Opus 5.5 claims 40% cheaper. Rerun your benchmarks?

Anthropic cut Opus pricing on Tuesday, and it did not cut it evenly.

Claude Opus 5.5, launched September 22, bills at $4 per million input tokens and $20 per million output, down 20% from Opus 5's $5 and $25. The cut that matters more for agent operators is elsewhere: cache reads fell 60%, to $0.20 per million, and cache writes dropped to $5 from $6.25. Anthropic claims typical workloads will cost 40% less to run than on Opus 5. Subscription users also got raised five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans, plus a rate-limit reset they can save for later.

So the reader question: should coding-agent operators rerun their model-cost benchmarks?

If Opus 5 was anywhere in your baseline, yes. The numbers changed twice, not once.

The first change is the tariff. A straight 20% cut on input and output tokens means every old per-task number is simply stale. For short, well-scoped work, you do not need a new benchmark. Multiply by 0.8 and move on.

The second change is where it gets interesting. Anthropic says Opus 5.5 finishes tasks with fewer tokens and fewer tool calls, and that cache reads now make up the majority of agentic and coding work costs. Its own usage data, published September 24, explains why the cache-read cut does the heavy lifting: over six months, the input-to-output token ratio in Claude Code sessions moved from 189:1 to 324:1, with 40% more model calls per prompt. When your session is 324 tokens of input context for every token of output, a 60% cut on cache reads beats a 20% cut on the tariff by a mile.

"On a well-scoped task, both models finish in about the same number of turns, and the price cut is all you get. The gap should be biggest on open-ended tasks, where a model can spend many turns on the wrong idea."

That is engineering leader Addy Osmani's verdict, quoted in Anthropic's companion analysis published September 22 on the Claude blog, and it is the right heuristic for deciding what to rerun. Well-scoped tasks get the 20%. Sprawling multi-turn loops might get the full 40%. Or they might not.

That last caveat is doing real work, because the 40% is a vendor claim, not a finding. Anthropic's own announcement pairs a pricing table measured at default effort with benchmark numbers recorded mostly at max effort, then warns that "benchmark margins have become a less reliable guide to real-world differences." Treat the 40% as a ceiling until your own logs confirm it. Independent per-task evidence is thin but not zero: Artificial Analysis measured Opus 5.5 at roughly 119k output tokens per task against ~73k for its predecessor on the Intelligence Index, landing the two roughly level on cost per task at max effort. That undercuts the 40% claim more than it supports it. Our price-war comparison of what Claude Code, Codex, and Gemini CLI actually cost has the per-token numbers; beyond that one independent data point, the per-task evidence is still Anthropic's alone.

The practical checklist:

Rerun if your last benchmark used Opus 5 as a cost baseline, or if your agent loops run long sessions with prompt caching enabled. Also worth a rerun if you previously ruled out Opus-tier models on price; they deserve a recheck against Fable 5.1 or GPT-6 Sol, which launched the same afternoon at $2/$10. Measure cost per completed task, not cost per million tokens. That is the metric Anthropic is implicitly competing on, and it is the one your old benchmark probably under-weighted.

Skip the rerun if your tasks are short and well-scoped, or if Opus was never in your shortlist. Two pieces of fine print apply either way. First, the cache-read discount applies to API pricing only; on plan usage the limits simply stretch about 25% further. Second, Claude Code's fast mode now runs up to 2.5x speed at $8/$40, which is a separate rate tier, not a discount.

One migration note before you flip the switch: on 5.5, thinking can no longer be disabled, and preserved thinking, the anti-distillation safeguard, applies to API accounts created after August 31. If your harness toggles either, update it. Teams standardizing their coding-agent setup, including the now-universal AGENTS.md instruction file, should treat the model swap and the harness changes as one job.

The wider picture is a price war, and this is one battle in it. OpenAI launched GPT-6 Sol at $2/$10 the same afternoon, half of Opus 5.5's tariff, and the September tools wave keeps lowering the cost of the infrastructure around long sessions. The tariff cuts are real and verifiable. The per-task savings are not, yet. Rerun the benchmarks that matter. Multiply the ones that do not.