Published September 8, 2026. OpenAI released GPT-6 Astra on September 3, 2026 — its most capable model to date, with major gains in computer use, coding, science, and CAD (OpenAI, 2026). On September 7, 2026, Nvidia CEO Jensen Huang tweeted 'AGI has arrived' in response to the launch.
Data last verified September 8, 2026 from OpenAI's announcement, BenchCAD scores, and developer demos.
Quick Answer
GPT-6 Astra is OpenAI's most capable model, launched September 3, 2026 (OpenAI, 2026). Major gains in computer use, coding, science, and CAD. Astra scored 95.9% on BenchCAD (vs 83.3% for GPT-5.6 Sol) — but only in an agentic setting. Pricing is $10 / $50 per million input/output tokens. Jensen Huang declared 'AGI has arrived' on launch day.
GPT-6 Astra — what changed vs GPT-5.6 Sol
| Capability | GPT-5.6 Sol | GPT-6 Astra |
|---|---|---|
| BenchCAD (agentic) | 83.3% | 95.9% |
| BenchCAD (plain) | Leader | No published score yet |
| Computer use | Strong | Significantly stronger (48-level CAPTCHA game demo) |
| Coding | Excellent | Improved for long-horizon tasks |
| Science reasoning | Excellent | Improved with better tool use |
| CAD / design | Capable | Capable + 95.9% BenchCAD (agentic) |
Source: OpenAI, BenchCAD project, 2026.
GPT-6 Astra — pricing
GPT-6 Astra costs (OpenAI, 2026):
- Input: $10 per million tokens (~$7.40 / £7)
- Output: $50 per million tokens (~$37 / £37)
- Cache reads: Discounted
A single 7-hour design run can burn a few hundred dollars in usage. For comparison, GPT-5.6 Sol is roughly $3 / $15 per million tokens — Astra is ~3x more expensive.
GPT-6 Astra — computer use demo (CAPTCHA game)
On September 7, 2026, a developer shared a video of GPT-6 Astra completing all 48 levels of Neal Agarwal's 'I Am Not a Robot' browser game (OpenAI / X, 2026):
- Astra used visual understanding + Computer Use capabilities to work through the challenges.
- It examined the browser interface, understood the objective, and performed actions (clicking, dragging, typing).
- The full game takes a human 30-60 minutes — Astra completed it autonomously.
- OpenAI Head of Applied Research Boris Power commented: 'True AGI can go everything a human can do on the computer.'
GPT-6 Astra — CAD / design caveat
The BenchCAD 95.9% score carries an important caveat (BenchCAD project, 2026):
- The 95.9% score was only achieved in an agentic setting — Astra could render, measure, and retry before submitting.
- In the plain (non-agentic) setting, Astra has no published score yet; GPT-5.6 Sol still leads.
- Engineers testing the tool warned that 95.9% geometric similarity can still produce the wrong part — one example had 96.1% similarity but a later feature closed an opening the design was meant to keep.
- A separate benchmark (neuralCAD-Edit) found the best model finished 53 percentage points behind professional designers on real edits.
Astra is impressive, but it sits above CAD software rather than replacing it — read instructions and write code while the CAD system holds the exact geometry.
GPT-6 Astra — safety signals
Two days after launch, OpenAI's chief scientist Jakub Pachocki issued a public warning (OpenAI / Pachocki, 2026):
'Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer. I expect and hope for voluntary slowdowns to become commonplace until shared safety bars are established. And I believe that international coordination on future AI development needs to become a top priority for governments around the world.'
The warning carries significant weight — Pachocki calls international coordination on future AI a top priority.
GPT-6 Astra — access
- API: platform.openai.com (OpenAI, 2026)
- ChatGPT: chat.openai.com (paying users get priority access)
- Rate limits: Per API plan tier; high-volume access requires the Pro / Enterprise tier.
Resources & next steps
For the official GPT-6 Astra announcement, API access, and pricing, go to openai.com and platform.openai.com. The BenchCAD project (github.com/benchnark) carries the public benchmark scores.
Verify current pricing, rate limits, and feature set on the official source: platform.openai.com (OpenAI, 2026).
Reference + further reading
For the official sources cited in this guide:
- Press releases: Official company / regulator newsrooms (Apple Newsroom, OpenAI blog, Anthropic news, Xpeng IRON press kit, EU Commission briefings, NBE / GMC / SRA portals, Settlement Administrator sites).
- Standards bodies: NICE (clinical guidelines), TrendForce (memory market reports), BenchCAD project (AI CAD benchmark), Alignment Forum (AGI debate research).
- Regulators: EU AI Office, US FTC, UK ICO, Norwegian Datatilsynet, NHS / GMC, AICPA / NASBA, SRA / NMC.
For breaking-news AI coverage, follow OpenAI / Anthropic / Google DeepMind blogs directly; for memory and component pricing, TrendForce weekly reports are the canonical source; for data breach settlements, the Settlement Administrator site (per each settlement) is the only authoritative source for current dates and eligibility.
Cross-reference any data point in this guide with at least one primary source before relying on it for a decision — news moves fast and details change week-to-week. Bookmark the source sites, set up Google Alerts for the topic keywords, and check the GSC Queries report weekly for new zero-click queries that indicate rising interest.
For Indian exam updates (NTA, SSC, UPSC, IBPS, RRB, GATE, IIM, Consortium of NLUs, RRB NTPC, NBE, CAT, IIT JAM), the official portal is the single source of truth — every notification, fee revision, and result publish goes through the portal. Recheck the portal before applying or appearing; date slips are common.
This guide was authored on September 8, 2026. Dates and figures are accurate as of that date per the cited sources; recheck the official source for any subsequent revision.
