OpenAI
GPT-6 Astra
OpenAI calls GPT-6 Astra its most capable model yet and the first to cross its own "Critical" cybersecurity threshold. Of the eight tracked benchmarks with an independent score, Humanity's Last Exam separates it from GPT-5.6 Sol on a clean same-tier comparison, by 5.2 points, and Agents' Last Exam by 4.6 at matched Max effort (a tie at XHigh); ARC-AGI-2 and AA-AnalystAgent land inside the noise band.
GPT-6 Astra benchmarks and pricing, every number sourced: 9 tracked GPT-6 Astra benchmark scores (8 independently run, 1 still resting on a vendor’s own claim), priced at $10.00 per million input tokens and $50.00 per million output.
- Released
- 2026-09-03
- License
- proprietary
- Context window
- 1M tokens
- Knowledge cutoff
- 2026-04-30
- Verified
- sources checked 2026-09-04–2026-10-08
- Parameters
- Not disclosed
- Architecture
- Not disclosed
GPT-6 Astra’s verified record
GPT-6 Astra’s most-compared rival is Kimi K3: 5 leads, 1 tie, and 2 not callable across their 8 shared comparisons. GPT-6 Astra is priced at $10.00/$50.00 per 1M tokens (in/out) vs Kimi K3’s $3.00/$15.00.
Against the 205 head-to-head comparisons GPT-6 Astra shares with other tracked models:
A gap counts for GPT-6 Astra only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — GPT-6 Astra trails on 3 of them.
HLE · no tools Reasoning
±2 is noiseTerminal-Bench 4.0 · max Terminal ops
±12.4 is noiseAnalystAgent Spreadsheet & document analysis
±11.2 is noiseARC-AGI-2 · max Compositional visual reasoning
±9.2 is noiseDeepSWE Long-horizon coding
±9.5 is noiseAgents' Last Exam Professional work
±3.2 is noiseTerminal-Bench 2.1 Terminal ops
±10.6 is noiseNo verdict for GPT-6 Astra anywhere on GPQA Diamond ( saturated); OSWorld 2.0 · v2026 08 08 offline ( nothing independently confirmed on both sides).
GPT-6 Astra API pricing
$10.00 in / $50.00 out per 1M tokens — official pricing source
What GPT-6 Astra costs per job
| Workload | Tokens in / out | Cost |
|---|---|---|
| One long chat turn | 100K / 10K | $1.50 |
| A codebase review | 1,000K / 100K | $15.00 |
| A day of agent work | 10,000K / 1,000K | $150.00 |
Computed from GPT-6 Astra’s list rates above — cache discounts and batch tiers are not applied.
GPT-6 Astra is one of 6 OpenAI models tracked on this site, at these official list prices.
| Model | In / 1M | Out / 1M |
|---|---|---|
| GPT-6.1 Sol | $2.00 | $10.00 |
| GPT-6 Luna | $0.10 | $0.50 |
| GPT-6 Sol(previous version) | $2.00 | $10.00 |
| GPT-6 Astra | $10.00 | $50.00 |
| GPT-5.6 Luna(previous version) | $0.20 | $1.20 |
| GPT-5.6 Sol(previous version) | $4.00 | $20.00 |
GPT-6 Astra benchmark scores
| Benchmark | Score |
|---|---|
| HLE(no tools)[1] Reasoning · ±2 is noise | |
| GPQA Diamondsaturated[2] Expert science Q&A — not ranked at any gap size | |
| Terminal-Bench 2.1[3] Terminal ops · ±10.6 is noise | |
| DeepSWE[4] Long-horizon coding · ±9.5 is noise | |
| ARC-AGI-2(max)[5] Compositional visual reasoning · ±9.2 is noise | |
| AnalystAgent[6] Spreadsheet & document analysis · ±11.2 is noise | |
| Agents' Last Exam[7] Professional work · ±3.2 is noise | |
| Terminal-Bench 4.0(max)[8] Terminal ops · ±12.4 is noise | |
| OSWorld 2.0(v2026 08_08_offline)[9] Long-horizon computer use · ±9.7 is noise |
Who ran these numbers: 8 of 9 independent — artificialanalysis.ai (3), tbench.ai (1), deepswe.datacurve.ai (1), arcprize.org (1), snorkel.ai (1), vals.ai (1); vendor self-reported (1).
- HLE: AA's own run of GPT-6 Astra at max effort, text-only subset. AA also lists xhigh (54.59%), high, medium, low, and non-reasoning tiers for this model.
- GPQA Diamond: AA's own run of GPT-6 Astra at max effort. AA also lists xhigh (96.26% — its single highest GPQA Diamond score to date, a tier-naming inversion), high (94.95%), and medium (93.94%). This benchmark is saturated on this site; every comparison on it is tainted regardless of gap size.
- Terminal-Bench 2.1: tbench.ai's own canonical board, "GPT-6 Astra (high)" run via the Codex agent, rank 1 of 18 (±1.8%, 95% CI). Artificial Analysis has since added its own Terminal-Bench 2.1 runs for this model (88.4% at max per its model page, checked 2026-09-30); the stored row stays tbench.ai's, a different harness. GPT-5.6 Sol's own tracked score for this benchmark comes from Artificial Analysis, not tbench.ai — a different harness, recorded as such. Cross-check (2026-10-01): vals.ai's archived Terminus 2 table lists it at 87.27.
- DeepSWE: 74%±3% at xhigh effort (the model's own best config on this board's v1.1 table, $6.52/task, 30k output tokens, 29 steps), tied for the board's top spot with Gemini 3.8 Flash (high) and Claude Opus 5 (max). GPT-5.6 Sol's own best config on the same board is max effort, 73%±3% — a cross-tier comparison by construction, though checking Astra's own max-tier row (73%±1%, per the board's full effort-level view) shows the two are functionally tied at matched tiers too.
- ARC-AGI-2: GPT-6 Astra's official ARC-AGI-2 leaderboard row (Max tier), 95.0% — 2.5 points above GPT-5.6 Sol's own tracked 92.5% at the same tier, a gap this benchmark's 9.2-point noise band swallows as a tie.
- AnalystAgent: AA's own run, board row 'GPT-6 Astra (max)'; the board rounds to one decimal. pass@1 60.0, pass@5 67.5.
- Agents' Last Exam: Overall pass rate 34.2 (Codex, Max, the board's best effort for this model; score 59.3; est. cost $1,100). Per-effort results: Max 34.2, XHigh 32.2, Medium 32.2, High 31.6, Low 29.6.
- Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 59.596 at effort max. Artificial Analysis same-conditions run (Max): 59.09. Official board (grant-funded, Codex, max effort, 5-trial pass@1): 58.18 ± 2.79 CI — 1.4 points below this run. The official board also lists this model at xhigh/high 57.88, medium 54.24, and low 50.61, all under Codex.
- OSWorld 2.0: OpenAI launch chart (2026-09-03), row labeled "OSWorld 2.0 (v2026.08.08, offline set, partial score)". No effort tier is attached to this figure; the surrounding prose frames it as a latency simulation at roughly 40 minutes per task, against GPT-5.6 Sol's 65.7 at roughly 75 minutes. OpenAI's later DevDay chart (2026-09-29) shows Astra max effort at 73.49 on the same slice — a re-measurement OpenAI does not reconcile with the 72.6. Footnote on the same page: Claude figures quoted there use the official settings, not the Fable 5.1 System Card's modified tasks and grading. Google's Argon table quotes the same 72.6 from OpenAI's blog. Verified via reader-rendered text and Google's independent republication.
Notes on the record
Launched 2026-09-03 (OpenAI announcement and System Card); OpenAI calls it "the most capable model we have ever broadly deployed." Not marked as superseding GPT-5.6 Sol, Terra or Luna: no OpenAI model or deprecation page uses successor language and all three stay on sale.
Pricing $10 in / $50 out per 1M tokens up to 272K input tokens — the same threshold Sol and Luna carry — then $20/$75 for the whole request above it. Context window 1,050,000; max output 128,000; knowledge cutoff 2026-04-30 (OpenAI model page). No parameter count disclosed.
Eight of fourteen boards have an independent score: HLE 54.68 (max, AA), GPQA Diamond 96.06 (max, AA; saturated), Terminal-Bench 2.1 87.4 (high, tbench.ai, rank 1 of 18), DeepSWE 74 (xhigh, tied for first), ARC-AGI-2 95.0 (Max, arcprize.org), AA-AnalystAgent 51.25 (max, AA, 2026-09-29), Agents' Last Exam 34.2 (Codex, Snorkel, 2026-10-01), Terminal-Bench 4.0 59.6 (vals.ai, 2026-10-08). LiveCodeBench, SWE-bench Verified, Toolathlon-Verified, LiveBench and HMMT had no Astra row as of 2026-09-04. OSWorld 2.0's row is OpenAI's self-reported 72.6 (2026-10-08). Against Sol, three clean comparisons (same tier, same harness): HLE, where Astra leads by 5.19 against a 2.0 band — a real gap; ARC-AGI-2 (2.5-point lead, 9.2 band) and AA-AnalystAgent (3.75-point lead, 11.2 band) — ties. Agents' Last Exam: real at matched Max (34.2 vs 29.6), a tie at XHigh; DeepSWE ties either way. Terminal-Bench 2.1 can't be compared: different harnesses (tbench.ai vs AA).
OpenAI's System Card says Astra is the first model to cross the "Critical" cybersecurity threshold of its Preparedness Framework: 100% on its internal ExploitBench (Sol 78.5%), two previously unknown zero-days, and a working browser sandbox escape. Proof-of-concept exploit requests are refused by default (91.5% on OpenAI's cyber-jailbreak set; Sol 59%); OpenAI also reports Astra's monitorability is lower than Sol's. The exploit-unlocked "Daybreak Red" tier runs on a separate model, GPT-5.6-Cyber; Astra's own reduced-refusal tier isn't broadly available yet (OpenAI Help Center). In ChatGPT the model is listed as "GPT-6 Pro."
Compare with
FAQ
Is GPT-6 Astra actually better than GPT-5.6 Sol?
Three, yes. Humanity's Last Exam: Astra's 54.68% beats Sol's 49.49% by 5.19 — same tier, same evaluator, real past the 2-point band. Terminal-Bench 4.0 (vals, 2026-10-08): 59.6 vs 37.88, matched max — real past the 12.4 band. ARC-AGI-2 and AA-AnalystAgent are the other clean, matched comparisons: Astra leads by 2.5 (inside the 9.2-point band) and by 3.75 (inside the 11.2-point band) — both ties. Agents' Last Exam is real at matched Max (34.2 vs 29.6) but a tie at XHigh. DeepSWE can only be compared at mismatched tiers (and ties even at matched ones); Terminal-Bench 2.1 has two different evaluators behind the two scores; GPQA Diamond favors Astra but is saturated, so the gap doesn't count.
Why did OpenAI restrict GPT-6 Astra's cybersecurity capabilities?
Because OpenAI's own System Card says Astra is the first model to cross the "Critical" threshold for cybersecurity under its Preparedness Framework — evidence includes a 100% score on OpenAI's internal ExploitBench (versus 78.5% for GPT-5.6 Sol), including two previously-unknown zero-day vulnerabilities the model found and that have since been reported to maintainers. By default, Astra refuses advanced cybersecurity tasks like building proof-of-concept exploits (a 91.5% refusal rate on OpenAI's own cyber-jailbreak eval, versus 59% for Sol), and a production monitoring layer watches tool-using sessions in ChatGPT, Codex, and the API. OpenAI also discloses, without downplaying it, that Astra's own monitorability — how reliably its reasoning can be watched for deception — has gone down compared to Sol.
Does GPT-6 Astra replace GPT-5.6 Sol?
No — not by OpenAI's own materials. None of its model, pricing or deprecation pages calls Astra a successor to Sol, Terra or Luna, and all three stay on sale at unchanged prices with no retirement date — unlike Claude Fable 5 to 5.1, where the vendor's own table said "successor to." OpenAI's launch page does benchmark Astra only against Sol, and its "choosing a model" guidance now leads with Astra and omits Sol entirely — informal signals that Astra has taken the flagship slot. OpenAI publishes nothing about the older Sol models, so this site says only what it can check: GPT-6 Sol is a previous version because GPT-6.1 Sol shipped.
How much does GPT-6 Astra cost, and does the price change for long prompts?
$10 per million input tokens and $50 per million output tokens for prompts up to 272,000 tokens — the same threshold GPT-5.6 Sol and Luna already carry on this site. Cross that line and the entire request, not just the excess, bills at double the input rate and 1.5 times the output rate ($20/$75 per million tokens), per OpenAI's own pricing page.
Has GPT-6 Astra been independently benchmarked?
Partially — eight of the fourteen benchmarks this site tracks carry an independently-run score: Humanity's Last Exam, GPQA Diamond, Terminal-Bench 2.1, DeepSWE, and ARC-AGI-2 within a day of the 2026-09-03 launch, plus AA-AnalystAgent (51.25 pass^5, Artificial Analysis, max effort), added 2026-09-29, and Agents' Last Exam (34.2 overall pass rate, Snorkel), added 2026-10-01, plus vals.ai's Terminal-Bench 4.0 run (59.6, max effort), added 2026-10-08. LiveCodeBench, SWE-bench Verified, Toolathlon-Verified, LiveBench, and HMMT Feb 2026 had not added the model as of 2026-09-04, checked live on each. None of the eight independently-run scores come from OpenAI's own claimed numbers — the one vendor-claimed row on this page is OSWorld 2.0's 72.6 (offline slice), OpenAI's own number.
Is GPT-6 Pro the same as GPT-6 Astra?
Yes — same underlying model, different name. In ChatGPT's model picker Astra does not appear under its own name; OpenAI's Help Center lists it as "GPT-6 Pro," available on the Pro ($100 and $200), Business, and Enterprise plans, with its own usage caps kept separate from "GPT-5.6 Sol Pro." The API name is gpt-6-astra. One caveat this site can't resolve: OpenAI's own announcement says Astra will reach all ChatGPT Plus users "over the coming days," while its Help Center states GPT-6 Pro is not included with Plus in Chat — two OpenAI pages, published the same window, disagreeing on Plus access.
Further reading
- GPQA Diamond leaderboard 2026 — GPT-6 Astra is one of the 23 models it compares.
- Models with 10M token context windows 2026 — GPT-6 Astra is one of the 40 models it compares.
Benchmark guides
- Humanity's Last Exam — what GPT-6 Astra’s reasoning score on it does and doesn’t prove.