What it is
GPT-6 Astra is OpenAI's top model, launched in early September 2026 alongside a higher-compute Astra Pro. OpenAI calls it its most capable model for the most demanding work: complex reasoning, coding, computer use, research and document creation. It reads text and images, writes text, and has a 1,050,000-token context window with 128,000 output tokens and an April 2026 knowledge cutoff.
It is a reasoning model with five effort levels: low, medium, high, xhigh and max. The API price is $10 per million input tokens and $50 per million output tokens, with cached input at $1 and cache writes at $12.50. Prompts over 272K input tokens are billed at twice the input rate and 1.5 times the output rate for the whole request. Batch and Flex are half price; Fast mode costs double.
OpenAI says Astra is the first model to cross the "Critical" cybersecurity threshold in its Preparedness Framework, and ships it with tighter access controls as a result. It is available through the OpenAI API and in ChatGPT on paid plans.
Our review
Astra is the model to pay for only when a problem sits at the very edge of what models can reason through, and those problems are rarer than its launch suggested. Artificial Analysis gives it 53 on its Intelligence Index, seventh of 214 models, one point above GPT-6.1 Sol (52) and five below Claude Opus 5.5 (58). Its cost per index task is $3.26 against $0.72 for Sol.
OpenAI's headline numbers need context. Astra's 97.6% on FrontierMath Tier 4 covers the 41 private problems in a 43-problem tier, and Epoch AI, which runs the benchmark, notes that OpenAI funded it and has exclusive access to part of it. The 98.6% on ARC-AGI-3 used a harness that keeps reasoning between turns, which The New Stack points out measures OpenAI's agent system as much as the model.
On coding it is good rather than leading. It scored 74.1% on DeepSWE v1.1 against 70.8% for Sol, and on Terminal-Bench 4.0 Anthropic's comparison table puts Astra at 57.9% to Opus 5.5's 66.4%, with Opus priced at $4 and $20 per million tokens. Artificial Analysis also measures Astra at about 56 output tokens a second, among the slower frontier models.
Operational details narrow the field further. EU data residency is unavailable with Fast mode, and Ultrafast is US-only, while Sol supports both regions in every mode. OpenAI says the Critical rating brings tighter safeguards, so security workloads should be tested for refusals before a migration. Astra's real advantage is in research-grade maths and science, where OpenAI reports 64.6% on Terminal-Bench Science against 52.6% for Claude Fable 5.1.
Where it earns its place
A quantitative research group at a hedge fund
Check proofs and derive models where a wrong step is expensive
Run Astra at xhigh or max effort through the Batch API at $5/$25 per million tokens and compare against Sol on the same problems. The gap is largest on competition-grade maths, so measure before committing a budget.
An engineering manager running 10,000 agent tasks a month
Decide between Astra and Sol for a coding agent
At 50K input and 5K output tokens per task, Astra costs about $7,500 a month and Sol about $1,500 at list rates. Sol scores within 3.3 points on DeepSWE, so Astra needs to fix tasks Sol cannot to earn the difference.
A biotech lab's computational team
Run long command-line analyses across scientific datasets
Astra's reported 64.6% on Terminal-Bench Science suits multi-step research scripts. Keep prompts under 272K tokens; above that the whole request is billed at 2x input and 1.5x output.
Before you commit
- Setup
- Available to all paid API tiers via the Responses and Chat Completions APIs; the model ID is gpt-6-astra and there is no free-tier access.
- Integrations
- Responses API tools include web and file search, code interpreter, hosted shell, apply patch, computer use, MCP and skills; function calling and structured outputs supported; no fine-tuning.
- Data and privacy
- OpenAI does not train on API data by default and keeps abuse-monitoring logs for up to 30 days; Zero Data Retention needs prior approval. EU data residency is supported, except with Fast mode.
- Cost at scale
- A 50K-input, 5K-output agent task costs about $0.75 (Sol: $0.15); 10,000 a month is $7,500, or $3,750 on Batch. Artificial Analysis measures $3.26 per index task.
- Limits
- Context 1.05M, output 128K; rate limits from 5,000 RPM and 1M TPM on the Build tier to 15,000 RPM and 40M TPM on Grow; long prompts over 272K tokens cost more; cyber requests face extra safeguards.
What's good
- 97.6% on FrontierMath Tier 4, per OpenAI
- 1.05M-token context, 128K output
- Five effort levels from low to max
- Batch and Flex at half price
What's not
- $3.26 per Artificial Analysis task, against $0.72 for GPT-6.1 Sol
- Behind Claude Opus 5.5 on Terminal-Bench 4.0 and the AA index
- Slow at about 56 tokens a second
- No EU residency in Fast mode; Ultrafast US-only
Best for
Research, maths and science teams that need the strongest reasoning OpenAI offers and can justify the price.
Alternatives
GPT-6 Astra: common questions
- Is GPT-6 Astra worth it?
- A record-setting maths and science model that costs four and a half times as much per task as a sibling scoring one point less. We rate it 4.2 out of 5.
- How much does GPT-6 Astra cost?
- API: $10 per 1M input tokens, $50 output; cached input $1; Batch/Flex half price; prompts over 272K input tokens cost 2x input and 1.5x output.
- What are the downsides of GPT-6 Astra?
- $3.26 per Artificial Analysis task, against $0.72 for GPT-6.1 Sol; behind Claude Opus 5.5 on Terminal-Bench 4.0 and the AA index; slow at about 56 tokens a second; no EU residency in Fast mode; Ultrafast US-only.
- Who is GPT-6 Astra best for?
- Research, maths and science teams that need the strongest reasoning OpenAI offers and can justify the price.
- How much does GPT-6 Astra cost for a team or in production?
- A 50K-input, 5K-output agent task costs about $0.75 (Sol: $0.15); 10,000 a month is $7,500, or $3,750 on Batch. Artificial Analysis measures $3.26 per index task.
- Is GPT-6 Astra safe for company data?
- OpenAI does not train on API data by default and keeps abuse-monitoring logs for up to 30 days; Zero Data Retention needs prior approval. EU data residency is supported, except with Fast mode.
- What are the best GPT-6 Astra alternatives?
- The closest alternatives we have reviewed are GPT-6.1 Sol, Claude Fable 5.1, Claude Opus 5.5 and Gemini 4 Argon.
Answers checked Oct 22, 2026.
Make GPT-6 Astra? Show this rating on your site
