OpenAI

GPT-6 Astra

Provider

OpenAI

Status

Current

Context

1,050,000 tok

Price

$10 / $50 /MTok

Knowledge

2026-04-30

GPT-6 Astra is OpenAI’s flagship generational release, which began rolling out on 3 September 2026 — one day after OpenAI rated it the first of its models to cross the Critical cybersecurity threshold in its own Preparedness Framework. It is priced at $10/$50 per million tokens, 2.5x GPT-5.6 Sol’s current rate and identical to Claude Fable 5.1’s.

The independent measurement moved in its first week. Artificial Analysis scored Astra 61 on its Intelligence Index at launch — level with Sol and Grok 4.6, five points under Fable 5.1 — then re-based the index twice (v4.2 on 4 September, v4.3 on 7 September), dropping the saturated GPQA Diamond and adding harder agentic work. On v4.3 Astra and Fable 5.1 tie at 53, above Claude Opus 5 (51) and Fable 5 (50), and Astra gets there at $3.26 per completed task against Fable 5.1’s $7.63 — on about a third of the output tokens. Its Coding Agent Index is 62, level with Fable 5.1 and above Opus 5 (60).

The cyber capabilities that earned the Critical rating — a perfect 100% on ExploitBench, two zero-day vulnerabilities found unaided in evaluation — are not in the box you buy: they are gated to vetted organisations in OpenAI’s Daybreak Blue program. What ships generally is the same model with those capabilities withheld.

Quick specs

CHECK STATUS →

What actually changed

Agentic coding is the real generation jump. On Artificial Analysis’s independent Coding Agent Index, Astra scores 62 — level with Claude Fable 5.1 and above Opus 5 (60) — at about 60% of Fable 5.1’s per-task cost on that index, and on AA’s own run of Terminal-Bench 4.0 it leads outright at 59.1% against Fable 5.1’s 52.0%, Opus 5’s 49.0% and Sol’s 39.9%. AA measures it using one-third the tokens of Sol (max) and one-fifth of Opus 5’s on coding benchmarks, a roughly 70% token-efficiency gain generation-on-generation. OpenAI’s own table agrees on direction: Terminal-Bench 4.0 rises to 57.7% (past Fable 5.1’s 55.8%), AutomationBench more than doubles Sol (41.4% vs 18.1%).

Computer use moves furthest. ScreenSpot-Pro jumps from Sol’s 76.9% to 92.7%, OSWorld 2.0 from 65.7% to 72.6%, and Agents’ Last Exam from 53.6% to 59.3% — all past the Anthropic frontier models on OpenAI’s table.

General intelligence: it depends which index you read. On the launch-week v4.1 index Astra read 61, the same as Sol; on the re-based v4.3 index — which dropped the saturated GPQA Diamond and weights agentic work harder — it reads 53, six points clear of Sol (47) and tied with Fable 5.1. The regressions survive the re-base: Artificial Analysis measures Astra about 45 Elo down on GDPval-AA v2 knowledge work against Sol and about 100 behind Fable 5.1 (1662 to 1764), and below Fable 5.1 on SciCode (56% to 63%) and AA-Omniscience accuracy (81% to 85%). On OpenAI’s own table it trails Fable 5.1 on Humanity’s Last Exam with tools, 57.2% to 65.0%. Astra is an agentic generation that now also ties the composite; it is not a knowledge-work leap.

Hallucination roughly halves. On AA’s hallucination evaluation Astra’s rate falls to 51% against Sol’s 92% — the largest safety-adjacent improvement in the independent data.

The cyber capabilities are the headline OpenAI wrote for itself. Astra is the first OpenAI model rated Critical for cybersecurity under its Preparedness Framework: a perfect score on ExploitBench (turning known vulnerabilities into working exploits), two zero-days found unaided in a separate evaluation, and 88.0% on SRE-Bench against Opus 5’s 12.5%. OpenAI reports the model declines 91.5% of cyber-related jailbreak attempts against Sol’s 59%. Full cyber capability goes only to the Daybreak Blue program.

Benchmark performance

All figures OpenAI-reported and vendor-run from the 3 September launch table, except where marked independent.

Coding and agents

The DeepSWE row deserves its caveat: Astra’s 74.1% leads Gemini 3.8 Flash (73.8%) and Opus 5 (73.7%) by fractions of a point on OpenAI’s own table, and Meta’s launch table the same day claims 75.4% for the preview-only Muse Spark 1.3 max — four vendors’ numbers within 1.7 points, none independently adjudicated.

Computer use and long context

The independent numbers

Artificial Analysis scores Astra (max) at 53 on Intelligence Index v4.3 — tied first of the roughly 200 models it tracks with Claude Fable 5.1, above Opus 5 (51), Fable 5 (50) and Sol (47) — and 62 on the Coding Agent Index, level with Fable 5.1 and above Opus 5 (60). The launch-week reading on the v4.1 index was 61, level with Sol; the re-base (v4.2 on 4 September, v4.3 on 7 September) removed the saturated GPQA Diamond, added AA-Briefcase, GDP.pdf and AutomationBench-AA, and upgraded Terminal-Bench to 4.0 — changes that favour exactly the agentic work Astra is built for. Cost per task on the new suite is $3.26 against Fable 5.1’s $7.63 and Opus 5’s $5.86, on about 27k output tokens per task to Fable 5.1’s 78k; against Sol ($1.99) it is still about 60% dearer per task, because the 2.5x price rise outruns the token savings. Measured throughput is 54 tokens per second with a 346-second time to first token at max effort — slower to start than Fable 5.1 (309 seconds).

What OpenAI did not publish

No SWE-bench figure of any kind — the second frontier launch in a week to skip the industry’s most-watched coding benchmark, after Fable 5.1 did the same. The highest SWE-bench Pro score remains Claude Fable 5’s 80.3%, a June number. Also unpublished at launch: max output tokens and training cutoff; Artificial Analysis has since measured throughput at 54 tokens per second.

The Critical cyber rating, and what it means for buyers

Astra crossed the Critical threshold of OpenAI’s Preparedness Framework on cybersecurity — the first OpenAI model to do so, announced 2 September, the day before rollout began. The evidence OpenAI cites: a perfect score on ExploitBench and two zero-day vulnerabilities discovered without human intervention in a separate evaluation against recently disclosed flaws.

Practically, this splits the product in two:

• The general release ships with those capabilities withheld. OpenAI says its safeguards “sufficiently minimize the risk of severe harm” for release, and reports 91.5% refusal of cyber jailbreak attempts in testing, up from Sol’s 59%.
• The full cyber capability goes only to vetted organisations in the Daybreak Blue program — the same week Google gated Gemini 3.8 Flash Cyber behind its Fairwind Program, and a month after OpenAI gated GPT-5.6-Cyber behind Daybreak Red. Frontier cyber capability is now uniformly access-controlled; the question labs are deciding is who qualifies, not whether to build it.

How to access GPT-6 Astra

Rollout began 3 September 2026 with enterprise customers in the Daybreak programme. ChatGPT Plus, Pro, Business and Enterprise, the API and AWS follow over the coming days, per OpenAI. Fast mode — the same model at higher throughput — is priced at double the standard rate ($20/$100).

Rollout to the paid ChatGPT tiers completed on 5 September, and the API is live (Artificial Analysis has measured its throughput). OpenAI has not published a maximum output length; check OpenAI’s pricing page for the current state before committing a workload.

How GPT-6 Astra compares

vs GPT-5.6 Sol

Six points clear on the re-based intelligence index (53 to 47; the two were level at 61 on the old scale), and a very different profile. Astra is far stronger on agents and computer use (ScreenSpot-Pro 92.7% vs 76.9%, AutomationBench 41.4% vs 18.1%), dramatically better on long context, and hallucinates roughly half as often on AA’s measurement — but it costs 2.5x Sol’s promotional rate and about 60% more per completed task ($3.26 to $1.99), and it regresses on knowledge work (~45 Elo on GDPval-AA v2). Sol stays on sale at $4/$20 to at least 21 November. If your workload is chat, drafting or analysis, Sol is the better buy; Astra earns its premium on agentic and computer-use work.

vs Claude Fable 5.1

Identical pricing ($10/$50), and a tie on the re-based independent index (53 apiece; Fable 5.1 led 66 to 61 on the old scale). Fable 5.1 leads on knowledge work (GDPval-AA v2 1764 vs 1662), SciCode (63% vs 56%), factual accuracy (85% vs 81% on AA-Omniscience) and Humanity’s Last Exam with tools (65.0% vs 57.2% — on OpenAI’s own table). Astra leads on Terminal-Bench 4.0 (59.1% vs 52.0% on AA’s run; 57.7% vs 55.8% on OpenAI’s), AutomationBench-AA (68% vs 59%), Terminal-Bench-Science (64.6% vs 52.6%), FrontierMath Tier 4 (97.6% vs 87.8%) and everything computer-use — and it does it at $3.26 per index task against $7.63. Neither published a SWE-bench figure. The tie is on the composite; Fable 5.1 wins the non-agentic components.

vs Claude Opus 5

Opus 5 at $5/$25 is half the token price, and on the re-based index it is now behind: 51 to Astra’s 53, 60 to 62 on the Coding Agent Index, and $5.86 per index task to Astra’s $3.26. What Opus 5 still has is a published SWE-bench Pro figure (79.2%) where Astra has none, a slightly lower hallucination rate, and a steadier record on following instructions in early reviews. Astra beats it on computer use, long context and every agentic terminal benchmark.

vs Gemini 3.8 Flash

The odd couple of the week: Gemini 3.8 Flash matches Astra to within half a point on DeepSWE (73.8% vs 74.1%, OpenAI’s table) at $0.75/$3.75 — one-thirteenth of Astra’s price. Astra is far more capable on the re-based index (53 vs 41) and far stronger on computer use; the Flash undercuts it catastrophically on cost for agentic coding workloads that fit its profile.

Known limitations

No SWE-bench figure of any kind. The pattern from Fable 5.1’s launch repeats. On coding-weighted comparisons the burden of proof sits on vendor benchmarks alone.

2.5x Sol’s price, about 60% more per task. The re-based index now separates them (53 to 47), but the per-task cost is still about 60% higher at max effort ($3.26 to $1.99).

Knowledge work regresses. About 45 Elo down on GDPval-AA v2 versus Sol and about 100 behind Fable 5.1, with declines on scientific coding and factual accuracy in AA’s suite; independent reviewers also report weaker writing than Sol.

Instruction-following complaints. Early reviewers report Astra explaining a task instead of doing it, and taking liberties with custom instructions in ways they did not see from Sol or Fable (Peter Yang, 9 September).

The headline cyber capability is not purchasable. Critical-rated capabilities live behind Daybreak Blue vetting; the general model is deliberately weaker than the one in the announcement.

Vendor-run headline benchmarks. Every figure except the Artificial Analysis measurements is OpenAI’s own.

Version history

Frequently asked questions

Is GPT-6 Astra the best AI model?

No. On the re-based Artificial Analysis Intelligence Index (v4.3) it scores 53 — level with Claude Fable 5.1 and Gemini 4 Argon, behind Claude Opus 5.5 (58) and Claude Sonnet 5.5 (56), and ahead of Opus 5 (51), Fable 5 (50) and GPT-5.6 Sol (47) — at $3.26 per completed task against Fable 5.1’s $7.63. Fable 5.1 leads it on knowledge work, scientific coding and factual accuracy. It publishes no SWE-bench figure.

How much does GPT-6 Astra cost?

$10 per million input tokens and $50 per million output — 2.5x GPT-5.6 Sol’s current promotional rate and identical to Claude Fable 5.1. Fast mode is double that, at $20/$100. Cache reads carry a 90% discount. Artificial Analysis measures it at $3.26 per completed task on its re-based index — roughly 60% more than Sol ($1.99) despite much better token efficiency, and well under Opus 5 ($5.86) and Fable 5.1 ($7.63).

Why was GPT-6 Astra rated a Critical cyber risk?

It is the first OpenAI model to cross the Critical cybersecurity threshold in OpenAI’s own Preparedness Framework: it scored 100% on ExploitBench, which measures turning known vulnerabilities into working exploits, and found two zero-day vulnerabilities unaided in a separate evaluation. Those capabilities are withheld from the general release and gated to vetted organisations in OpenAI’s Daybreak Blue program.

What is GPT-6 Astra’s SWE-bench score?

OpenAI did not publish one — no SWE-bench Pro or Verified figure of any kind appears in the launch material, making Astra the second frontier launch in a week (after Claude Fable 5.1) to omit it. The nearest published coding-agent evidence is vendor DeepSWE v1.1 (74.1%) and the independent AA Coding Agent Index (62).

What is the context window?

1,050,000 tokens, with a knowledge cutoff of 30 April 2026, per OpenAI’s API docs. OpenAI’s long-context retrieval numbers are strong: 100% on its MRCR v2 8-needle test at 256–512K and 96.3% at 512K–1M, against Sol’s 91.5% and 73.8%.

Should I upgrade from GPT-5.6 Sol?

Only if your workload is agentic. Astra’s gains are concentrated in coding agents, computer use and long context; it regresses on knowledge work (about 45 Elo on GDPval-AA v2), reviewers report weaker writing, and it costs about 60% more per completed task. Sol remains on sale at $4/$20 until at least 21 November 2026.

Last verified 11 September 2026. Benchmark figures are OpenAI-reported from the 3 September 2026 launch table and vendor-run unless marked independent; Intelligence Index (v4.3, re-based 7 September 2026), Coding Agent Index, Terminal-Bench 4.0 (AA’s run), token-efficiency, cost-per-task, throughput and hallucination figures are Artificial Analysis measurements. OpenAI published no SWE-bench figure for this model. Pricing and availability subject to change.