GPT-6 Astra: What OpenAI Released, Pricing and Benchmarks

OpenAI announced GPT-6 Astra on September 3, 2026, with most coverage landing the following day. It did not reach everyone at once: enterprises in OpenAI's application-based Trusted Access Program got it first, and the API plus ChatGPT Plus, Pro, Business and Enterprise followed over the days after, alongside availability through Amazon Web Services. Sources: CNBC, Al Jazeera.

It is OpenAI's current flagship model, and the name matters: Astra is the release, GPT-6 is the generation. OpenAI pitches it as the "world's most intelligent and aligned" model and aims it squarely at long-horizon agentic work, meaning tasks that run for many steps without a human correcting course: computer use, software engineering, research and multi-step professional workflows. Source: OpenAI via CNBC.

On the API, $10 per million input tokens and $50 per million output tokens, with reused prompt prefixes billed at $1.00 per million cached input tokens. That caching discount is not a detail: for agents that resend a large system prompt on every turn, it is the difference between a workable bill and an unworkable one. ChatGPT subscribers reach the model through their existing Plus, Pro, Business or Enterprise plan instead. Sources: OpenRouter, CloudZero.

1,050,000 tokens of input and up to 128,000 tokens of output. Roughly, a million tokens is a few thousand pages of text, so an entire codebase or a full case file fits in a single request. The output ceiling is the one people forget: a model can read a million tokens and still be capped at what it writes back in one response. Sources: OpenRouter, llm-stats.

On OpenAI's own published benchmarks, Astra scores ahead of both its predecessor GPT-5.6 Sol and Anthropic's Claude Fable 5. Read that with the usual caution: these are the vendor's numbers on the vendor's chosen evaluations, and benchmark leads have repeatedly failed to survive contact with real workloads. The honest answer is that the two are close enough that the choice usually comes down to price, latency and how each behaves on your specific task. Source: Al Jazeera.

OpenAI reports 99.9% on ARC-AGI-3, 97.6% on FrontierMath Tier 4, 72.6% on OSWorld 2.0 and 100% on ExploitBench. The most telling number is long-context retrieval: 96.3% on MRCR v2 against 73.8% for GPT-5.6 Sol. A jump that size is what actually makes a million-token window usable rather than decorative, since a context you cannot reliably retrieve from is just expensive. Source: OpenAI.

Astra exposes five reasoning-effort levels, so you choose how long the model thinks before answering. Low effort is cheaper and faster for routine extraction or formatting; high effort spends far more output tokens on internal reasoning and suits proofs, debugging and planning. Since output tokens are billed at $50 per million, the setting is a direct cost lever, not just a quality one.

Three routes. ChatGPT, if you are on Plus, Pro, Business or Enterprise, where it appears in the model picker. The OpenAI API, using the model identifier `gpt-6-astra`. Amazon Web Services, for teams already routing models through AWS. The earliest access went to companies accepted into OpenAI's cybersecurity-focused Trusted Access Program, which is an application, not a purchase. Sources: CNBC, OpenRouter.

Because the capability that makes Astra useful is the same capability that makes it worrying. A model that scores 100% on ExploitBench is, stated plainly, very good at finding software vulnerabilities, which cuts both ways. OpenAI's announcement gave unusual space to alignment and misuse, and it staged the rollout through a vetted cybersecurity program rather than opening the gates. Independent scrutiny of frontier releases has grown sharper through 2026, and the framing reflects that. Source: Al Jazeera.

It depends on what you run. If your workload is short prompts and simple outputs, the older, cheaper models still do the job and Astra's per-token price is hard to justify. If you run agents that work for many steps, read large documents, or write and debug code, the long-context and computer-use gains are where the difference shows up. The pragmatic move is to route by task rather than migrating everything at once.