Availability (staged rollout)
Source: OpenAI announcement (GPT-6 Astra: A new generation of intelligence) (verified 2026-09-04)
Pricing (Standard $10/$50, verified 2026-09-04)
Verified against OpenAI's official model docs (developers.openai.com). All prices in USD per 1M tokens.
| Tier | Input | Cached Input | Cache Writes | Output | Notes |
|---|---|---|---|---|---|
| Standard (Short context) | $10.00 | $1.00 | $12.50 | $50.00 | ≤272K input tokens. Default rate |
| Standard (Long context) | $20.00 | $2.00 | $25.00 | $75.00 | >272K input tokens. 2× input, 1.5× output |
| Batch / Flex | $5.00 | $0.50 | $6.25 | $25.00 | 50% off Standard |
| Fast mode (Short context) | $20.00 | $2.00 | $25.00 | $100.00 | Up to 2× speed. 2× Standard (Long context applies the Fast multiplier to long-context rates) |
Primary sources: GPT-6 Astra Model Docs · developers.openai.com/api/docs/pricing (verified 2026-09-04)
Model Specifications
Sources: OpenAI GPT-6 Astra Model Docs (verified 2026-09-04). reasoning.effort supports low / medium / high / xhigh / max. Audio/video not supported (image input only). Responses API supports Web search, File search, Image generation, Code interpreter, Hosted Shell, Apply Patch, Skills, Computer Use, MCP, and Tool search.
Performance — New Records Across Key Benchmarks
OpenAI positions GPT-6 Astra as state-of-the-art in computer use, browsing, software engineering, cybersecurity, science and professional work. Figures below are from the official announcement (2026-09-03).
- FrontierMath Tier 4 (v2): 97.6% — new high in mathematics (GPT-5.6 Sol 83.0%, Claude Fable 5.1 87.8%).
- ARC-AGI-3: 99.9% — human-parity abstract reasoning (Sol 7.8%). ARC Prize Foundation: "effectively reaching human parity."
- ExploitBench: 100.0% — perfect score in exploit development (Sol 78.5%).
- Terminal-Bench Science 0.1: 64.6% — scientific workflows (Sol 22.4%, Claude Fable 5.1 52.6%).
- Agents' Last Exam: 59.3% — complex professional tasks (Claude Opus 5 55.5%, Sol 53.6%).
- OSWorld 2.0: 72.6% — computer use (Sol 65.7%), in ~47% less time per task.
- Terminal-Bench 4.0: 57.9% — terminal-based agents (Sol 37.3%, Claude Fable 5.1 55.8%).
- GPQA Diamond: 96.0% — graduate-level science reasoning (Sol 94.6%).
- Artificial Analysis Intelligence Index v4.1.1: 61.2 — composite index.
OpenAI ran many benchmarks at maximum effort. At lower-cost settings Astra still beats Sol's best on some tasks (e.g. GPQA 94.9% vs 94.6% at ~37% lower estimated API cost). Benchmarks are published figures and do not guarantee real-world experience.
Source: OpenAI announcement (GPT-6 Astra) (verified 2026-09-04)
Competitor Price Comparison
All prices in USD per 1M tokens (standard, short context). GPT-6 Astra sits in the top pricing tier.
| Model | Provider | Input | Output | vs Astra | Notes |
|---|---|---|---|---|---|
| GPT-6 Astra | OpenAI | $10.00 | $50.00 | — | New flagship (baseline) |
| Claude Fable 5.1 | Anthropic | $10.00 | $50.00 | Same | Mythos-class. Detail page |
| Claude Opus 5 | Anthropic | $5.00 | $25.00 | Opus cheaper | 50% cheaper input & output |
| GPT-5.6 Sol | OpenAI | $4.00 | $20.00 | Sol cheaper | Promo (thru 11/21). Prior flagship |
| GPT-5.5 | OpenAI | $5.00 | $30.00 | 5.5 cheaper | Prior flagship |
| Grok 4.6 | xAI | $2.00 | $6.00 | Grok cheaper | Cheapest flagship-class in absolute terms |
| Gemini 3.8 Flash | $0.75 | $3.75 | Gemini cheaper | Flash tier (different class). Intro price (thru 12/31) |
* GPT-6 Astra sits in the top pricing tier. For cost-sensitive workloads, Sol ($4/$20 promo) or Gemini 3.8 Flash ($0.75/$3.75) are strong alternatives. OpenAI argues Astra uses fewer output tokens per task, but published per-task data is limited.
Sources: official pricing pages from each provider (verified 2026-09-04). See full pricing comparison.
Ideal Use Cases
High-failure-cost workloads that need top-tier reasoning and computer use.
Computer Use & Browsing
Filling forms, updating CRMs, organizing calendars — GUI automation. OSWorld 72.6%.
Software Engineering
57.9% on Terminal-Bench 4.0. Long-horizon coding and refactoring with Codex.
Scientific Research & Math
FrontierMath 97.6%, GPQA 96.0%. Assisted solving open problems in mathematics.
Cybersecurity (defensive)
ExploitBench-level capability. Secure code review and patching (advanced offensive tasks refused).
Professional & Document Work
Template-faithful slides, spreadsheets and documents. BenchCAD 95.9%.
Long-horizon Autonomous Agents
Multi-step automation with MCP, computer use and tool calling.
OpenAI Flagship Selection Guide
| Model | Input | Output | When to Use |
|---|---|---|---|
| GPT-6 Astra ⬅ | $10.00 | $50.00 | Hardest tasks, computer use, science, high-failure-cost work |
| GPT-5.6 Sol | $4.00 | $20.00 | High-accuracy reasoning at lower cost (promo thru 11/21) |
| GPT-5.6 Terra | $2.00 | $12.00 | Standard coding, analysis, production agents |
| GPT-5.6 Luna | $0.20 | $1.20 | High-throughput, cost-sensitive, routine processing |
💡 Strategy: escalate from Luna → Terra → Sol, reserving GPT-6 Astra for the hardest tasks, computer use and scientific research.
Estimate Your Actual Costs
Compare token costs for GPT-6 Astra and all major models. Just enter your token counts.
🧮 Open Token Calculator →📊 Full Pricing Comparison · ⚡ Compare with GPT-5.6 Sol · 🗓️ Release Timeline · 🔗 OpenAI Official Pricing
⚠️ Disclaimer
- Pricing verified against OpenAI's official page (developers.openai.com) on 2026-09-04. Subject to change without notice.
- As of 2026-09-03, GPT-6 Astra rollout is limited to a set of organizations; API/ChatGPT availability follows "over the coming days." It may not yet be broadly available.
- Benchmark scores are OpenAI's published figures and do not guarantee real-world experience.
- This site is informational only. Not affiliated with any provider.