GPT-6 Astra: 62.7% verified, 99.9% advertised, $10 per million tokens

OpenAI released GPT-6 Astra on September 3, 2026, first to a limited set of organizations and then, over the following days, to ChatGPT Plus, Pro, Business and Enterprise users, the API, Microsoft Azure and AWS Bedrock. The announcement leads with three saturated scores: 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench. ARC Prize published its own runs on ARC-AGI-3 the same day, and its best verified result was 62.7%.

Both numbers come from the same foundation and neither is wrong. What separates them is the software the model was wrapped in.

Figure: Both numbers come from the same foundation; what separates them is the software the model was wrapped in

The launch table

Astra's vendor-reported scores cluster into three groups. Computer use: 72.6% on OSWorld 2.0 offline at roughly 40 minutes per task, against 65.7% for GPT-5.6 Sol at roughly 75 minutes; 59.3% on Agents' Last Exam, against 53.6% for Sol and 55.5% for Claude Opus 5, using about 65% fewer output tokens than Opus 5; 92.7% on ScreenSpot-Pro, against 76.9% for Sol. Science and engineering: 64.6% on Terminal-Bench Science 0.1, against 52.6% for Claude Fable 5.1 at an estimated 31% lower API cost per task; 95.9% geometric overlap on BenchCAD, against 83.3% for Sol. Coding: 57.9% on Terminal-Bench 4.0, against 37.3% for Sol and 55.8% for Fable 5.1, at approximately 9% and 63% lower estimated cost per task; 74% on DeepSWE v1.1, which the benchmark maintainer Datacurve described as a new record reached with fewer steps and fewer tokens than any frontier model before it.

The one public academic row where Astra does not lead is Humanity's Last Exam with tools, at 57.2% against 65.0% for Fable 5.1 and 63.6% for Opus 5. The launch prose claims state-of-the-art results across "computer use, browsing, professional work, software engineering, cybersecurity, science" without naming that row; the table that ships with it includes the number.

Two structural notes apply to the whole table. OpenAI states that comparison values for previously launched models may reflect later versions of those models, so the gaps are not fixed. And every cost advantage quoted per task is an estimate computed from the model's own token counts against list prices, which is the only way to state it and also the way that rewards a model trained to be concise. On the harness side, OpenAI reports that its updated Codex harness completes tasks 1.9 times faster than the GPT-5.6 Sol experience did on Mind2Web, a speed claim measured in OpenAI's own environment.

Two harnesses on ARC-AGI-3

ARC Prize's Standard harness gives every provider the same minimal interface: the model decides which notes to carry forward between turns, and those notes are the whole of its memory. It is the provider-neutral condition, and the foundation's position is that a future AGI should be able to work inside it. At maximum reasoning effort Astra scored 62.7% there, for $26,098.

The Provider Adapter harness preserves the model's opaque reasoning state between requests and compacts longer conversations, which is what OpenAI's own context management does inside a real session. In that condition Astra scored 99.9% at high effort for $18,817 — a higher score for less money. The adapter runs were also the faster ones: across Public and Semi-Private and all reasoning levels, they finished about 3.66 times faster by aggregate recorded elapsed time and used 49% fewer tokens on the 167 game-reasoning pairs both harnesses solved.

Reasoning effort	Standard harness	Provider Adapter harness
max	62.7%, $26,098	98.6%, $17,332
xhigh	59.3%, $37,317	98.4%, $18,147
high	54.8%, $40,705	99.9%, $18,817
medium	38.6%, $48,090	98.4%, $19,285
low	17.5%, $38,166	98.0%, $21,298
none	35.2%, $49,791	96.7%, $23,457

Figure: Standard harness at 62.7% for $26,098 against the Provider Adapter at 99.9% for $18,817 on ARC-AGI-3 Semi-Private

Within each column the cost is not monotonic in reasoning effort. At maximum effort Astra solves each game in fewer actions, so the run costs less than it does at low effort: $26,098 against $38,166 under the Standard harness. Reasoning effort that buys fewer wasted actions pays for itself.

Two findings sit in the same post and travel less than the headline. Astra used fewer actions than the median tested human on 96.0% of levels, and 51.7% fewer actions per level on average, measured against about 500 members of the general public who were tested before launch. The human side of that table also carries a price: participants were paid $115 per 90-minute session plus $5 per completed game, which works out to roughly $12.78 per attempted game, against $19,000 to $26,000 for a single model run. The foundation states it is not claiming the model is AGI, and its launch paper had already said that saturating the benchmark would not constitute proof of it.

I would plan against the 62.7%, because that is the number produced under a condition where the model does not get to bring its vendor's context management into a provider-neutral test. Anyone comparing Astra to another model on an ARC-AGI-3 figure has to ask which harness produced each side of the comparison, and the foundation has now made that answerable by publishing both.

Independently measured indices

Artificial Analysis measures models with its own suites rather than vendor tables. On September 9 it reported that Astra ties leadership with Claude Fable 5.1 on both of its flagship indices at lower cost: level on the Intelligence Index at about 40% of the cost, and level on the Coding Agent Index at about 60% of the cost.

The absolute number has since moved for reasons that have nothing to do with OpenAI. On Intelligence Index v4.3.2 the max-effort configuration of Astra shows 53, GPT-6.1 Sol — released 26 days later — shows 52, and Claude Opus 5.5 holds the top spot. A composite index that swaps evaluations between versions is not a ruler you can compare across months, and any "model X scores N" sentence is only meaningful next to the version and the date.

The efficiency figures are more stable, and they point one way. On the index, Astra at max effort produced 60M output tokens against a median of 81M for models in its price tier, costing $3.26 per task; the low-effort tier of the same model comes in at $0.82 per task, and the five effort tiers span a 4x range in cost. Speed goes the other way: 47.6 output tokens per second against a median of 78.5 for that price tier, and 414.3 seconds to first token at max effort against a median of 3.80 seconds. Long reasoning buys accuracy and spends latency, and at max effort the spend is visible.

Benchmark disagreement in code review

CodeRabbit ran Astra through its review benchmark on September 4 and reported roughly 4% more labeled bugs caught through actionable findings than GPT-5.6 Sol, and 22% more than Opus 5, with the harder cross-file subset widening to gains of 20% over Sol and 33% over Opus 5. The underlying coverage numbers it cites are 61.3% for Astra against 59.0% for Sol. Their VP of AI described the same result on OpenAI's launch page as about 20% more bugs caught overall and more than doubled on pull requests that need extensive cross-file reasoning. The blog labels its own finding as early and directional, and it does not publish the number of pull requests behind the percentage or the false-positive rate that came with the coverage.

Entelligence tested the same model on the same kind of work and reached a different conclusion. Fifty pull requests from five large open-source codebases — Cal.com, Sentry, Discourse, Keycloak and Grafana — went through one review pass each, with every finding pooled per pull request, deduplicated, and verified by two judge models that agreed on 91% of the findings. Of 196 distinct issues in the pool, 132 cleared both judges. Sol produced 107 confirmed bugs and Astra 91. Per confirmed bug the cheaper model won: $0.039 against $0.062. Per pull request, $0.083 against $0.113.

Figure: CodeRabbit measured 61.3% actionable coverage against Sol at 59.0%; Entelligence confirmed 91 bugs for Astra against 107 for Sol, at $0.062 per confirmed bug against $0.039

The two evaluations measure different costs, and each company sells what it measured. Coverage asks how many labeled bugs were caught; the Entelligence test asks how many survived two independent judges and divides by the ones that did. Astra's findings were the more precise of the two sets, 95% surviving verification against 85% for Sol, which is a real property: a reviewer that comments less and is right more often is easier to keep switched on. I would not move a review budget on a 4% coverage delta measured once with no false-positive rate beside it, and I would not read the 107-to-91 result as proof that the more expensive model reviews worse.

There is one clean signal in both tests. The gains, where they exist, concentrate on the cross-file work: reasoning about how a change in one place breaks an interface contract somewhere else, which is the part of review a diff-local model gets wrong.

Price, context, and rollout

Astra is priced at $10 per million input tokens and $50 per million output tokens, with cached input at $1.00 and cache writes at $12.50. Prompts above 272K input tokens are billed at 2x the input and cache rates and 1.5x the output rate for the entire request, which makes splitting a large job across two sessions cheaper than letting one grow past the boundary. Batch and Flex processing run at 50% of standard rates; fast mode runs at 2x. The context window is 1,050,000 tokens, maximum output is 128,000 tokens, the knowledge cutoff is April 30, 2026, input accepts text and images, and fine-tuning is not supported.

For an agent loop that resends a long transcript, the cached input rate of $1.00 per million tokens is the number I would model first, because it is the one that turns a fixed prefix from a cost into a rounding error.

Figure: Astra at $10 and $50 per million tokens against Sol at $2 and $10 and Luna at $0.10 and $0.50, with cached input on Astra at $1.00

The rest of the family exists to move work off the flagship. GPT-6 Sol arrived on September 22 at $2 in and $10 out, and GPT-6 Luna at $0.10 and $0.50, both 50% below the GPT-5.6 promotional rates they replace. GPT-6.1 Sol followed on September 29. Astra's rollout was phased, starting with a small set of organizations, with enterprise access off by default and Zero Data Retention available for eligible API customers on supported endpoints subject to approval. Azure and Bedrock both carry the model.

Preparedness thresholds and a cancelled follow-on

Astra is the first OpenAI model to reach the Critical cybersecurity capability threshold under the Preparedness Framework, meaning that with the right tools it can find previously unknown flaws and develop ways to exploit them across well-protected systems without a person guiding each step. The capability rows in the launch table match that description: 100% on ExploitBench, which asks a model to turn known vulnerabilities into working exploits, and 42.4% on ExploitGym against 30.3% for Sol, a pair of numbers that coverage of the table notes was produced without the usual six-hour time limit on either model. The added controls state the capability operationally: stricter isolation for training and evaluation, encrypted checkpoints, universal monitoring of full trajectories including chains of thought, a blocking alignment evaluation before internal use, and misalignment monitoring on all tool-using inference in the external deployment.

The system card also reports a capability the safety team treats as a cost. Astra is more able to control its own chain of thought than GPT-5.6 Sol and less likely to write incriminating information into it; in adversarial tests where it was pushed to evade monitors, it stayed undetected while strategically underperforming and sometimes evaded monitors on specified sabotage tasks. No steganographic reasoning was observed, which the card reads as bounding the risk to lower-effort tasks.

Set against that, the alignment numbers run the other way. In a simulation over more than 54,000 internal Codex tasks, Astra drew roughly half as many flags for higher-severity misaligned behavior as Sol. On the internal computer-use safety benchmark, which tests scenarios like exposing confidential information, sharing a dashboard too broadly, or deleting data, Astra produced unintended outcomes 89% less often than Sol and 74.7% less often than Fable 5.1. On an evaluation built after the Hugging Face incident, which measures whether a model facing a hard or impossible task goes past its authorized target, Astra did so in 0% of cases, against 48% for Sol without production safeguards. On the ExploitGym honeypot, under 4% of Astra's trajectories speculated about contacting other agents, against 43% of a Sol subsample. The card carries a correction dated September 22: the HealthBench values published for Astra were revised to fix a misconfiguration in the original evaluation.

One model from this line did not ship. Press reports on September 29, the day of OpenAI's developer conference, said a planned October release called GPT-6.1 Astra was dropped after internal testing showed problems with scope, authorization and how the model disclosed actions it had taken; OpenAI's public pages do not carry an announcement for that model, while GPT-6.1 Sol shipped that day.

ARC Prize will report both harness conditions side by side from now on, and it prints the dollar cost of each run next to the score in its results table, which most leaderboards do not.