OpenAI's most capable deployed model sets records on computer use, terminal work, long context, and cybersecurity, and lands roughly level with the Anthropic frontier on general coding and reasoning. It is also the first model OpenAI has classified Critical for cyber capability under its Preparedness Framework, which changes the access story as much as the benchmark story.
ScreenSpot-Pro goes from 76.9% to 92.7% against GPT-5.6 Sol, OSWorld 2.0 from 65.7% to 72.6%, and SRE-Bench single-attempt solve rate from 55.9% to 88.0%. On the Artificial Analysis Intelligence Index it scores 61.2, below Claude Fable 5.1 at 65.7 and Opus 5 at 63.1. Both statements are true at once, and which one matters depends entirely on what you are building.
Terminal-Bench 4.0 is a genuine win at 57.7% against 55.8% for Fable 5.1. But FrontierCode 1.1 Main puts Astra at 53.3% against 53.5% for Fable 5 and 53.4% for Opus 5, and the Artificial Analysis Coding Agent Index has all three inside a point and a half. If you route on coding alone, this release does not move your decision.
$10 input and $50 output per million tokens, matching Claude Fable 5.1 exactly. Cached input is $1, which is four times what Fable 5.1 charges for a cache read. Cross 272K input tokens and the entire request reprices at 2x input and 1.5x output, so a single long request costs $20 / $75.
OpenAI says Astra can find previously unknown security flaws and develop new ways to exploit them across many well-protected systems. It scores 100% on ExploitBench and achieves arbitrary code execution on 39.0% of a contamination-controlled set of 20 high-severity V8 vulnerabilities, against 5.5% for Sol. Access to cyber and biology capability runs through separate trust-based tiers.
The system card reports a substantial decrease in chain-of-thought monitorability against previous models, and says that if the trend continues across generations they would soon have significantly reduced confidence in detecting many forms of misaligned behaviour. That is a candid disclosure and a real operational risk for anyone running long autonomous agents.
OpenAI's launch numbers set against GPT-5.6 Sol, which Astra replaces at the top of the lineup, and against the Anthropic frontier two days earlier. Read the caveats under the table before you read the table: several of these were run under harnesses that are not directly comparable across vendors.
| Capability | GPT-6 Astra | GPT-5.6 Sol | Claude Fable 5.1 | Claude Opus 5 |
|---|---|---|---|---|
Agentic terminal coding Terminal-Bench 4.0 | 57.7% +20.4 pts vs Sol | 37.3% | 55.8% | 52.3% |
Scientific terminal work Terminal-Bench Science Sol and Opus 5 figures as Anthropic published them | 64.6% +42.2 pts vs Sol | 22.4% | 52.6% | 29.0% |
GUI grounding ScreenSpot-Pro | 92.7% +15.8 pts vs Sol | 76.9% | not published | not published |
Agentic computer use OSWorld 2.0 variant not stated; Anthropic reports partial and strict separately | 72.6% +6.9 pts vs Sol | 65.7% | 77.9% partial / 41.7% strict | 75.4% partial / 39.6% strict |
Agentic automation AutomationBench | 41.4% +23.3 pts vs Sol | 18.1% | 31.4% | 26.9% |
Incident response SRE-Bench, single attempt 99.2% within four attempts, against 68.7% for Sol | 88.0% +32.1 pts vs Sol | 55.9% | 12.5% | not published |
Long-context retrieval MRCR v2, 8 needles, 512K to 1M | 96.3% +22.5 pts vs Sol | 73.8% | not published | not published |
Frontier mathematics FrontierMath Tier 4 (v2) Epoch AI notes OpenAI funded the benchmark's development | 97.6% | not published | 87.8% | 73.2% |
Abstract reasoning ARC-AGI-3 99.9% on OpenAI's Provider Adapter harness, 62.7% on the standard one | 99.9% +92.1 pts vs Sol | 7.8% | not published | 30.2% |
Multidisciplinary reasoning Humanity's Last Exam, with tools omitted from the launch narrative | 57.2% | not published | 65.0% | 63.6% |
General intelligence Artificial Analysis Intelligence Index v4.1.1 third-party index, Astra places fourth | 61.2 | not published | 65.7 | 63.1 |
Exploit development ExploitBench run without production safeguards on both models | 100.0% +21.5 pts vs Sol | 78.5% | 70.0% | not published |
Every figure above was run at maximum effort unless the row says otherwise, which is not how anyone runs a model in production. Four caveats travel with this table. The ARC-AGI-3 headline of 99.9% is on OpenAI's own Provider Adapter harness and drops to 62.7% on the standard one, while rival models were scored under different setups. Epoch AI, which runs FrontierMath, notes that OpenAI funded its development. The BenchCAD comparison ran Claude under modified evaluation settings. The ExploitGym comparison removed the usual six-hour time limit, and the cyber benchmarks ran without production safeguards on both models. We re-run our own evals on customer tasks before recommending a switch, which is the only number that has ever decided a routing call for us.
What changes for the engineering team. Two comparisons carry the decision: Astra against the GPT it replaces, and Astra against Claude Fable 5.1, which arrived two days earlier at exactly the same headline price.
| Dimension | vs GPT-5.6 Sol | vs Claude Fable 5.1 |
|---|---|---|
Computer use and browser agents | The clearest upgrade in the release. ScreenSpot-Pro grounding goes from 76.9% to 92.7% and OSWorld 2.0 from 65.7% to 72.6%. If you run a browser or desktop agent, this is the reason to look at Astra at all. | Astra leads on the GUI grounding numbers each vendor published, though the OSWorld variants are not stated consistently enough to declare a winner. Anthropic reports Fable 5.1 higher on OSWorld 2.0 partial credit. Run both on your own flows. |
Coding | A large jump on terminal work (37.3% to 57.7% on Terminal-Bench 4.0) and a modest one elsewhere. DeepSWE v1.1 moves 72.7% to 74.1%, which is inside the noise most teams can measure. | Effectively a tie. Astra edges Terminal-Bench 4.0, Fable 5 edges FrontierCode 1.1, and the Artificial Analysis Coding Agent Index puts Fable 5 at 68.1%, Fable 5.1 at 67.2%, and Astra at 67.0%. Nobody should re-platform a coding agent on this. |
Cost | Two and a half times Sol on both input and output ($10 / $50 against $4 / $20). Sol stays the right default for most production work, and Astra is a routing target for the tasks that justify the rate. | Identical headline pricing. The difference is in caching: Astra reads cached input at $1 per million, Fable 5.1 at $0.25. On a long agent session that re-reads a big cached prefix every turn, that is a four-to-one gap in Anthropic's favour on the largest line of the bill. |
Long context | MRCR v2 eight-needle retrieval in the 512K to 1M band goes from 73.8% to 96.3%, which is the difference between a window you can advertise and one you can rely on. Same 1.05M window on both. | Fable 5.1 prices its full 1M window at the standard per-token rate with no surcharge. Astra reprices the entire request at 2x input and 1.5x output above 272K input tokens, so a genuinely long request costs $20 / $75. If your workload lives past 272K, this is the deciding line. |
Operational risk | Better on most safety measures: about 53% fewer severity-3-or-above misalignment flags across more than 54,000 internal tasks, indirect prompt-injection robustness at 99.79% against 96.23%, and an internal hallucination rate of 4.2% against 12.2%. Against that, a disclosed regression in chain-of-thought monitorability. | Different postures. Anthropic gates cyber capability down and reports Fable 5.1 as its most prompt-injection-robust model. OpenAI ships a Critical cyber classification with trust-tiered access. For regulated buyers this is a procurement conversation, not a benchmark one. |
Inside a Kensink build, model selection is a routing decision the agent makes at runtime, not a vendor commitment frozen at design time. Astra changes our routing for computer use and incident response, and changes nothing about where we send general coding work until a customer eval says otherwise.
Reasoning effort now runs low, medium, high, xhigh, and max. Every headline benchmark in the launch was run at maximum effort, which is worth remembering when you compare a launch number to what you see at the level you can actually afford. Reasoning tokens bill as output at $50 per million.
The window is 1,050,000 tokens with a 922,000 maximum input and 128,000 maximum output. The number that matters is MRCR v2 eight-needle retrieval at 96.3% in the 512K to 1M band, against 73.8% for Sol. A window you can advertise is not the same as a window you can put an agent inside.
Requests above 272,000 input tokens are billed at 2x input and cache rates and 1.5x output for the full request, not just the excess. Cache writes bill at 1.25x the uncached input rate. Anyone planning to use the full window should model cost at the long-context rate of $20 / $75 rather than the headline $10 / $50.
Astra is the first model to reach Critical for cybersecurity under OpenAI's Preparedness Framework. The public model is restricted and refuses in cyber areas; genuine security research runs through separate trust-based access tiers. OpenAI also delayed the release after its July 2026 Hugging Face incident to add safeguards.
OpenAI added a misalignment monitoring system across tool-using inference, plus production classifiers layered on top of base model behaviour, confirmation policies for consequential actions, and age-appropriate boundaries for under-18 users. This is infrastructure around the model, not behaviour inside it, and it can refuse things your evals passed.
The system card reports a substantial decrease in chain-of-thought monitorability compared with previous models, and states plainly that a continued trend would soon leave OpenAI with significantly reduced confidence in detecting many forms of misaligned behaviour. Publishing that is the right call, and it belongs in your risk register if you run long autonomous agents.
Astra is not a drop-in for every GPT-5.6 call site. Nothing here throws on a simple text completion, but the modality surface is narrower than the model it sits above, and the cost model has a cliff in it.
Astra serves Chat Completions, Responses, and Batch. Realtime, Assistants, Fine-tuning, Embeddings, Image generation, Video, Audio, and Moderation are all unsupported. Any code path that treats the flagship model ID as a general-purpose default across the OpenAI surface will fail on the endpoints it does not serve. Route audio to the realtime and transcription models and embeddings to the embedding models, as separate concerns.
v1/chat/completions · v1/responses · v1/batchAstra accepts text and images and returns text. It does not generate images, audio, or video. If you were relying on a single model ID for multimodal output, that path needs splitting: gpt-image-2 for images, the realtime models for speech.
input: text, image · output: textTwo levels above high. They buy accuracy on the hardest tasks and they buy it with reasoning tokens, billed as output. Treat max as an escalation path for a retry rather than a default, and measure cost per completed task rather than cost per call.
reasoning: { effort: "low" | "medium" | "high" | "xhigh" | "max" }Cached input reads bill at $1 per million tokens against $10 uncached, a 90% discount, and cache writes bill at 1.25x the uncached input rate. Worth wiring properly for any agent with a large shared preamble, though note that Claude Fable 5.1 charges $0.25 for the equivalent read.
cached input: $1.00 / 1M · cache write: $12.50 / 1MBatch and Flex run at 50% of standard rates, and Fast mode at 2x. For anything that is not user-facing and not latency-bound, batch turns a $10 / $50 model into a $5 / $25 one, which is the cheapest lever available on this model.
batch / flex: 0.5x · fast: 2xStreaming, function calling, structured outputs, file search, web search, image input, and prompt caching all work as on GPT-5.6. If your integration only uses these, the model string is genuinely the only change.
streaming · function_calling · structured_outputs · file_search · web_searchRate limits are tier-gated and worth checking before a load test: tier 1 starts at 500,000 TPM, and tier 5 reaches 15,000 RPM, 40,000,000 TPM, and a 15,000,000,000 token batch queue. Rollout was staged from a limited partner set on day one to ChatGPT Plus, Pro, Business, and Enterprise, then the API and AWS.
Model IDs on every platform, the limits, and the two lines that decide whether procurement signs off. Kept here so nobody has to reconstruct it from three vendor pages.
| Model ID | gpt-6-astra |
|---|---|
| Context window | 1,050,000 tokens |
| Maximum input | 922,000 tokens |
| Max output | 128,000 tokens |
| Knowledge cutoff | 30 April 2026 |
| Modalities | Text and image in, text out. No audio, no image output |
| Reasoning effort | low, medium, high, xhigh, max. Reasoning tokens are billed as output |
| Endpoints | Chat Completions, Responses, Batch |
| Not supported | Realtime, Assistants, Fine-tuning, Embeddings, Image generation, Video, Audio, Moderation |
| Features | Streaming, function calling, structured outputs, file search, web search, image input, prompt caching |
| Long-context threshold | 272,000 input tokens. Above it the whole request reprices |
| Rate limit, tier 1 | 500,000 TPM |
| Rate limit, tier 5 | 15,000 RPM, 40,000,000 TPM, 15,000,000,000 batch queue |
| Availability | OpenAI API and AWS. Staged rollout from a limited partner set on day one |
Astra is an escalation target, not a global default. This is the order we work in when a customer asks whether to adopt it.
Computer use, browser and desktop agents, terminal work, incident response, long-context retrieval past 272K, and frontier maths. If your workload is general coding or general reasoning, the published numbers say Astra is level with what you already have, and the migration is not worth running.
The headline is $10 / $50, but above 272,000 input tokens the whole request reprices to $20 / $75. Take your p95 request size, not your median, and price it there. Then check the batch path: half rates on anything that is not user-facing.
Grep for anywhere the flagship model ID is used against audio, embeddings, image generation, moderation, assistants, or realtime. Those are separate models now and need separate routing.
Every launch number was measured at maximum effort. Run your eval suite at low, medium, and high, and find the level where the quality gain stops paying for the reasoning tokens. That level, not the launch slide, is your production setting.
Misalignment monitoring runs across tool-using inference, with production classifiers on top and confirmation policies for consequential actions. Any of these can refuse a request your evals passed. Handle refusals as a response shape and test the path.
OpenAI disclosed a substantial drop in chain-of-thought monitorability. If you run long autonomous agents, that argues for more external verification: independent checks on agent output, tighter tool permissions, and a human gate on consequential actions.
Which model for which job, across the current OpenAI lineup. Astra sits at the top as an escalation target, and most production traffic still belongs further down the list.
$10 / $50 standard, $20 / $75 above 272K input tokens. Reach for it when a cheaper tier has already failed your eval, and measure cost per completed task rather than cost per call.
$4 / $20, with the same 1.05M context as Astra. Astra's coding lead over Sol is real on terminal work and modest elsewhere, so most teams should stay here and route up selectively.
$2 / $12. Our default tier for the bulk of the token volume in a working system.
$0.20 / $1.20, fifty times cheaper than Astra on input. A surprising amount of an agent's traffic belongs here once you actually measure which steps need judgement.
Per million tokens, standard tier. Batch and Flex run at 50% of these rates and Fast mode at 2x. The line that catches teams out is the long-context threshold: cross 272,000 input tokens and the entire request bills at 2x input and cache and 1.5x output, so Astra becomes a $20 / $75 model rather than a $10 / $50 one. Cache writes bill at 1.25x the uncached input rate. For comparison, Claude Fable 5.1 matches the $10 / $50 headline but charges $0.25 for a cache read against Astra's $1, and applies no long-context surcharge across its 1M window.
Under OpenAI's Preparedness Framework, Critical means the model can find previously unknown security flaws and develop new ways to exploit them across many well-protected systems. It scores 100% on ExploitBench and reaches arbitrary code execution on 39.0% of a contamination-controlled set of 20 high-severity V8 vulnerabilities, against 5.5% for Sol. The publicly deployed model is restricted and refuses in these areas at a 91.5% cyber jailbreak refusal rate, against 59% for Sol, with genuine research routed through trust-based access tiers.
Astra logs roughly 53% fewer severity-3-or-above misalignment flags than Sol across more than 54,000 internal tasks, raises indirect prompt-injection robustness from 96.23% to 99.79%, cuts the internal hallucination rate from 12.2% to 4.2%, and scores 0.0% on an ExploitGym honeypot-cheating test where Sol scored 48.2%. The regression is monitorability: reasoning that is measurably harder to inspect.
Misalignment monitoring runs across all tool-using inference, with production classifiers layered above base behaviour, confirmation policies on consequential actions, and age-appropriate boundaries for under-18 users. This is deployment infrastructure rather than model behaviour, which means it can decline a request your evals passed and it can change without a model release. Handle refusals as a response shape and test that path.
Astra sets records on computer use, terminal work, long-context retrieval, and cyber, and sits fourth on a third-party general intelligence index behind Fable 5.1, Opus 5, and Fable 5. Calling it a generational leap is fair for the categories it targets and misleading everywhere else. We route on categories, so this is easy to act on.
The 99.9% on ARC-AGI-3 is 62.7% on the standard harness. FrontierMath was funded by OpenAI. The BenchCAD comparison ran Claude under modified settings. ExploitGym dropped its six-hour limit. None of this makes the model bad, and all of it makes the table unusable as a procurement document.
Browser and desktop agents, and incident response. A jump from 76.9% to 92.7% on GUI grounding and from 55.9% to 88.0% on single-attempt SRE-Bench is the kind of margin that changes what is buildable, not just what benchmarks well. General coding stays where it is until a customer eval moves it.
Astra and Fable 5.1 have identical headline pricing. On a long agent session the bill is dominated by re-reading a cached prefix, and there Fable 5.1 charges $0.25 per million against Astra's $1, with no long-context surcharge. If your workload is a long-running agent rather than a one-shot call, that gap is worth more than any row in the benchmark table.
OpenAI wrote down that its model's reasoning became substantially harder to monitor, and that continuing the trend would leave it with significantly reduced confidence in detecting misaligned behaviour. Publishing that is genuinely good practice. It also means external verification, tight tool permissions, and human gates on consequential actions are engineering requirements now, not governance theatre.
Every model we integrate runs through the same operating system. Three pillars, sixteen layers, one Compound Growth Loop. The methodology that keeps AI work from rotting after the first ship.
Read the K-FrameworkDirect API integration with the model. No LangChain, no orchestration vendor, no agent framework built on quicksand. Typed contracts, the same way we wire up Postgres.
An eval suite built from your real tasks gates every prompt and model change. Quality is measured before it ships, not vibed in a demo.
Governance, audit, and oversight wired in from day one. Who called what, with which prompt version, at what cost. Your auditors get answers, not screenshots.
A model in production without observability is roulette. We instrument every integration so engineering and finance can see the same numbers, and so a regression at 3am surfaces before a customer opens a ticket.
Tokens in, tokens out, dollars spent. Sliced by feature, tenant, and route. Budgets enforced where it matters.
Real distributions, not averages. We know which routes are slow, and why.
The same eval suite that gates a release runs continuously in production. A regression on real traffic surfaces fast.
PII scrubbed at the proxy, shipped to your SIEM. Retention controls match your compliance window.
Dashboards your team owns, not ours. At handoff you get the queries, the alerts, and the runbook. We are not in the path to read your metrics.