GPT-6 Astra

GPT-6 Astra is OpenAI's September 4, 2026 flagship. The company calls it its most intelligent and most aligned model yet, aimed at long, end-to-end work: computer use, browsing, software engineering, science, and professional documents. The API id is gpt-6-astra. Standard pricing is $10 per 1M input tokens and $50 per 1M output tokens. Context is 1,050,000 tokens, with up to 922,000 input tokens and 128,000 output tokens. Knowledge cutoff is April 30, 2026.

Compare GPT-5.6 Sol if you want the previous OpenAI flagship at $5/$30, or Claude Fable 5 if you want Anthropic's competing frontier SKU.

Key Features

• Computer use as the headline: OpenAI reports state-of-the-art scores on Agents' Last Exam (59.3%), OSWorld 2.0 offline (72.6% at about 40 minutes per task vs 65.7% / 75 minutes for GPT-5.6 Sol), and ScreenSpot-Pro without tools (92.7%).
• Reasoning effort without none: low, medium, high, xhigh, and max. Astra does not support none. Fast mode is unavailable for GPT-6 Astra with EU data residency.
• Async tools and mid-turn steering: function tools can run async: true while the model keeps working; WebSocket Responses can inject a correction mid-turn without dropping the original task.
• Alignment claims you should treat as vendor evals: OpenAI says Astra went beyond an authorized target in 0% of an internal "impossible task" eval, vs 48% for GPT-5.6 Sol without production safeguards. Production ChatGPT and Codex still wrap extra monitors.
• Cyber capability with a restricted launch: ExploitBench 100% in the research setup, and OpenAI places Astra at the Critical cybersecurity threshold in its Preparedness Framework. The shipping product refuses advanced offensive work such as proof-of-concept exploits; Daybreak is the path OpenAI describes for less-restricted defensive access.

Model Specifications

Pricing

From OpenAI's model page, Standard text tokens:

Prompts with more than 272K input tokens are billed at 2x input and cache rates and 1.5x output for the full request. Batch and Flex are 50% of Standard. Fast mode is 2x Standard. OpenRouter listed the same Standard rates on 2026-09-05, plus Flex at $5/$25.

ChatGPT Plus, Pro, Business, and Enterprise were rolling out in the days after launch. Sam Altman publicly called the rollout messy and said Pro subscribers would go first. Astra usage sits inside existing subscription allowances; extra credits are sold on top. Enterprise admins enable Astra (off by default at launch). Pro, Business, and Enterprise also get GPT-6 Astra Pro.

Use Cases

• Computer-use agents: forms, CRM updates, calendar, research-to-doc, frontend QA, and installing software on screen.
• Professional artifacts: decks, spreadsheets, and documents that follow an existing template.
• Agentic coding: Codex, Responses API tool loops, and repository work. OpenAI reports Terminal-Bench 4.0 at 57.9% vs 37.3% for GPT-5.6 Sol.
• Science and math: OpenAI reports FrontierMath Tier 4 (v2) at 97.6% and ARC-AGI-3 at 99.9% in its research harness. Treat those as first-party scores.

Limitation: per-token price is higher than Sol. OpenAI argues cost per task can still fall because Astra uses fewer output tokens, but you should measure that on your own traces.

Getting Started

1. Set model to gpt-6-astra on the Responses API. Tool calling needs Responses; Chat Completions still works for plain text.
2. If you used none or minimal on older models, start at low.
3. Remove temperature, top_p, and top_logprobs.
4. When migrating from GPT-5.5 or earlier, replace prompt_cache_retention with prompt_cache_options.ttl set to "30m".

Frequently Asked Questions

Is it in ChatGPT yet?

OpenAI said it would roll out to Plus, Pro, Business, and Enterprise over the days after September 4, 2026. Availability still depends on your plan and workspace admin.

Does Fast mode work everywhere?

No. Fast mode is 2x Standard price and is not available for GPT-6 Astra with EU data residency.

Should I drop GPT-5.6 Sol?

Keep Sol if you want the $5/$30 SKU or a calmer reasoning profile. Astra is the new top of the stack, not a rename of Sol.

Alternatives

• GPT-5.6 Sol: previous OpenAI flagship, cheaper per token.
• Claude Fable 5: Anthropic frontier coding and computer use.
• Claude Mythos 5.1: Anthropic's less-restricted sibling if your job is cyber eval, not ChatGPT.

Tips

1. Prompt Astra to finish authorized work before asking questions. The model asks more than Sol did.
2. Audit AGENTS.md and skills. Astra follows file instructions more strictly and can pause on conflicts.
3. Watch the 272K surcharge. Cache the stable prefix.

Conclusion

GPT-6 Astra is OpenAI's new flagship for computer use, coding, and long professional jobs, at $10/$50 and 1.05M context. Start from the official announcement and the API model page, then measure cost per task against Sol before you cut over.