Claude Opus 5.5 cuts token prices 20%. Anthropic's 40% claim needs one more number

The price cut is direct; the total saving is workload-dependent

Anthropic released Claude Opus 5.5 on 22 September 2026 as the first model in its Claude 5.5 family. The published API rates are $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. That is a 20% reduction in the two headline token prices.

Anthropic separately says Opus 5.5 costs 40% less than Opus 5 on a typical workload at default settings. The extra saving comes from its claim that the model needs fewer tokens to finish a task. Those are different measurements: one is a public unit price, while the other combines price, token consumption and the vendor's chosen task mix.

Caching is where agentic workloads receive the largest unit-price change

Anthropic's prices per one million tokens

Usage	Opus 5.5	Opus 5	Unit-price change
Input	$4	$5	20% lower
Output	$20	$25	20% lower
Cache reads	$0.20	$0.50	60% lower
Cache writes	$5	$6.25	20% lower
Fast mode input	$8	Not the standard rate	2x Opus 5.5 standard input
Fast mode output	$40	Not the standard rate	2x Opus 5.5 standard output

Cache reads fall by 60%, a material change for coding and agent workflows that repeatedly reuse a large codebase or instruction prefix. Cache writes fall by 20%. The economic benefit therefore depends on whether an application actually caches stable context and achieves a high reuse rate.

Fast mode is a separate trade. Anthropic says it can deliver up to 2.5 times the speed, but it doubles the standard Opus 5.5 input and output rates to $8 and $40. It may be rational for an interactive incident or a developer waiting on a critical path. It is harder to justify for unattended work where latency has little business value.

The benchmark table is broad, but it is still a vendor release table

Anthropic reports gains across agentic coding, knowledge work, computer use and chart recognition. Examples include 66.4% on Terminal-Bench 4.0, 54.4% on the main FrontierCode 1.1 result and 81.8% on a partial OSWorld 2.0 result. The company also says output is more than 30% faster than Opus 5.

The footnotes are essential. Most Opus 5.5 headline benchmark results use adaptive thinking at maximum effort, while some comparisons use scores reported by other model vendors. AutomationBench was run by Zapier, and several tasks could fall back to other Claude models when safeguards intervened. Anthropic itself warns that benchmark margins at this capability level are becoming less reliable guides to real-world differences.

That does not make the results unhelpful. It means procurement should reproduce a small number of representative jobs with identical prompts, tools, time limits and acceptance tests before translating leaderboard gaps into a production forecast.

Measure cost per accepted task, not cost per million tokens alone

A practical evaluation record for an agentic workload

Measure	Why it matters
Input, output, cache-read and cache-write tokens	Reconstructs the bill instead of relying on one blended token count
Retries and failed tool calls	Captures work that consumed money but produced no accepted result
Wall-clock time	Shows whether standard or fast mode changes a business bottleneck
Human review minutes	A cheaper run can cost more if it creates extra checking or rework
Acceptance rate	Converts raw model output into completed useful tasks

For a code migration, an accepted task might mean that the patch passes the repository's tests and review. For research, it might require citations that match the claimed evidence. For a browser agent, it could require reaching the correct final state without an irreversible mistake.

Run the same sample with Opus 5 and Opus 5.5, keep the harness and permissions fixed, and record the full task cost. That test reveals whether the vendor's token-efficiency improvement appears in the environment that will pay the bill.

Stronger safety claims still require application controls

Anthropic describes Opus 5.5 as its strongest model on an automated behavioural audit and says it is more resistant to prompt injection than Opus 5. It also says the model is less likely to take hard-to-reverse actions or operate outside its assigned boundaries. These are vendor evaluation claims, not a replacement for permission boundaries in the product using the model.

A stronger model can execute a mistaken plan more effectively. High-impact tools should still use narrow credentials, explicit approval points, reversible operations and logs that show what changed. Model capability and system authority remain separate decisions.

The upgrade case is strongest where context is reused and completion is expensive