokens. IDENTICAL to Claude Fable 5 (now legacy) on all four lines Anthropic publishes a comparable figure for: $10.00 base, $12.50 5-minute cache write, $1.00 cache read, $50.00 output, plus both batch legs at $5.00/$25.00. Against Claude Fable 5.1 the ONLY rate-card difference is the cache read, $1.00 vs $0.25 (4.00x, preserved under batch at $0.50 vs $0.125), because Anthropic dropped Fable 5.1 and Mythos 5.1 to a 0.025x multiplier on 2026-09-01 while Astra sits on the 0.100x that eight other models across both price lists share. Second difference is structural, not priced: Fable 5.1 bills its full 1,000,000-token window flat with no long-context tier, so on 400,000-token prompts at an 85% hit rate Astra runs 2.46x Fable 5.1 ($2,180.00 vs $885.00 per 1,000 turns) while at 60,000-token prompts it runs only 1.126x ($341.00 vs $302.75). CONTEXT: 1,050,000 is 922,000 max input plus 128,000 max output, a sum of two ceilings rather than one addressable window; the 272,000 threshold therefore sits at 29.5% of fillable input, so about 70% of the input range bills at double. Rate limits make most of that unreachable: Tier 1 is 500,000 TPM against a 922,000-token max prompt, so one maximum-length request is 84% larger than the whole minute; Tier 2 (1,000,000 TPM) fits exactly one. Free tier not supported. reasoning.effort supports low, medium, high, xhigh, max. Knowledge cutoff April 30, 2026. AVAILABILITY as of 2026-09-04: Trusted Access Program enterprises only; OpenAI says API plus Plus/Pro/Business/Enterprise are coming in the coming days, and names Microsoft Azure and AWS Bedrock as venues, but neither has published a rate (Azure's OpenAI pricing page carries an undated pending-publication banner) and OpenRouter's full 427-model catalogue had no Astra row when pulled 2026-09-04. Marked partner for that reason; flip to generally available once the API opens. OPENAI-REPORTED BENCHMARKS: Terminal-Bench 4.0 57.9 (Sol 37.3, Fable 5.1 55.8), Terminal-Bench Science 0.1 64.6 (Fable 5.1 52.6), BenchCAD with tools 95.9 (Sol 83.3, Fable 5.1 84.3), Agents' Last Exam 59.3 (Opus 5 55.5, Sol 53.6), GPQA Diamond 96.0, OSWorld 2.0 72.6 at about 40 min/task (Sol 65.7 at about 75 min), FrontierMath Tier 4 v2 97.6 (Sol 83.0, Fable 5.1 87.8; the launch prose separately says Astra saturates Tier 4 at 98% without a version qualifier), ARC-AGI-3 99.9 (Sol 7.8), DeepSWE v1.1 74.1 (Sol 72.7), ExploitBench 100 (Sol 78.5, Fable 5.1 70), Exploit Gym 42.4 (Sol 30.3, Fable 5.1 30.4), SRE-Bench 88.0 pass@1 and 99.2 within four (Sol 55.9/68.7, Fable 5.1 12.5), GPQA Diamond Fable 5.1 93.7. The 70.2 on the OSWorld row is Opus 5, not Fable 5.1, and is widely misattributed. The ExploitBench and ExploitGym figures were measured WITHOUT production safeguards; the shipping model refuses those tasks. COST CLAIMS ARE TOKEN CLAIMS: OpenAI quotes per-task savings of 9% vs Sol and 63% vs Fable 5.1 on Terminal-Bench 4.0, 31% vs Fable 5.1 and 27% vs Sol on Terminal-Bench Science, 43% vs Sol and 86% vs Fable 5.1 on BenchCAD, 57% vs Sol on DeepSWE v1.1, 37% vs Sol on GPQA at a lower-cost setting, and about 65% fewer output tokens than Opus 5 on Agents' Last Exam. Since Sol's card is exactly 0.4x and Fable 5.1's is 1.0x on input and output, these invert to implied token ratios of 2.75x, at least 2.70x, at least 1.45x, 3.42x, 4.39x, at least 7.14x, 5.81x and 3.97x respectively, a factor-of-five spread against the same competitor within one post. Tool calling is Responses-API-only. Model ID: gpt-6-astra.
Input / 1M tokens
$10
Output / 1M tokens
$50
Context Window
1.1M
Max Output
128K
Price History
Launched at current price on 2026-09-03. No price changes recorded since.
Frequently Asked Questions
Common questions about GPT-6 Astra pricing and usage