Claude Opus 5.5 Lands at 40% Lower Cost, the Same Week OpenAI Cut GPT-6 Pricing in Half

Anthropic launched Claude Opus 5.5 on September 22, cutting list pricing 20% to $4/$20 per million input/output tokens and cache reads 60% to $0.20/million, while the model runs roughly 40% cheaper than Opus 5 to complete comparable work. On Terminal-Bench 4.0 it scored 66.4% against Opus 5’s 52.3%. Anthropic cited an early tester completing a 680,000-line code migration in under 24 hours — normally weeks of work — and auditing a 200,000-line codebase in 3 hours instead of 20-plus.

Anthropic engineer Addy Osmani’s own cost breakdown of the release found a roughly 31% cost reduction on matched sessions, driven by turn count, cache-read volume, and which model handles a given subtask. His recommendation for teams running agentic workflows: default to Opus 5.5 for routine work, escalate to the heavier Fable 5.1 model only when a task stalls, and offload simple lookups to smaller Sonnet or Haiku models rather than running everything on the flagship.

The same week, OpenAI cut GPT-6 pricing roughly in half with two new models, Sol and Luna, priced at $2/$10 and $0.10/$0.50 per million tokens respectively. OpenAI’s own benchmarks showed Sol scoring within 1.1 points of Claude Fable 5 on the DeepSWE v1.1 coding benchmark at about 80% lower cost — but that comparison was built against Claude Opus 5, not the 5.5 pricing and performance Anthropic shipped the same week.

For anyone building or selling on top of these platforms, the takeaway isn’t which lab “won” — it’s that the cost floor for frontier-level agentic work moved twice in the same seven days, from both major labs simultaneously. A vendor comparison published this week is stale by the time a buyer reads it; the number worth tracking isn’t a benchmark score but how fast the price-per-task keeps resetting under it.