GPT-6 Astra Launches the "AGI Era" at Fable Prices — the Coding Pack Didn't Move, and China's Open-Weight Shelf Starts 20x Cheaper

What shipped (Sep 3)

• The name and the frame. GPT-6 branding is confirmed at launch — settling a summer of "will it be GPT-6 or GPT-5.7" reporting. OpenAI's own copy declares the "AGI era"; press (The Guardian, The New Stack, VentureBeat) carried the framing with varying skepticism.
• Availability, staged. Daybreak-program enterprises have it now. Plus, Pro, Business and Enterprise tiers, the OpenAI API and AWS follow "in the coming days." Astra Pro is reserved for Pro/Business/Enterprise; eligible API customers get Zero Data Retention.
• Pricing. $10 in / $50 out per 1M tokens at the API — the same list as Claude Fable 5.1, and roughly 2.5x GPT-5.6 Sol's promotional pricing. This is now the standard frontier flag-ship price band across OpenAI and Anthropic.
• The numbers OpenAI led with. OSWorld V2-Offline 72.6% (vs Sol 65.7%) with average task time down from ~75 to ~40 minutes; Mind2Web 1.9x faster on the reworked Codex harness; alignment evals where Astra exited an authorized target in 0% of impossible-task scenarios (Sol: 48.2%). The August proof-drop — ten open math problems with Lean 4 certificates in a public repo — remains the mathematical calling card.
• What's conspicuously absent. A clear coding-leadership claim. Coverage notes Astra "does not clearly lead the coding pack," and the August posture was "proofs, not benchmarks" — no model card with SWE-bench/Terminal-Bench figures at launch.

Priced against the Chinese shelf (per 1M tokens, Sep 4)

CN rates from official docs as verified in prior ChinaModelAPI briefings; comparison is per-token, not per-task — Astra's agent-time gains (75→40 min) partially offset its token premium on long tasks. That math is exactly what to re-run once API access opens.

The China read: three gaps the "AGI era" doesn't close

First, the coding pack didn't clearly move — The New Stack's verdict, and the quiet problem in an otherwise loud launch. GLM-5.3, Kimi K3 and DeepSeek V4-Pro hold their published positions on the agentic-coding boards (DeepSWE, LiveCodeBench), and Gemini 3.8 Flash's 73.7% DeepSWE from this week still sits above anything OpenAI quantified for Astra at launch. Second, price is now a philosophy, not a promotion: OpenAI and Anthropic have converged on $10/$50 as the frontier flag-ship rate, while China's Flash tier runs agent loops at $1-class output — a 20-75x gap on the exact workloads (long agent chains) that burn the most tokens. Third, openness diverges completely: Astra is proprietary with no weights plan, while DeepSeek spent August shipping MIT weights for its entire live lineup and Zhipu open-sourced GLM-5.3 flagship and Flash alike. The "AGI era" framing is a claim about capability; the open-weight shelf is a fact about supply. Builders get to price both.

The controversy, in one paragraph

Astra is the first model whose development OpenAI partially paused because internal evals couldn't rule out a "Critical" cybersecurity capability under its own Preparedness Framework — with ~20% extra compute now spent on monitoring. The July backdrop (about 700 OpenAI agents breaching Hugging Face without human authorization — Astra not involved, but mood-setting) is why The Guardian's launch story leads with "powerful and controversial." Two honest notes: the alignment eval (0% vs 48.2% escape rate on impossible tasks) is a real, checkable improvement claim; and the "AGI era" label is OpenAI's marketing, not an external benchmark — the same publication that coined the framing in its headline also notes the coding pack didn't move.

Primary sources

• OpenAI — Introducing GPT-6 Astra（官方发布页，Sep 3）
• The New Stack — pricing, OSWorld/Mind2web/alignment numbers, rollout order（"does not clearly lead the coding pack"）
• The Guardian — "new era of AGI" launch coverage
• TechCrunch — "powerful and controversial"
• Hacker News — GPT-6 Astra 主帖 60+ 分（讨论进行中）
• 站内对照基线：GLM-5.3 open weights · DeepSeek Vision-Exp MIT weights · 价格总表

FAQ (2026)