GPT-6 Astra tops ARC-AGI-3 while slashing token costs, DeepLearning.AI reports
DeepLearningAI · x · 2026-09-14
- DeepLearning.AI's The Batch covers OpenAI's new flagship vision-language model GPT-6 Astra, which the company says is the first to hit the "critical" cybersecurity tier of its Preparedness Framework; its most advanced cyber abilities are restricted to select organizations.
- Key specs and features:
- Up to 1,050,000 token input, 128,000 token output at 71.3 tokens/s; knowledge cutoff April 30, 2026
- Five reasoning levels (low to max), switchable mid-conversation without invalidating cache
- Asynchronous tool calls let the model keep reasoning while tools run
- Retained reasoning memory across API calls, context compaction, mid-turn steering, and experimental Codex note-taking/search over earlier context
- Benchmarks: first on ARC-AGI-3 and Arena AI's WebDev leaderboard, tied for first on Intelligence Index v4.3 (53), all at a fraction of the token usage and cost of the few models that beat it on some indexes.
Related event: GPT-6 Astra Ships with 1M Context, Tops ARC-AGI-3(2 posts)→
More from Models
- Grok trains on user data by default; business plans can opt out — carlosdponx · 2026-09-15
- Bolt Forge launches free until Oct 14 with GLM, DeepSeek and Kimi plus up to 50x more usage — HeyAmit_ · 2026-09-15
- GPT-6 Astra tested on robot control: impressive on simple tasks, limited dexterity — DJiafei · 2026-09-15
- SOTA Inference Is Nearly Free for Consumers, So the Local-Model Trend May Reverse — mobileraj · 2026-09-15
- Cursor user switches to Claude Code, burns through quota by day 3 — jdluk87 · 2026-09-15
- Claude Max and Codex tiers are creating a computing power gap that locks out $20-budget newcomers — IndraVahan · 2026-09-15