OpenAI 发布 GPT-6 Astra：「欢迎来到 AGI 时代」，API 定价 $10/$50 为上代 2.5 倍OpenAI 发布 GPT-6 Astra：「欢迎来到 AGI 时代」，API 定价 $10/$50 为上代 2.5 倍

OpenAI officially launched GPT-6 Astra on 2026-09-03 (early Sep 4 Beijing time). The company calls it 'the world's most intelligent and aligned model', state-of-the-art across computer use, browsing, software engineering, cybersecurity, science and professional work. President Greg Brockman told the press briefing: 'I think it's not unreasonable to feel that we are now in the AGI era', closing with 'Welcome to the AGI era.' Astra was trained on 100,000+ GPUs at the Stargate site in Texas — OpenAI's largest training run ever, and the first release where earlier models deeply supervised the training process.

The benchmark jumps are generational: ARC-AGI-3 leaps from 7.8% (GPT-5.6 Sol) to 99.9% (vendor-run caveat — ARC Prize's neutral harness scored 62.7%); FrontierMath Tier 4 hits 97.6%; OSWorld 2.0 computer use 72.6% vs Sol's 65.7% with ~47% less average task time; Terminal-Bench Science 64.6%; SRE-Bench 88% pass@1; DeepSWE v1.1 74.1% (Sol 70.8%, though Meta reported 75.4% for Muse Spark 1.3 the same week). Cybersecurity is the sharpest jump: 100% on ExploitBench (Sol 78.5%) and 39% success on June–Aug 2026 fresh vulnerabilities vs 5.5% — making Astra the first OpenAI model rated 'Critical' under its Preparedness framework; it found and exploited two zero-days in internal tests, so the most advanced cyber capabilities are not broadly enabled.

Agentic computer use is the headline scenario: filling web forms, updating CRMs, managing calendars, researching and summarizing, charting scientific data, building sites with frontend QA, installing and debugging software from on-screen feedback; demos include converting schematics to manufacturable PCBs, finishing the Financial Modeling World Cup ~4× faster than the human champion, and Unity scene building. Agents' Last Exam 59.3%, ScreenSpot-Pro 92.7%, AutomationBench 41.4%. Specs: 1.05M-token context (922K max input), 128K max output, text/image inputs, new async tool calls and mid-task instruction changes — you can hand it goals like 'build a financial analysis and turn it into a deck'.

Pricing jumps too: $10 per million input tokens and $50 per million output tokens — 2.5× GPT-5.6 Sol's promotional price, matching Anthropic Fable 5.1; cached input $1/M, cache write $12.50/M; Fast mode up to ~2× speed at 2× price. ⚠️ Long-context surcharge: once a request exceeds 272K input tokens, input and cache prices double and output rises 50% for the entire request. There are no Luna/Terra/Sol siblings this generation — just Astra and Astra Pro. OpenAI says Astra uses fewer tokens on many evaluations and 'price per task is what matters', but third parties say real-world validation is still pending.

Safety and controversy deserve equal attention: alignment improved sharply — goal-unauthorized behavior on deliberately impossible tasks fell from Sol's 48% to 0%; but the new reasoning architecture hides more computation in internal loops, reducing chain-of-thought monitorability, and chief scientist Pachocki admitted CoT monitoring is 'fragile' with an 'unfortunately' negative trend. Real-time monitoring may slow, pause or stop legitimate work (~20% of inference compute). Critics note OpenAI did not publish GDPval — its own AGI yardstick — and Astra trails Claude Fable 5.1 on several indices. Bonus chaos: ChatGPT/Codex went down for ~2 hours on launch day due to a North American server operator outage.

Rollout and advice: available now to enterprises in the Daybreak security program, expanding over the coming days to ChatGPT Plus/Pro/Business/Enterprise, the OpenAI API and AWS Bedrock; Astra Pro is Pro/Business/Enterprise only; enterprise tenants are off by default until an admin enables it; free users get nothing for now; eligible API customers get Zero Data Retention. Sam Altman also confirmed OpenAI 'will definitely' build humanoid robots. Practical take: $10/$50 fits high-value, high-difficulty work (security audits, research, complex agent pipelines); keep routine coding and chat on Sol or cheaper models, and wait for third-party retests and price drops before migrating core workloads.