GPT-6 Astra Deep-Dive
A generational leap — or a carefully staged narrative?

Three findings that matter

• ① The genuine leap is in delivering work and in security — not in test-taking: research workflows (Terminal-Bench Science) are up 188.4% on the prior Sol generation, office automation (AutomationBench) up 128.7%, and average time per OSWorld task falls from roughly 75 minutes to 40 (−47%). These are structural gaps.
• ② The most dazzling scores carry clear scaffolding gains, and coding is not a rout: on ARC-AGI-3 Astra scores 99.9% under OpenAI's own Provider Adapter but just 62.7% under the ARC Prize neutral harness; on long-horizon coding (DeepSWE), Meta's Muse Spark 1.3 (75.4%) actually edges past Astra (74.1%), and on Terminal-Bench it leads Claude Fable 5.1 by only ~2.1 points — inside the same confidence band.
• ③ Cybersecurity is both the hardest capability and the one to watch most carefully: a 39% success rate on high-severity V8 bugs newly disclosed in Jun–Aug 2026 (Sol 5.5%), and the evaluation found and exploited two previously unknown zero-days. That is precisely why full offensive capability is withheld and released only through controlled, gated channels — for the first time, "the stronger the capability, the narrower the distribution" becomes a product design.

The one-line takeaway

If your core value is end-to-end completion of complex computer tasks, research pipelines or security work, Astra is today's only model that has pulled clearly ahead. If you are writing code, answering questions or optimizing cost per token, its edge over Fable 5.1, Muse Spark and Gemini Flash sits within the margin of error — at 8–12× the price. Decide by scenario in Section 08; do not be led by a single headline score.

What happened

On September 3, 2026 local time, OpenAI released GPT-6 Astra ("Astra," Latin for star), billing it as "the most intelligent and most aligned model in the world today," spanning computer use, web browsing, software engineering, cybersecurity, scientific research and professional work. Note the sequencing: on September 1 OpenAI first confirmed that Astra had crossed the "Critical" cybersecurity threshold in its internal Preparedness Framework and would restrict that capability; the public launch came only on September 3 — the safety classification preceded the product reveal.

Two signals matter more than parameter counts. First, president Greg Brockman closed the launch with "Welcome to the AGI era," while conceding AGI is "a fuzzy concept without a clear boundary," no longer tied to the trigger clauses in the old Microsoft contract but rather "a mission-level, spiritual notion." Second, Astra is OpenAI's first publicly deployed model to reach the Critical cyber threshold — a backhanded distinction, meaning it can autonomously discover unknown vulnerabilities and build exploit chains with minimal human help.

Four launches in one week: frontier labs enter the "scenario split" phase

• Anthropic: shipped Claude Fable 5.1 (its Mythos/Fable line faces U.S. government deployment curbs over safety concerns).
• Google: released Gemini 3.8 Flash, doubling down on multimodal reach, distribution and low-price high-throughput.
• Meta: released Muse Spark 1.3, betting on extreme low cost and open weights.
• OpenAI: released Astra, betting on Computer Use and high-value long tasks.

Three questions this report answers

• Which advantages are structural, and which are illusions created by how a score was produced? (Sections 05–06)
• Does the high-price strategy of "selling task cost, not tokens" hold up financially? (Sections 07–08)
• What does a Critical safety rating and the AGI narrative mean for industry rules and procurement? (Sections 08–09)

Source tiers

Three "protocol traps" to recognize before reading the scorecard

• ① A different harness is a different score: the same model, under a vendor adapter that carries opaque reasoning state across turns and compresses context, can score 37 points above its result under a neutral harness identical for every vendor. The former measures the system capability of model plus agent framework; the latter is closer to bare-model capability.
• ② Held-out tests are where truth shows: aggregated question banks can be contaminated by training data; tests on vulnerabilities disclosed 6–8 months before release are far more credible. ExploitBench's 100% aggregate and the 39% on novel bugs must therefore be read separately — and the former is a "capability coverage rate," not a pass rate.
• ③ Different test-set versions are not directly comparable: Anthropic's 77.9% on OSWorld used a different versioned subset, and its authors explicitly say it cannot be set against Astra's 72.6%.

Comparison models

Analytical framework

5.1 Research-grade math & science: near saturation, but the field is strong too

• The absolute level is real: 97.6% puts a research-grade math benchmark near saturation, and OpenAI says it has helped resolve several long-standing math problems.
• But it is not a generational gap: Fable 5.1 already reaches 87.8%, and Gemini Flash trails by just 0.7 on GPQA. On this dimension Astra is "the top-scoring student," not "the only one who passed."

5.2 Abstract reasoning: both 99.9% and 62.7% are true — because they measure different things

• The 37.2-point gap is a "system bonus": the adapter lets the model carry opaque intermediate reasoning state into the next turn and compress context — essentially the score of a model plus a strong agent framework. The standard harness allows only the brief notes the model itself chooses to keep.
• Yet 62.7% is still a huge advance: against Sol's 7.8% and Opus 5's 30.2%, Astra is a decisive first under the neutral protocol and beats the human median on action efficiency — leading even after wringing out the inflation is more persuasive than 99.9%.
• Methodological significance: this is the textbook case for reporting "model capability" and "system capability" separately — and the reason this report keeps returning to protocol.

5.3 Cybersecurity: the hardest claim here — and the one to watch most warily

• This is a rating "with teeth": full offensive capability is not open to the public, server-side safety checks can terminate a task outright, production monitors out-of-bounds behavior in real time, and security customers are supplied first through the controlled Daybreak / Trusted Access channels.

5.4 Computer use & engineering delivery: here is the real gap

Rather than compare absolute scores one by one, look at Astra's rate of improvement over the prior Sol generation — comparing growth across benchmarks pinpoints exactly where the leap happened.

• The leap concentrates in carrying complex flows through to completion: research workflows +188.4%, office automation +128.7%, terminal engineering +55.2% — Astra's strength is long-horizon, multi-step, cross-tool process capability.
• Computer Use moves from demo to production: in an official demo a pediatrician-research lookup took 2 min 54 sec (versus ~5 hours for a human); average time per OSWorld task dropped from ~75 minutes to 40 (−47%). Codex can now preserve working notes across contexts and ask questions mid-task without blocking unrelated work.
• Yet raw operation accuracy improves only modestly (OSWorld +10.5%, ALE +10.6%): it is faster and better at finishing, but per-step accuracy does not jump as dramatically as the headline totals imply.

5.5 Coding: neck and neck — even overtaken by a cheaper model

• Astra is not top on test-taking: Meta's Muse Spark 1.3 edges ahead at 75.4% on DeepSWE; Gemini 3.8 Flash and Claude Opus 5 also sit near 74%, with overlapping confidence bands.
• Only ~2.1 points clear of Fable on Terminal-Bench: across ~113 tasks that is the swing of one or two problems — not a significant difference.
• Astra's real coding edge is closed-loop delivery inside the terminal — at lower cost: OpenAI estimates its per-task cost runs ~63% below Fable 5.1 and ~57% below Sol. The advantage comes from less rework and fewer calls, not a higher single-shot accuracy.

5.6 Alignment: the numbers improve markedly — yet monitoring gets harder

Pricing & specs: selling "task cost," not tokens

• Specs: 1.05M-token context, 128k max output, knowledge cutoff 2026-04-30; input past 272k tokens triggers tiered surcharges, and Fast mode doubles the price.
• The core defense (Brockman): "What matters is how much it costs to complete a task, not the price per token" — on high-config tasks OpenAI estimates per-task API cost ~57% below Sol and ~63% below Fable.
• Rollout cadence: first to Daybreak/Trusted Access cybersecurity customers, then within days to Plus/Pro/Business/Enterprise, the API and AWS Bedrock.

Five structural consequences for the LLM ecosystem

• ① The agent battleground shifts from conversation to operating the computer: "screen + mouse + keyboard" becomes the universal interface; enterprise integration, RPA and office automation look set to be rebuilt within 6–12 months (a Legora-style "reconcile 41 financial documents in minutes and surface all four planted errors" is among the first use cases to land).
• ② "The stronger the capability, the narrower the distribution" becomes a new paradigm: Astra is the first model whose power changed its product form; safety ratings move from an internal framework into a product feature and a regulatory variable.
• ③ The benchmark-credibility crisis deepens: model-plus-scaffolding system capability and bare-model capability must be reported separately, amplifying the value of independent ARC-Prize-style retests — precisely the opening for third-party evaluators.
• ④ The open/low-cost camp faces a stress test: if "the expensive model saves on the total bill" is confirmed by real invoices, the high-price route is vindicated; if rework is high and workflows are often false-blocked, the low-cost camp will quickly take the long tail.
• ⑤ AGI shifts from a technical judgment to a brand asset: Brockman removes AGI from a falsifiable contractual definition while using it to endorse the product. Once AGI is marketing language, what is genuinely scarce is third-party, re-checkable, condition-transparent evaluation.

Three confident calls

• Call one: the agent battleground has formally shifted from conversation to operating the computer; enterprise integration, RPA and office automation will be rebuilt within 6–12 months, and Astra currently leads this axis by half a length.
• Call two: "the stronger the capability, the narrower the distribution" is the new frontier paradigm; safety ratings move from internal framework toward product form and regulatory variable, and Critical may become a standard threshold for later flagships.
• Call three: the benchmark-credibility crisis deepens, system and bare-model capability must be reported separately, and independent third parties such as ARC Prize will keep gaining authority — which is exactly why Modelspectra cross-checks multiple sources and labels every protocol.

Signals to track next cycle (falsifiable)

• Real enterprise bills (next two quarters): does "lower per-task cost" replicate on the third-party/user side, or get eaten by rework and false blocks?
• More neutral retests: do third parties beyond ARC Prize reproduce the official numbers on OSWorld, Terminal-Bench and ExploitBench?
• Does the coding overtake persist: can Muse Spark 1.3 and later open models stay ahead of Astra on DeepSWE and similar benchmarks?
• Safety-rating regulation lands: Daybreak access terms, US/UK AISI findings, and whether competitors' security capability also triggers Critical.
• Computer-Use substitution data: do RPA/middleware vendors show quantifiable signs of displacement in products and results?

Whether the "AGI era" begins today depends on whom you ask — but it is certain that OpenAI would like you to believe it does. Strip away the narrative and Astra remains a genuine, high-grade leap on the axis of delivering work; it is simply far from the omnipotence its perfect-score poster suggests.