GPT-6 SOL AND LUNA ARE OUT!!!

에피소드 타임라인

1. 1:27

   Introduction and overview of Claude Opus 5.5 release

   • Anthropic released Claude Opus 5.5, the first model since Dario's essay calling for pacing the frontier, and it performs at Fable 5.1 level for most tasks at 40% lower cost than Opus 5.
   • Opus 5.5 is faster, cheaper, and on many benchmarks actually takes the frontier position from Fable 5.1, which the host calls crazy to see.
   • The host notes this is the first major model release since Anthropic began pacing the frontier, and external evaluators Meter and Frontier Design tested it before release.
2. 10:50

   Benchmark deep dive: Terminal Bench, Frontier Code, Cursor Bench, GDPval

   • Opus 5.5 dominates Terminal Bench 4.0 with 66.4% versus Astra's 57.9% and Fable 5.1's 55.8%, a 10-point jump over Fable.
   • On GDPval 2.1, OpenAI's benchmark for real-world knowledge work, Opus 5.5 scores 1846 ELO versus Fable 5.1's 1735, a 300+ point jump over GPT6 Astra's 1542.
   • On Frontier Code V1.1 Opus 5.5 hits 54.4% versus Fable's 50% and Astra's 53.3%, and on Cursor Bench it scores 57.8% versus Fable 5.1's 51%.
   • Opus 5.5 did not take first place on Automation Bench (Astra 41.4% vs 40%) or Terminal Bench Science, where GPT6 Astra still leads.
3. 21:32

   Pricing, speed, and cost-per-task efficiency analysis

   • Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, a 20% price decrease from Opus 5, with cache reads at 20 cents versus 50 cents.
   • The host argues cost per task completed matters more than cost per token, since a cheap model that uses 10x tokens is still expensive overall.
   • On Automation Bench and Frontier Code, Opus 5.5 sits in the top-left quadrant of quality versus cost, with medium effort beating max effort on cost efficiency.
   • Opus 5.5 generates output more than 30% faster than Opus 5 and requires less compute to serve.