What is GPT-6 Astra?
GPT-6 Astra is the top model in OpenAI's GPT-6 family, above GPT-6 Sol and GPT-6 Luna. It reads text and images, writes text, and works across a 1.05M-token context with up to 128K tokens of output; its knowledge stops at April 2026. OpenAI built it for jobs that run from a request to a finished result across code, browsers and professional software, and says it reaches better results with fewer output tokens than earlier models.
It is OpenAI's current flagship. Developers reach it on the OpenAI API and Amazon Bedrock, and ChatGPT Plus, Pro and Business users get it in Codex and ChatGPT Work.
What changed from GPT-5.6 Sol?
More control over long agent runs, a more careful working style, and a few breaking API changes.
- New API controls
- Async tool calling, mid-turn steering, and reasoning changes mid-conversation that keep the prompt cache.
- Working style
- OpenAI says it stays coherent longer on long tasks, follows instructions more closely, and asks where earlier models guessed.
- Breaking changes
- No none effort, no custom temperature, top_p or logprobs, and tool calls need the Responses API.
- Safety checks
- Misalignment monitoring reviews agent work in the Responses API and can stop a conversation for review.
Who is GPT-6 Astra best for?
GPT-6 Astra is best for long end-to-end jobs, very large inputs, agents that use a computer, and paid ChatGPT subscribers who want OpenAI's strongest model.
Long end-to-end jobs
OpenAI points to Astra for complete workflows across code, apps and research that need reasoning over many steps and tools.
Very large inputs
It reads about 1.05 million tokens of context and can write up to 128,000 tokens in one reply, though prompts over 272K tokens cost more.
Agents that use a computer
Through the Responses API it supports computer use, a hosted shell, skills, MCP servers, and web and file search.
Paid ChatGPT subscribers
It is included from the $20-a-month Plus plan in Codex and ChatGPT Work, so you can try it before paying API rates.
What's included with GPT-6 Astra?
Five effort levels, async tool calling, mid-turn steering, effort changes that keep the cache, a full set of hosted tools, cheaper Batch and Flex processing, and Amazon Bedrock.
- Five effort levels
- Reasoning runs from low to max; there is no none setting, and ChatGPT suggests starting at Light, its name for low.
- Async tool calling
- Astra keeps reasoning or answers other parts of a request while your application runs a slow tool.
- Mid-turn steering
- You can send a correction over a WebSocket while it works, and the Responses API keeps the work already done.
- Effort changes that keep the cache
- A configuration_update item raises or lowers reasoning effort mid-conversation without rewriting the cached prompt.
- Hosted tools
- Web and file search, code interpreter, a hosted shell, apply patch, skills, computer use, MCP and tool search all work on the Responses API.
- Batch, Flex and Fast mode
- Batch and Flex cost half the standard rate for work that can wait; Fast mode costs double.
- Amazon Bedrock
- Astra is in Bedrock's catalog, and OpenAI says Bedrock's price in commercial regions matches its own.
How much does GPT-6 Astra cost?
GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens on the OpenAI API, and in ChatGPT it comes with Plus at $20 a month and higher paid plans, not with Free or Go.
- Cached input $1, cache writes $12.50 per million tokens
- Prompts over 272K tokens: $20 / $75 for the whole request
- Batch and Flex at half price, Fast mode at double
- About 5 to 45 GPT-6 Astra messages every five hours in Codex and ChatGPT Work
- Buy ChatGPT credits when you run out
- Same GPT-6 Astra allowance as Plus
- Admin controls, SAML SSO and MFA
- No training on business data by default
- 5x or 20x the usage of Plus
- About 25 to 225, or 100 to 900, GPT-6 Astra messages every five hours
On the API, Batch and Flex cost $5 / $25 per million tokens and Fast mode $20 / $100; data-residency endpoints add 10%, and the free API tier does not include GPT-6 Astra. ChatGPT Business is $20 per user per month billed yearly, for 2 or more users; Enterprise and Edu are priced by sales. ChatGPT Free ($0) and Go ($8 a month) get GPT-6 Luna instead of Astra. The message counts are OpenAI's estimates for Codex and ChatGPT Work, not fixed limits, and the page does not list the monthly price of the Pro 20x tier. Full pricing on developers.openai.com. We checked these prices on .
How do readers rate GPT-6 Astra?
No one has rated GPT-6 Astra yet. Be the first.
Sign in to rate. One rating per account.
GPT-6 Astra pros and cons
It finishes hard jobs in fewer tokens than earlier OpenAI models, but it costs five times as much as GPT-6 Sol and drains ChatGPT Plus limits fast.
Pros
- Needs fewer tokens per task than earlier OpenAI models, and several early testers found its cost per result close to GPT-5.6 Sol's despite the higher token price.
- Early Hacker News and Reddit users say it found and fixed bugs that GPT-5.6 Sol, Terra and Grok 4.6 had missed.
- Strong at visual work: users praise pages rebuilt from design images, detailed SVG drawings and 3D scenes in Blender and Three.js.
- Several users say its writing reads more naturally than earlier GPT models, with fewer stock AI phrases.
Cons
- Drains ChatGPT limits fast: OpenAI estimates 5 to 45 Astra messages every five hours on Plus, and many Plus users report one task using a whole five-hour allowance.
- At $10 / $50 per million tokens it costs five times GPT-6 Sol, and a prompt over 272K tokens bills the whole request at $20 / $75.
- OpenAI says it asks for clarification more often than earlier models, and some early users had to keep telling it to carry on or found it slower per step than GPT-5.6 Sol.
- Older API code can break: there is no none effort, no custom temperature, top_p or logprobs, tool calls need the Responses API, and misalignment monitoring can stop a conversation for good.
What are the best GPT-6 Astra alternatives?
Claude Fable 5.1 for long prompts at the same price, Claude Opus 5.5 for a lower bill, and Muse Spark 1.3 for the cheapest 1M-token context.
Claude Fable 5.1PaidChoose it when your prompts run long: it costs the same $10 / $50 per million tokens but bills its whole 1M-token context at that rate, and cache reads cost $0.25 against Astra's $1.
Claude Opus 5.5PaidChoose it when cost matters more than peak capability: it charges $4 / $20 per million tokens for a 1M-token context and 128K output, and Anthropic recommends it for most work.
Muse Spark 1.3PaidChoose it when you want a cheap long-context model: $1.25 / $4.25 per million tokens on Meta's API with a 1M-token context, and it also reads video and PDFs.
Questions about GPT-6 Astra
What is GPT-6 Astra?
GPT-6 Astra is OpenAI's most capable model, built for long jobs across code, research, computer use and documents, and sold on the API and in paid ChatGPT plans.
Who is GPT-6 Astra for?
GPT-6 Astra is best for long end-to-end jobs, very large inputs, agents that use a computer, and paid ChatGPT subscribers who want OpenAI's strongest model.
Is GPT-6 Astra free?
No. ChatGPT Free and Go use GPT-6 Luna, and the OpenAI API's free tier does not include Astra. You need ChatGPT Plus at $20 a month or a higher plan, where it runs in Codex and ChatGPT Work, or paid API usage.
How much more does GPT-6 Astra cost than GPT-6 Sol?
GPT-6 Astra costs five times as much as GPT-6 Sol on the OpenAI API: $10 against $2 per million input tokens, and $50 against $10 per million output tokens. OpenAI suggests Sol for everyday and complex work and Astra when a task needs the most capability across many steps and tools.
What are the best GPT-6 Astra alternatives?
The closest alternatives on Recaply are Claude Fable 5.1, Claude Opus 5.5 and Muse Spark 1.3.
Where can I try GPT-6 Astra?
On its official website, openai.com.
Why does GPT-6 Astra use up my ChatGPT Plus limit so fast?
Astra uses five times as many ChatGPT credits per token as GPT-6 Sol, so OpenAI estimates only 5 to 45 Astra messages every five hours on Plus, against 15 to 150 for Sol. Starting at Light effort helps, and Plus and Pro users can buy extra credits.
Why does GPT-6 Astra keep stopping to ask me questions?
OpenAI designed Astra to ask when your answer could change the result, so it pauses more often than earlier models. OpenAI's guide suggests telling it in the prompt to infer your intent and carry the task to completion unless an action is destructive.
Can I use GPT-6 Astra on Amazon Bedrock?
Yes. GPT-6 Astra is in Amazon Bedrock's catalog, billed through AWS, and OpenAI says Bedrock's price in commercial regions matches its own. Codex and ChatGPT Work can also be set up to send Astra requests through Bedrock.
Our verdict
GPT-6 Astra earns its price on the hardest multi-step jobs, but GPT-6 Sol should stay your default.
Early users say it cracks bugs and visual tasks that other models miss and needs fewer tokens to get there, yet at $10 / $50 per million tokens it costs five times as much as GPT-6 Sol.
- Skip it for
- ChatGPT Plus users who want to run it all day, since OpenAI estimates only 5 to 45 Astra messages every five hours.
- Check first
- On the API, move tool calls to the Responses API and test your longest prompts, which cost double above 272K tokens.
GPT-6 Sol
Gemini 4 Carbon
Muse Wonder
GPT-5.4
GPT-5.5