OpenAI Launches GPT-6 Astra With Record Benchmarks

OpenAI announced GPT-6 Astra on September 3, 2026, claiming 97.6% on FrontierMath Tier 4 v2 and 100% on ExploitBench. The ARC-AGI-3 result comes with a caveat.

by HowAIWorks Team
On this page

Introduction

OpenAI announced GPT-6 Astra on September 3, 2026, and framed it as the beginning of what the company calls the "AGI era." The launch numbers include 97.6% on FrontierMath Tier 4 v2, 96% on GPQA Diamond, 100% on ExploitBench, and a near-saturation result on ARC-AGI-3, the benchmark built specifically to resist this kind of progress.

The scores are real. The most-quoted one comes with an asterisk that matters more than the number.

What OpenAI Reported

The claimed results at launch, as reported by OpenAI at the time of publication:

BenchmarkScoreWhat it measures
FrontierMath Tier 4 v297.6%Research-level mathematics
GPQA Diamond96%Graduate-level science questions
BenchCAD95.9%Programmatic CAD — generating executable CadQuery programs across 106 industrial part families
DeepSWE v1.174.1%113 long-horizon engineering tasks in live open-source repos, checked by program-based verifiers
ExploitBench100%Cyber-exploitation, scored as a 16-flag ladder from reaching vulnerable code to arbitrary code execution
ARC-AGI-398.6%Learning the rules of an unfamiliar interactive environment

Only one of the near-perfect scores is a step change, which is clearer against what the field scored a few months ago. On GPQA Diamond, Claude Opus 5 sits at 94.2% and Gemini 3.1 Pro at 94.3%; Astra's 96% is a 1.7-point move on a converged benchmark and means little. FrontierMath Tier 4 is the opposite — GPT-5.5 scored 35.4%, Claude Opus 4.7 22.9% — so 97.6% is a different order of result. Astra ran the v2 revision rather than the cut those models took, so treat the gap as indicative; nothing about the version explains a jump that size.

Cybersecurity is the outlier. OpenAI restricted those capabilities at launch and routed expanded access through a separate program it calls Daybreak Blue. It has not published the reasoning, but a perfect score on a ladder ending at arbitrary code execution is the obvious context.

Coding is the honest number. DeepSWE v1.1 at 74.1% is strong but nowhere near saturation, and its tasks come from live repositories — which makes it the score most predictive of whether the model helps on a real codebase.

The defining product capability is not on this list. GPT-6 Astra ships with computer use — driving browsers, spreadsheets, and desktop applications through the same interfaces a person uses, without a purpose-built API integration for each one.

The ARC-AGI-3 Asterisk

ARC-AGI-3 tests whether a model can learn the rules of an unfamiliar interactive environment from scratch, which is why a near-perfect claim on it carries so much rhetorical weight for a reasoning model. ARC Prize ran and published its own evaluation, as it does for every frontier submission.

The independent results split in two:

  • 62.7% on the ARC-AGI-3 Semi-Private set using ARC Prize's standard harness, the configuration every provider runs identically. That run cost roughly $26,000 in compute.
  • 99.9% using a provider adapter harness, which preserves OpenAI's opaque reasoning state between requests and compacts long conversations so the model can reuse its own prior work. That run was about 3.66x faster and used 49% fewer tokens.

Same model, same tasks, a 37-point gap. The adapter is not cheating — reusing reasoning state across turns is a real platform capability — but the near-perfect figure measures the model plus OpenAI's proprietary scaffolding, while 62.7% is what survives an apples-to-apples comparison against other providers. The 98.6% that led the launch materials sits between the two and matches neither.

ARC Prize was blunt about the interpretation: saturating the benchmark "would not represent 'proof of achieving AGI,'" because ARC-AGI-3 has a tightly bounded scope and deterministic, closed-ended mechanics. It also reported a striking secondary result — Astra used fewer actions than the median human tester on 96% of completed levels, so on efficiency, not just accuracy, it is past a human baseline drawn from about 500 general-public participants.

VentureBeat flagged a second gap: GDPval, OpenAI's own measure of performance on real occupational tasks, is absent from the launch materials. For a release marketed on economic generality, that is the omission worth watching.

Availability and Pricing

The rollout started with enterprise customers already in OpenAI's Daybreak program, and extends over the following days to ChatGPT Plus, Pro, Business, and Enterprise, plus the OpenAI API, Microsoft Azure, and AWS Bedrock. There is no free tier. Enterprise administrators have to enable the model per workspace — it is off by default at launch.

API pricing, per million tokens:

ModeInputCached inputOutput
Standard$10$1$50
Fast$20$100

Fast mode delivers up to twice the throughput at twice the price. Pro, Business, and Enterprise plans also get a higher-effort GPT-6 Astra Pro variant.

The model ID is gpt-6-astra. It carries a 1,050,000-token context window — matching GPT-5.6 — with 128,000 max output tokens and an April 30, 2026 knowledge cutoff. Input is text and image; output is text only. Tooling includes web search, code interpreter, computer use, and MCP.

Conclusion

GPT-6 Astra is a substantial release: a million-token context, native computer use that changes what an AI agent can be pointed at, and a FrontierMath result that is genuinely discontinuous with what frontier models scored six months ago. Whether it is the start of an "AGI era" is a claim the benchmarks cannot support, and ARC Prize — whose benchmark supplied the headline — said so first.

For anyone actually evaluating the model, the saturated scores no longer discriminate between frontier systems and can be skipped. DeepSWE's 74.1%, and whatever computer-use reliability you can measure on your own workflows, are the numbers with headroom left in them, which makes them the only ones still carrying information.

Sources

Frequently Asked Questions

GPT-6 Astra is OpenAI's frontier model announced on September 3, 2026. It is a reasoning model with a 1,050,000-token context window, text and image input, and built-in computer use, and OpenAI positions it as the start of what the company calls the AGI era.
OpenAI reported 97.6% on FrontierMath Tier 4 v2, 96% on GPQA Diamond, 95.9% on BenchCAD, 74.1% on DeepSWE v1.1, 100% on ExploitBench, and 98.6% on ARC-AGI-3. All are vendor-reported figures at the time of publication except ARC-AGI-3, which ARC Prize evaluated independently.
It is not disputed so much as harness-dependent. ARC Prize measured 62.7% using its standard harness, which every provider runs the same way, and 99.9% using a provider adapter that preserves OpenAI's opaque reasoning state between requests. The two numbers describe the same model under different scaffolding.
Standard pricing is $10 per million input tokens, $1 per million cached input tokens, and $50 per million output tokens. A Fast mode runs at roughly twice the speed for twice the price.
The rollout began with enterprise customers in OpenAI's Daybreak program and extends to ChatGPT Plus, Pro, Business, and Enterprise over the following days, plus the OpenAI API, Microsoft Azure, and AWS Bedrock. There is no free-tier access, and the cybersecurity capabilities are gated behind a separate access program.
No benchmark result settles that question, and ARC Prize said explicitly that saturating ARC-AGI-3 would not represent proof of AGI. The benchmark has deterministic, closed-ended mechanics that do not reflect open-ended real-world work.

Continue Your AI Journey

Explore our glossary and model catalog to deepen your understanding.