GPT-5.6 Boosts Capabilities & Cuts Costs: ARC-AGI-3 Score Jumps to 38.3%
新智元 · wechat · 2026-08-14
OpenAI released the Builder's Guide to GPT-5.6, showcasing significant advancements in model capability and cost optimization.
- ARC-AGI-3 Score Surge: By switching to the Responses API with "reasoning persistence" and "compression" enabled, GPT-5.6 Sol's score on ARC-AGI-3 jumped from 13.3% to 38.3%, while output tokens were reduced by 6x.
- Luna (Extra-Small Model): On the BrowseComp benchmark, GPT-5.6 Luna achieved 84.04% for just $1.33, closely matching GPT-5.5's 84.36% which cost $33.27—a 25x cost reduction.
- Engineering Best Practices: OpenAI recommends three cost-saving methods: Programmatic Tool Calling (executing JS in a sandbox to reduce interaction rounds), native multi-agent parallel processing, and prompt caching extended to 30 minutes.
- Ultrafast Mode: Powered by Cerebras, GPT-5.6 Sol achieves up to 750 tokens/s. On Humanity's Last Exam, it ran nearly 7x faster than Claude Fable 5 while maintaining the same flagship intelligence level.
Related event: OpenAI and Cerebras Preview Ultrafast Mode for GPT-5.6 Sol at 750 Tokens/s(20 posts)→
More from coding & agent
- A vector DB is not LLM memory: long thread maps write/read paths, failure modes and evals — techNmak · 2026-10-16
- OpenUI 1.0 launches open source, turning agent text answers into live interactive UIs — heyshrutimishra · 2026-10-16
- 10 GitHub repos to see what your AI agent really did, from Langfuse to claude-code-log — victor_explore · 2026-10-16
- Mosaic: multi-session agentic RL with hierarchical advantage estimation — LARK-Lab · 2026-10-16
- Recursive Environment Scaling tanks six models' success rates by 57–77 points — Lingxiao Jing · 2026-10-16
- Harvard's MPT: decentralized agent swarms learn self-orchestration via bidding — Harvard · 2026-10-16