Anthropic went CRAZY (Opus 5.5)

Cronologia dell'episodio

1. 0:00

   Anthropic releases Claude Opus 5.5, the first frontier model since calling for pacing the frontier, delivering top benchmark scores at lower cost and higher speed.

   • Opus 5.5 is the first major Anthropic release since Dario's essay calling for pacing the frontier, yet it still lands as the absolute frontier model, beating Fable 5.1 on most benchmarks.
   • It performs at Fable 5.1 level for most tasks while costing 40% less to run than Opus 5 and generating output over 30% faster.
   • The last .5 release, Opus 4.5, was the inflection point where coding models became capable of handling large multi-hour autonomous tasks, so expectations for 5.5 are high.
2. 3:30

   Benchmark results show Opus 5.5 dominating coding, knowledge work, and reasoning evaluations.

   • On Terminal Bench 4.0, Opus 5.5 scores 66.4% versus 57.9% for Astra and 55.8% for Fable 5.1, a 10-point jump over the previous frontier.
   • On GDPval 2.1, OpenAI's real-world knowledge work benchmark, it hits an 1846 ELO score versus 1735 for Fable 5.1, a 300+ point jump.
   • It also leads on Frontier Code V1.1 (54.4), Cursorbench (57.8), and Humanity's Last Exam (67%), though it places second to Astra on Automation Bench and Terminal Bench Science.
3. 6:14

   Pricing drops modestly per token, but cost-per-task efficiency is the real story.

   • Opus 5.5 costs $4 per million input tokens versus $5 for Opus 5, and $20 versus $25 per million output tokens, with cache reads at 20 cents versus 50 cents.
   • The 40% total cost reduction on typical workloads comes from combining lower token prices with fewer tokens needed to complete a task.
   • On quality-versus-cost-per-task charts for Automation Bench, Frontier Code, GDPval, and Terminal Bench, Opus 5.5 consistently occupies the ideal upper-left quadrant.

Concetti chiave

• Opus 5.5— Anthropic's new frontier model that dominates most benchmarks while being cheaper and faster than Opus 5.
• pacing the frontier— Anthropic's stated policy of slowing the release of its most capable models while making frontier intelligence broadly accessible.
• Terminal Bench 4.0— The key agentic coding benchmark where Opus 5.5 scored 66.4%, a 10-point jump over Fable 5.1.

Citazioni rilevanti

• It performs at the level of Fable 5.1 for most tasks, but it costs 40% less to run than Opus 5.

  🤯

  — A frontier model that is simultaneously cheaper and more capable overturns the assumption that top intelligence must cost more.

  Vai a 1:07

• That is over a 300 point ELO jump.

  🤯

  — A 300-point ELO gain on GDPval is an enormous leap that reveals how fast frontier capability is compounding.

  Vai a 4:06

Conclusioni applicabili