OpenAI's first GPT-6 model and its answer to Claude Fable 5.1, two days later and at the same price. The new ceiling on computer use, agentic coding and frontier math. Also searched as ChatGPT 6 and GPT-6.
Verified September 4, 2026.Pricing, availability and every benchmark on this page come from OpenAI's launch post; context, max output and cutoff come from OpenAI's model documentation. Where OpenAI published no number, the table says so. Nothing here is estimated.
Choose your main workload, then estimate the monthly API premium.
Its strongest advantage is reliable, long-horizon work across browsers, terminals, and professional software.
List price only; cached input, Batch API, tools, and provider pricing can change the total.
import OpenAI from "openai";
const openai = new OpenAI();
const response = await openai.responses.create({
model: "gpt-6-astra",
reasoning: { effort: "high" },
input: "Analyze this task and produce a production-ready result."
});
console.log(response.output_text);The confirmed changes over GPT-5.6 Sol, in OpenAI's own numbers.
72.6% on OSWorld 2.0 in roughly 40 minutes per task, against GPT-5.6 Sol's 65.7% at roughly 75 — about 47% less time. 92.7% on ScreenSpot-Pro. OpenAI calls it the world's best computer-use model, and the numbers back that up.
57.9% on Terminal-Bench 4.0 against Claude Fable 5.1's 55.8%, Opus 5's 52.3% and Sol's 37.3%. In Codex, Astra can keep notes across context windows instead of compacting, and earlier windows stay searchable.
97.6% on FrontierMath Tier 4 (v2) against Fable 5.1's 87.8% and Sol's 83.0%, with two new results on prime gaps published alongside. 96.0% on GPQA Diamond. 64.6% on Terminal-Bench Science, up from Fable 5.1's 52.6%.
100% on ExploitBench and 88.0% on SRE-Bench. Astra meets the Critical cybersecurity threshold under OpenAI's Preparedness Framework, so advanced offensive tasks are refused outside the OpenAI Daybreak program.
2.4% misaligned outcomes on OpenAI's computer-use safety benchmark against Sol's 22.0%; 0% scope creep on impossible cyber tasks against 48%; a third as many false claims about its own capabilities.
Astra's written reasoning is harder to monitor than Sol's, per the system card. Misalignment monitoring runs in production and can pause or stop legitimate agent work, including in the API.
OpenAI's published launch table, reproduced. Rival figures are OpenAI's comparisons, except the SWE-bench row, which is from our catalog.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol | Claude Fable 5.1 | Claude Opus 5 |
|---|---|---|---|---|
| OSWorld 2.0 (offline set) | 72.6% | 65.7% | — | 70.2% |
| ScreenSpot-Pro (no tools) | 92.7% | 76.9% | — | — |
| Terminal-Bench 4.0 | 57.9% | 37.3% | 55.8% | 52.3% |
| DeepSWE v1.1 | 74.1% | 72.7% | 67.4% | 73.7% |
| AutomationBench | 41.4% | 18.1% | 31.4% | 26.9% |
| Terminal-Bench Science 0.1 | 64.6% | 22.4% | 52.6% | 30.0% |
| FrontierMath Tier 4 (v2) | 97.6% | 83.0% | 87.8% | 73.2% |
| GPQA Diamond | 96.0% | 94.6% | 93.7% | 93.7% |
| Humanity's Last Exam (with tools) Fable 5.1 leads | 57.2% | — | 65.0% | 63.6% |
| ARC-AGI-2 | 95.0% | 92.5% | 90.0% | 90.4% |
| ARC-AGI-3 OpenAI harness; ARC Prize's stateless runs score far lower | 99.9% | 7.8% | — | 30.2% |
| ExploitBench | 100.0% | 78.5% | — | 70% |
| Artificial Analysis Intelligence Index v4.1.1 Fable 5.1 leads | 61.2 | 60.9 | 65.7 | 63.1 |
| SWE-bench Verified / Pro | not published | 96.2% / 64.6% | not published | ~96% / 79.2% |
Sol figures refer to the API version. OSWorld 2.0 is the v2026.08.08 offline set with partial scoring. ARC-AGI-3 used OpenAI's responses-API harness; ARC Prize's standard stateless runs of the same model score far lower.
Published by OpenAI at launch. Per 1M tokens, input / output.
A representative month of 10M input and 2M output tokens costs about $200 on Astra at list price, against $80 on GPT-5.6 Sol. In ChatGPT, Astra usage sits inside the existing Plus, Pro, Business and Enterprise allowances, with credits available for more. See ChatGPT Plus limits and the live API pricing table.
Rollout as OpenAI described it on launch day.
Agent builders: Astra ships with production misalignment monitoring that can pause a task in ChatGPT or Codex and stop it outright in the API. Advanced offensive-security work is refused unless you are in OpenAI Daybreak. Plan for interruptions on long autonomous runs until OpenAI tunes the thresholds it says it is still calibrating.
Humanity's Last Exam with tools: 57.2% against Claude Fable 5.1's 65.0% and Opus 5's 63.6%.
Artificial Analysis Intelligence Index v4.1.1: 61.2 against Fable 5.1's 65.7 and Opus 5's 63.1.
No SWE-bench Verified or Pro score published, so the most-cited coding benchmark has no Astra entry.
2.5 times GPT-5.6 Sol's current price per token, and Fast mode doubles Astra's price again.
OpenAI's own finding that Astra's written reasoning is harder to monitor than Sol's.
UseRightAI separates vendor claims from independent comparisons. These are the official pages used for launch facts and current specifications.
GPT-6 Astra is OpenAI's frontier model released on September 3, 2026 — the first model in the GPT-6 generation. OpenAI describes it as its most intelligent and aligned model, with state-of-the-art results on computer use, browsing, software engineering, cybersecurity, science and professional work. It arrived two days after Anthropic's Claude Fable 5.1 and is priced identically at $10 per million input tokens and $50 per million output tokens.
OpenAI announced and began rolling out GPT-6 Astra on Thursday, September 3, 2026. There is no separate product called ChatGPT 6: Astra is the model, and it is becoming available inside ChatGPT for Plus, Pro, Business and Enterprise subscribers over the days after launch. Searches for ChatGPT 6 and GPT-6 both refer to this release.
On the OpenAI API, standard pricing is $10 per million input tokens and $50 per million output tokens, with separate rates for cache reads and writes. Fast mode delivers up to 2x the speed at 2x the price, so about $20/$100. In ChatGPT, Astra is included in Plus, Pro, Business and Enterprise allowances, with additional usage purchasable as credits. That is 2.5 times GPT-5.6 Sol's current $4/$20 price and exactly Claude Fable 5.1's $10/$50.
OpenAI's model documentation lists a 1,050,000-token context window, 128,000 maximum output tokens and a knowledge cutoff of April 30, 2026, with reasoning-effort settings up to 'max'. On OpenAI's MRCR long-context test, Astra scores 96.3% on the 8-needle 512K–1M range against Sol's 73.8%.
On OpenAI's published table Astra wins the agentic and math rows: Terminal-Bench 4.0 (57.9% vs 55.8%), Terminal-Bench Science (64.6% vs 52.6%), FrontierMath Tier 4 (97.6% vs 87.8%), GPQA Diamond (96.0% vs 93.7%) and AutomationBench (41.4% vs 31.4%). Fable 5.1 keeps Humanity's Last Exam with tools (65.0% vs 57.2%) and the Artificial Analysis Intelligence Index (65.7 vs 61.2). Both cost $10/$50. Neither has a published SWE-bench figure. For computer use and math, Astra; for broad reasoning under the same budget, it is genuinely close.
Astra beats Sol on every benchmark OpenAI published, by 20 points on Terminal-Bench 4.0 (57.9% vs 37.3%), by 7 points on OSWorld 2.0 while taking about half the time, and by nearly 15 points on FrontierMath Tier 4. It costs 2.5 times as much per token: $10/$50 against Sol's current $4/$20. Switch for computer-use agents, hard math and long autonomous coding runs. For high-volume everyday coding, Sol remains the lower-cost option.
Astra scored 100% on ExploitBench and 88.0% on SRE-Bench, and in expert testing without safeguards it used previously unknown vulnerabilities to achieve code execution in hardened browsers. That meets the Critical cybersecurity threshold under OpenAI's Preparedness Framework. The public model does secure code review and patching but refuses to build proof-of-concept exploits; less restrictive safeguards for defensive work come through the OpenAI Daybreak program.
It trails Claude Fable 5.1 on Humanity's Last Exam and on the Artificial Analysis index, it ships with no SWE-bench Verified or Pro score, and it is expensive. OpenAI also reports that Astra's written reasoning is harder to monitor than Sol's, and that its production misalignment monitoring can pause or stop legitimate work, including agent runs in the API.