OpenAI has officially announced GPT-6 Astra, its flagship model for complex reasoning, coding, computer use, research and professional work. This page reports what OpenAI has published, and keeps our own analysis clearly separate from it.
Specifications verified against OpenAI’s official documentation on 19 September 2026.
GPT-6 Astra is OpenAI’s officially released flagship model, available through ChatGPT, the OpenAI API, Microsoft Azure and AWS Bedrock under the model ID gpt-6-astra. OpenAI describes it as its most capable model, built for the hardest end-to-end work — complex reasoning, coding, computer use, research and document creation.
Confirmed specifications: a 1,050,000-token context window, 128,000 maximum output tokens, a knowledge cutoff of 30 April 2026, and reasoning.effort levels of low, medium, high, xhigh and max. Standard API pricing is $10.00 per million input tokens and $50.00 per million output tokens, with cached input at $1.00 and cache writes at $12.50. Every figure on this page comes from OpenAI’s own announcement and model documentation, linked in the sources section.
| Specification | GPT-6 Astra |
|---|---|
| Status | Officially announced and rolling out |
| Model ID | gpt-6-astra |
| Context window | 1,050,000 tokens |
| Maximum output | 128,000 tokens |
| Knowledge cutoff | 30 April 2026 |
| Reasoning effort | low, medium, high, xhigh, max |
| Input price | $10.00 per 1M tokens |
| Cached input | $1.00 per 1M tokens |
| Cache writes | $12.50 per 1M tokens |
| Output price | $50.00 per 1M tokens |
| Modalities | Text input and output; image input only. Audio and video not supported |
| Tools (Responses API) | Web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, tool search |
| Other features | Function calling, structured outputs and streaming supported; fine-tuning not supported |
OpenAI announced GPT-6 Astra in September 2026. In its announcement OpenAI stated the model was “rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API, Microsoft Azure, and AWS Bedrock.”
The documented availability channels are therefore:
gpt-6-astra.Because a staged rollout was explicitly described, access on any particular plan, region or account may differ from the announcement at any given moment. Check OpenAI’s own documentation for what is available to you rather than assuming universal access.
OpenAI positions computer use as a headline capability, calling Astra “the world’s best computer use model” and describing “a new frontier in the speed, accuracy, and safety of computer use.” The published examples include filling out online forms, updating customer records in a CRM, organising a calendar, conducting online research and drafting summaries in an email or document editor, analysing scientific data and generating plots, creating a website and running frontend QA checks, and installing and testing software.
OpenAI states plainly that “GPT-6 Astra is the best model for software engineering to date.” On Terminal-Bench 4.0, which tests agents on terminal-based software engineering, system configuration and data analysis tasks, OpenAI reports 57.9% for Astra against 37.3% for GPT-5.6 Sol and 55.8% for Claude Fable 5.1.
Alongside the model, OpenAI describes a change to how Codex handles long sessions: rather than repeatedly compacting history into a summary, Astra can keep notes across context windows while earlier context remains searchable. OpenAI describes this as an experimental feature enabled in the Codex configuration, becoming the default in the following weeks.
GPT-6 Astra exposes a reasoning.effort setting with five levels:
Because reasoning tokens are billed, effort is a direct quality-versus-cost control. OpenAI’s own benchmark notes repeatedly cite results “at a lower-cost setting” that still beat the previous flagship’s best score, which suggests the lower levels are worth testing before defaulting to the top of the range.
OpenAI describes targeted training for professional environments, combining multistep workflows with the production of “polished documents, spreadsheets, and presentations.” It highlights template adherence — matching your writing and visual style and following existing formats — and says Astra is trained to pull only relevant context into outputs rather than repeating unnecessary information.
On BenchCAD, which tests reconstructing 3D objects from multi-view renders by generating CAD code, OpenAI reports a 95.9% geometric-overlap score with tools, against 83.3% for GPT-5.6 Sol and 84.3% reported for Claude Fable 5.1. For scientific reasoning, OpenAI reports 96.0% on GPQA Diamond.
The figures below are published by OpenAI. They are real, specific numbers, and they are also vendor-run under conditions OpenAI selected.
| Benchmark | GPT-6 Astra | Comparison published by OpenAI |
|---|---|---|
| FrontierMath Tier 4 (v2) | 98% | Described by OpenAI as saturating the benchmark |
| ARC-AGI-3 | 99.9% | ARC Prize reports parity with its human action-efficiency baseline on 96% of levels |
| ExploitBench | 100% | 78.5% for GPT-5.6 Sol |
| ExploitGym | 42.4% | 30.3% for GPT-5.6 Sol |
| GPQA Diamond | 96.0% | At a lower-cost setting, 94.9% vs GPT-5.6 Sol’s best of 94.6% |
| Terminal-Bench 4.0 | 57.9% | 37.3% for GPT-5.6 Sol; 55.8% for Claude Fable 5.1 |
| Terminal-Bench Science 0.1 | 64.6% | 52.6% for Claude Fable 5.1 |
| Agents’ Last Exam | 59.3% | 55.5% for Claude Opus 5; 53.6% for GPT-5.6 Sol |
| BenchCAD (with tools) | 95.9% | 83.3% for GPT-5.6 Sol; 84.3% reported for Claude Fable 5.1 |
| OSWorld 2.0 (latency sim) | 72.6% at ~40 min/task | 65.7% at ~75 min/task for GPT-5.6 Sol |
OpenAI’s published standard rates, per million tokens:
| Token type | Price per 1M tokens |
|---|---|
| Input | $10.00 |
| Cached input | $1.00 |
| Cache writes | $12.50 |
| Output | $50.00 |
Pricing is usage-based, so what you pay depends on how many tokens you send and receive. Three published rules change the effective rate:
Current rates are published on OpenAI’s GPT-6 Astra model page. For how this compares across providers, see our AI model API pricing comparison.
A factual comparison of the current flagship against OpenAI’s previous one. Both values in every row come from each model’s own OpenAI documentation page.
| Feature | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| Positioning | Latest flagship — “hardest end-to-end work” | Previous flagship — “complex professional work” |
| Model ID | gpt-6-astra | gpt-5.6-sol (the gpt-5.6 alias routes here) |
| Context window | 1,050,000 tokens | 1,050,000 tokens |
| Max output | 128,000 tokens | 128,000 tokens |
| Knowledge cutoff | 30 April 2026 | 16 February 2026 |
| Reasoning effort | low, medium, high, xhigh, max | none, low, medium (default), high, xhigh, max |
| Input price | $10.00 / 1M | $4.00 / 1M |
| Cached input | $1.00 / 1M | $0.40 / 1M |
| Output price | $50.00 / 1M | $20.00 / 1M |
Two details worth noting rather than glossing over. First, GPT-5.6 Sol supports a none reasoning-effort level that Astra does not, which matters if you have a latency-sensitive path that wants no reasoning tokens at all. Second, OpenAI describes Sol’s pricing as promotional, available at least through 21 November 2026 — so the 2.5x price gap shown above reflects a promotional rate on one side and may narrow.
We are not declaring a winner. The two models share a context window and output ceiling; they differ on price, knowledge cutoff, effort range and capability, and which of those matters is a property of your workload.
Cases where the capability is most likely to justify the price:
The counterweight is cost. At $10 input and $50 output per million tokens, Astra is 2.5x the current token price of GPT-5.6 Sol, and far more than the smaller models in OpenAI’s lineup. For classification, routing, simple extraction or high-volume chat, a cheaper model will often produce the same outcome for a fraction of the bill.
The right measure is cost per completed task, not cost per token. A more expensive model that finishes correctly on the first attempt can be cheaper overall than a cheap one that needs several attempts plus human review — and on easy work the reverse is just as often true. GPT-6 Astra is not automatically the right default for every workload, and OpenAI does not claim it is.
Benchmarks establish that a model is in the right capability class. They do not tell you whether it works for what you are building. A short, repeatable evaluation on your own tasks settles that in an afternoon:
For a cross-vendor view, see our comparison of Claude, GPT, Gemini and the leading Chinese models, the per-token pricing comparison, or our explainer on Claude Fable 5.1, which appears in several of OpenAI’s published comparisons above.
While the frontier moves, getting your business online does not have to wait. XAIWebsites generates a complete business website — copy written for your business, mobile-ready layout, and hosting — from a short questionnaire.
GPT-6 Astra is OpenAI's flagship model, which OpenAI describes as its most capable model, built for the hardest end-to-end work. OpenAI lists it for complex reasoning, coding, computer use, research and document creation, and describes it as state-of-the-art on computer use, browsing, software engineering, cybersecurity, science and professional work.
Yes. OpenAI announced GPT-6 Astra in September 2026. OpenAI stated it was rolling out first to a limited set of organizations, and over the following days would become available to all ChatGPT Plus, Pro, Business and Enterprise users, as well as through the OpenAI API, Microsoft Azure and AWS Bedrock.
The API model identifier is gpt-6-astra. OpenAI's model documentation lists gpt-6-astra as both the model name and its current snapshot and alias, so the same string works for direct calls.
OpenAI's model documentation lists a 1,050,000-token context window and a maximum of 128,000 output tokens. Note that prompts above 272,000 input tokens are billed at 2x input and cache rates and 1.5x output for the full request, so a very large context has a pricing consequence as well as a capability one.
OpenAI's published standard rates are $10.00 per million input tokens, $1.00 per million cached input tokens, $12.50 per million cache-write tokens, and $50.00 per million output tokens. Batch and Flex are priced at 50% of standard rates and Fast mode at 2x. Pricing is usage-based, so the bill depends on your token volume.
OpenAI's model documentation lists a knowledge cutoff of 30 April 2026. For anything after that date the model needs the information supplied in context, or a tool such as web search, rather than relying on its training data.
OpenAI documents that reasoning.effort supports low, medium, high, xhigh and max on GPT-6 Astra. Higher effort generally means more thorough reasoning and more tokens consumed, so the setting is a direct quality-versus-cost control rather than a simple on/off switch.
OpenAI positions it for complex reasoning, coding, computer use, research and document creation. Its documentation lists support for function calling, structured outputs and streaming, and tools including web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search. Fine-tuning is listed as not supported.
Both list a 1,050,000-token context window and 128,000 max output tokens. Astra costs $10 input and $50 output per million tokens against Sol's $4 and $20, and has a later knowledge cutoff (30 April 2026 versus 16 February 2026). Sol additionally supports a 'none' reasoning-effort level that Astra does not. OpenAI notes Sol's pricing is promotional at least through 21 November 2026.
Through the OpenAI API using the model ID gpt-6-astra, and via Microsoft Azure and AWS Bedrock. In ChatGPT, OpenAI stated availability for Plus, Pro, Business and Enterprise users following an initial limited rollout. Check OpenAI's own documentation for current availability in your region and plan.
Every specification, price and benchmark figure on this page comes from one of the following OpenAI pages. Where we add interpretation, it is marked as our own.
Updated 19 September 2026. Specifications, pricing and availability were verified against OpenAI’s official documentation on that date; all three change over time, so confirm against OpenAI’s own pages before relying on any figure. See our methodology.