GPT-6.2 Astra and Sol ship 20%-600% cheaper but draw mixed reviews; Anthropic still holds the top model
Latent Space · rss · 2026-11-18
Latent Space's AI News breaks down OpenAI's GPT-6.2 launch:
Models and pricing
- Astra scores 62 on AA's Intelligence Index (vs Fable 5.2's 65); $8/$0.40 cached/$40 per M tokens, 1M context; Enterprise config adds legal search (20.8% on Harvey LAB-AA vs 15.0% base).
- Sol matches September's Astra at 15% of the cost per task; $2/$0.10/$10.
- First externally released OpenAI model rated High for AI Self-Improvement; Epoch verified 7 of 68 FrontierMath Erdős solutions in Lean (previous best: 3).
Polarized early reviews
- Matt Shumer called Astra "essentially unusable"; Pietro Schirano blamed the harness; Mercor's OpenWorldEngineer puts Astra at 67.2% vs Fable 5.2's 54.6%; Astra leads 3D/robotics; on WANDR Astra hits 0.783 at $12.54/task vs Sol 0.746 at $2.65.
- Latent Space's verdict: some backlash is influencer overreaction, RL artifacts are real—Anthropic owns the best model, OpenAI the best intelligence-per-dollar.
Also covered
- Microsoft ships MAI Code 2/Flash in Copilot; Copilot Fusion hits 71.2% on Terminal-Bench 4.1 at 88% lower cost than Fable 5.2; MAI Image 2.7 tops image editing.
- Google: Hydrogen powers free Gemini Tasks in six countries; Antigravity robotics plugin with MuJoCo.
- Agent engineering: HF's Pi adapter lifts Qwen3.8-27B from 15 to 32 solves on the 51 hardest SWE-bench Pro tasks with 55% fewer tool calls; Baseten measures up to 4.4× cost spread across 22 harnesses; Google SIM shows harnesses overfit to models; in a 1000-agent run only 40-60 called the model at once (130:1 read/write tokens); SemiAnalysis: 1.8% of uncached requests drove 15% of spend.
- Industry: NVIDIA posts $115.8B quarterly revenue (+103%) under interim CEO Colette Kress, board won't reinstate Jensen Huang; Isomorphic Labs raises $5B; Genesis AI raises $500M at $3.5B; Zyphra reports 9-60× long-context speedups; vLLM's sliding-window draft boosts Kimi K3.1 Pro acceptance length up to 30%.
More from coding & agent
- Pawel Huryn open-sources 45 agent skills distilled from shipping three live products — PawelHuryn · 2026-11-18
- Google donates Agent Substrate to CNCF: 500ms restores, 10x density — krishnan · 2026-11-18
- Code model eval: Rust hits 95-100% accuracy as 13 languages see big gains — shuding · 2026-11-18
- Law school AI course wraps: 6 student teams ship working legal AI prototypes — jkubicki · 2026-11-18
- n8n's open-source agent Nod: secure remote access via private reverse proxy — victorialslocum · 2026-11-18
- Floot Design launches to kill AI slop: pick your UI before any code ships — heyshrutimishra · 2026-11-18