Try GPT-6 Astra — The Future Is Here
OpenAI's most intelligent and aligned model yet — welcomed by many as the opening of the AGI era. It matches human experts on frontier benchmarks, operates computers like people do, and writes production-grade software.
How to try GPT-6 Astra in minutes
Four verified deployment routes: consumer web, terminal & IDE coding workflows, API integration, and enterprise cloud.
ChatGPT
- 1.Open model picker in ChatGPT Web or Desktop app
- 2.Select GPT-6 Astra (rolling out to all paid tiers)
- 3.Pro & Enterprise plans also unlock GPT-6 Astra Pro
Codex (CLI & IDE)
- 1.Terminal agent: curl -fsSL https://chatgpt.com/codex/install.sh | sh
- 2.Run: codex -m gpt-6-astra or set model in ~/.codex/config.toml
- 3.Also in VS Code, Cursor, Windsurf & JetBrains extension sidebar
OpenAI API
- 1.Call the Responses API with model "gpt-6-astra"
- 2.Set reasoning effort: low, medium, high, xhigh or max
- 3.$10/$50 per 1M tokens up to 272K context; cached input $1
Azure & AWS Bedrock
- 1.Available on Microsoft Azure OpenAI and AWS Bedrock
- 2.Same model ID and pricing structure as OpenAI API
- 3.Enterprise access off by default — admin must enable
Benchmarks & the AGI debate
ARC-AGI-3 human parity, software engineering state-of-the-art, and what evaluators are actually saying.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol | Competing Frontier |
|---|---|---|---|
| ARC-AGI-3 (adapter harness) | 99.9% | 7.8% | 30.2%Claude Opus 5 |
| FrontierMath Tier 4 | 97.6% | 83.0% | — |
| ExploitBench | 100% | 78.5% | — |
| SRE-Bench (single try) | 88.0% | 55.9% | — |
| Terminal-Bench 4.0 | 57.9% | — | 55.8%Fable 5.1 |
| GPQA Diamond | 96.0% | 94.6% | — |
| Agents' Last Exam | 59.3% | — | 55.5%Claude Opus 5 |
“The best model we've ever tested.”
— Greg Kamradt, ARC Prize Foundation
“The story is: end of one era, start of another.”
— Greg Burnham, Epoch AI
“A noticeable step-function change in frontier model capabilities.”
— ARC Prize team, on GPT-6 Astra's ARC-AGI-3 run
The boundary has shifted, but open-endedness remains
- •OpenAI perspective: Marketed as "a new generation of intelligence" and its closest step toward AGI to date.
- •ARC Prize perspective: Succeeded in demonstrating human parity on deterministic interactive levels, but explicitly notes this is not a proof of real-world open-ended AGI.
- •The consensus: A clear step-function leap in computer use and reasoning efficiency.
Computer use
Astra sees the screen and works like a person: filling forms, updating CRMs, running frontend QA and installing software — finishing OSWorld tasks in about 47% less time than GPT-5.6 Sol.
Best-in-class coding
Described by OpenAI as the best software engineering model to date. A new Codex integration keeps searchable notes across context windows instead of compressing them away.
Science & math
Astra helped establish a new bound of 186 for short prime gaps — improving a term in the mathematics that had stood unchanged for over 80 years.
Million-token memory
Reads entire codebases and book-length corpora in one pass, scoring 96.3% on MRCR v2 (8-needle) across 512K–1M-token evaluations.
Agentic workflows
Plans and executes long professional tasks end to end. Partners report complex creative workflows completing with up to 20% fewer tokens than other frontier models.
Alignment & safety
On an impossible-task eval, Astra violated its scope 0% of the time vs 48% for GPT-5.6 Sol. During testing it discovered two real zero-days — and disclosed them to maintainers.
Token-saving guide & reasoning effort ladder
Astra's $10/$50 token pricing makes configuration discipline essential. Here is how production teams cut per-task costs by up to 60%.
| Tier | Score | Cost / Task | When to Use |
|---|---|---|---|
| low | 49 | $0.63 | Quick answers and easy edits — already outperforms GPT-5.6 Sol at high effort |
| mediumSWEET SPOT | 52 | $1.16 | The default for most work — beats GPT-5.6 Sol at max effort, for less money |
| high | 53 | $1.41 | Long autonomous runs (~20 min): Codex sessions, browser agents, tool pipelines |
| xhigh | 54 | $1.85 | Only after measuring a real failure rate at high effort |
| max | 55 | $2.57 | Hardest problems only — otherwise you're paying for a feeling, per benchmarks |
The 6 Rules of GPT-6 Astra Token Optimization
Actionable practices compiled from production deployments and developer benchmarks.
Keep prompt prefixes stable
Cached input costs $1 per million tokens — 90% off the $10 list price. Put fixed rules and system prompts first, dynamic content (paths, timestamps, per-task details) last.
Only cache what repeats
Cache writes cost $12.50/M. A cached block needs to be read at least twice to break even — one-off requests are cheaper uncached.
Stay under the 272K line
Once input passes 272K tokens, the entire request bills at long-context rates ($20/$75 per 1M) — not just the overflow. Trim context before you cross it.
Match effort to the task
Medium is the default; low is fine for simple Q&A; reserve high+ for long autonomous runs where fewer wasted steps pays for the pricier thinking.
Use Batch & Flex for async work
Both modes cost 50% of standard rates. If a job doesn't need a real-time answer — backfills, evaluations, bulk processing — don't pay full price.
Watch output length
Output tokens cost $50/M — 5× input. Ask for concise formats, cap max output tokens, and have Astra summarize instead of reproduce long context.
gpt-6-astra
GPT-6 Astra Pro tier included with Pro / Business / Enterprise
1,050,000 tokens
128,000 max output tokens
Text + images → text
Reasoning model with adjustable effort (low → max)
April 30, 2026
$10 / $50 per 1M tokens
Input / output up to 272K context · Batch & Flex −50% · Fast mode 2×
ChatGPT · API · Azure · AWS Bedrock
Rolling out to Plus, Pro, Business and Enterprise plans
Frequently asked questions & primary documentation
Direct answers to operational queries alongside links to official deployment cards and independent benchmark reports.
The official homepage and launch post for GPT-6 Astra.
Frontier safety update: critical capabilities and safeguards behind the launch.
Deployment safety hub: evaluations, limitations and monitoring notes.
Independent evaluation: human parity on action efficiency, harness details and cost.
Full technical specifications: context, pricing tiers and modalities.
Chinese coverage: "OpenAI 发布地球最强大模型 GPT-6，宣告欢迎来到 AGI 时代".
Ready to explore GPT-6 Astra?
OpenAI's most capable system is accessible today across ChatGPT Plus, Pro, Enterprise, and the Responses API.
Independent technical dossier published by Findry AI. GPT-6 Astra is a product of OpenAI.