GPT-6 Astra Review: Features, Benchmarks & Pricing

GPT-6 Astra is OpenAI's newest frontier model, released on September 3, 2026. It takes over the top of the lineup from GPT-5.6 Sol, and it is built less like a chatbot and more like an operator: it drives real software on your behalf, fills in forms, runs browser checks, builds slide decks, and works through scientific data.

This review covers what GPT-6 Astra does well, what its benchmark numbers mean in practice, what it costs, and where the limits sit. It also covers access: GPT-6 Astra is no longer available in Chat Smith, but GPT-6 Sol and GPT-6 Luna, which bring much of Astra's progress to faster and cheaper models, are.

OpenAI describes GPT-6 Astra as its most capable and most aligned model to date, with state-of-the-art results across computer use, browsing, software engineering, cybersecurity, science, and professional work. It ships with a context window of roughly 1.05 million tokens and can return up to 128,000 tokens in a single response.

The real change is in how it works rather than how it writes. Astra is trained to run long, multi-step jobs end to end: reading a screen, operating actual software, checking its own output, and carrying constraints from the start of a task through to the finish. In the API you can set reasoning effort anywhere from low to max, trading response speed against depth on hard problems.

It also reads instructions better. Earlier models tended to treat a mid-task correction as a brand-new goal. Astra folds new requirements into the job it already has, asks a focused question when the answer would change the outcome, and makes a sensible assumption when it would not.

Since launch, OpenAI has filled out the GPT-6 family with faster, cheaper models trained with similar methods: GPT-6 Sol and GPT-6 Luna arrived on September 22, 2026, and GPT-6.1 Sol followed on September 29, nearly matching Astra on coding, computer use, and professional work at one-fifth of its standard token price.

Four areas separate GPT-6 Astra from the generation before it.

This is where the jump is clearest. Astra can fill in online forms, update customer records in a CRM, organise a calendar, research a question and drop the summary into an email or a document, then run frontend QA on a site it just built. On OSWorld 2.0 it scored 72.6% at around 40 minutes per task, against 65.7% at around 75 minutes for GPT-5.6 Sol — better results in roughly 47% less time. On ScreenSpot-Pro without tools it reached 92.7%, up from 76.9%.

For business work, template adherence matters more than raw writing quality. Astra follows an existing deck or document template closely, matches a house style, and is trained to pull only the context that belongs in the output instead of padding it out. On AutomationBench, which tests real office automation, it scored 41.4% against 18.1% for Sol.

On Terminal-Bench 4.0, Astra scored 57.9% — a wide gap over Sol's 37.3% and ahead of Claude Fable 5.1 at 55.8%. In Codex it can also keep notes across context windows instead of squeezing a long session into one summary every time it runs out of room, so details like why a fix failed survive into the next stretch of work.

Astra scored 97.6% on FrontierMath Tier 4 and 96.0% on GPQA Diamond, and OpenAI credits it with contributions to two new results on gaps between prime numbers. On Terminal-Bench Science it reached 64.6%, nearly triple Sol's 22.4%. Long-context recall improved as well: 96.3% on the 512K–1M range of OpenAI's MRCR test, up from 73.8%. If your day mostly involves working through math problem sets or having to summarize a long PDF, that is the kind of gain you notice immediately.

The table below collects OpenAI's published figures for Astra, its predecessor, and Claude Fable 5.1 as an outside reference point. Scores are the best result at any reasoning effort.

Two patterns stand out. The gaps are widest on agentic, tool-driven tasks and narrowest on knowledge questions, where the previous generation was already near the ceiling. And Astra tends to reach these scores using fewer output tokens, so the cost per finished task can land lower than the headline price suggests.

In the OpenAI API, GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens, with separate rates for cache reads and writes. A fast mode runs at up to double the speed for double the price, and prompts above 272K input tokens are billed at higher multiples.

• ChatGPT: included in the Plus, Pro, Business, and Enterprise allowances, with extra usage sold as credits.
• GPT-6 Astra Pro: an extended version for the Pro, Business, and Enterprise plans.
• Developers: available as gpt-6-astra in the OpenAI API, plus Microsoft Azure and AWS Bedrock.
• Enterprise workspaces: access is off by default at launch, so an admin has to switch it on.
• Chat Smith: GPT-6 Astra is not available in Chat Smith. You can use GPT-6 Sol and GPT-6 Luna instead, with limited use on the free plan and higher usage limits on Chat Smith Pro at $9.99 per month or $39.99 per year.

Astra is the first OpenAI model to reach the Critical cybersecurity level under the company's Preparedness Framework, meaning that with the right access it can find unknown flaws and build working exploits. The version shipped to users refuses advanced offensive tasks such as writing proof-of-concept exploits, while defensive work like secure code review and patching stays available.

OpenAI also runs misalignment monitoring in production for Astra-class models. Those checks can slow, pause, or stop a task mid-run, including legitimate security work, and in ChatGPT or Codex you may be asked to review an action before it continues.

One result moved the wrong way. OpenAI reports that Astra's written reasoning is harder to monitor than Sol's, partly because it solves problems in fewer written steps, and says improving monitorability is still open research. Worth knowing if you plan to delegate work you cannot easily audit.

And the usual caveat holds: benchmark gains do not map neatly onto your own workflow. At $50 per million output tokens, Astra is a poor fit for high-volume, low-difficulty work — that is what cheaper models in the same family, such as GPT-6 Sol and GPT-6 Luna, are for.

GPT-6 Astra is no longer available in Chat Smith, but two newer members of the same family are. GPT-6 Sol is the balanced choice for complex coding, agentic workflows, and demanding everyday work, while GPT-6 Luna handles fast, high-volume tasks like writing, summaries, and quick research. Pick either from the model list and start chatting on Android, iOS, or in the browser — free users get limited use, and Chat Smith Pro raises those limits for heavier work.

The practical advantage is comparison. GPT-6 Sol and GPT-6 Luna sit alongside Claude Sonnet 5, Gemini 3.5 Flash, and Grok 4.5 in the same app, so you can put one question to several models and keep whichever answer holds up. Context carries over when you switch mid-conversation, and there is no second account or second subscription to manage.

A sensible split in daily use: send long documents, multi-step research, debugging, and anything that needs careful reasoning to GPT-6 Sol, and leave quick lookups and short rewrites to GPT-6 Luna. For a closer look at each, see our GPT-6 Sol review and GPT-6 Luna review.

GPT-6 Astra is the first model in this lineup you can reasonably hand a whole task to instead of a single question. The gains show up in agentic work — computer use, long coding sessions, document production, scientific workflows — far more than in everyday chat, and they arrive with a higher output price and stricter safety checks. For work that used to need someone watching every step, that trade is usually worth making.

If you want Astra-style capability without the Astra price, GPT-6 Sol and GPT-6 Luna are live in Chat Smith now, with higher usage limits on Chat Smith Pro.