Dek: Anthropic lowered the model’s published input and output rates by 20% and its cache-read rate by 60%. The company’s larger 40% typical-workload saving depends on how a task actually runs.
Anthropic introduced Claude Opus 5.5 on September 22 as the first model in its Claude 5.5 family. The launch came with a headline likely to catch an API buyer’s attention: Anthropic says the model costs 40% less than Opus 5 on typical token-billed workloads at default settings.
That figure is a task-cost estimate, not a universal price reduction. The published rate card shows a 20% cut in base input and output prices, a 20% cut in cache-write pricing, and a 60% cut for cache reads. Anthropic says the rest of the typical-workload improvement comes from Opus 5.5 using fewer tokens to complete tasks.
The practical question is therefore not whether every request becomes 40% cheaper. It is whether a buyer’s own mix of fresh input, cached context, output, retries, effort settings, and routing produces a similar reduction per accepted result.
The rate card starts with a 20% base cut
The standard API prices for Opus 5.5 are $4 per million input tokens and $20 per million output tokens, according to Anthropic’s launch announcement. Opus 5 is listed at $5 and $25 respectively. Both base rates are 20% lower.
Caching creates the largest percentage difference. An Opus 5.5 cache read costs $0.20 per million tokens, down from $0.50 for Opus 5, while a cache write falls from $6.25 to $5. That makes cache reads 60% cheaper and cache writes 20% cheaper.
| Price per 1M tokens | Opus 5.5 | Opus 5 | Published reduction |
|---|---|---|---|
| Input | $4 | $5 | 20% |
| Output | $20 | $25 | 20% |
| Cache read | $0.20 | $0.50 | 60% |
| Cache write | $5 | $6.25 | 20% |
The table explains why 40% cannot describe every bill. A workload with little cache reuse and similar token consumption would begin near the 20% base-rate reduction. One with a large share of repeated context could benefit more from the cheaper cache-read line. Neither outcome is guaranteed because task behavior can change alongside prices.
Why cost per task can move further
Anthropic attributes its 40% typical-workload estimate to lower rates and fewer tokens used to finish a task. That second component is consequential. When a task uses fewer tokens, its total cost can fall faster than the rate per token.
The reverse can also happen for a particular deployment. Long outputs remain more expensive than inputs, so output-heavy work can dominate the bill. A workflow that retries frequently, carries large uncached prompts, or fails an acceptance check will have different economics from Anthropic’s typical case. Buyers should measure the entire path to an accepted result rather than a single successful request.
This distinction also prevents a common category error. Anthropic’s claim concerns token-billed workloads. It does not announce a 40% reduction in fixed monthly Claude subscription prices. Readers comparing plans and quotas can use our guide to AI subscription usage limits and reset transparency for that separate purchasing question.
Fast mode changes the calculation
Opus 5.5 also has a fast mode in Claude Code and on the Claude Platform. Anthropic says it can deliver up to 2.5 times the speed, priced at $8 per million input tokens and $40 per million output tokens. Those rates are twice the model’s standard input and output prices.
That option turns latency into an explicit procurement choice. A team may decide faster completion is worth the premium for interactive coding or time-sensitive operations. It should not apply the 40% default-setting estimate to fast-mode traffic without measuring it. The relevant comparison is the value of reduced waiting against the higher token rates and the effect on completed-task volume.
Benchmarks add context, not a clean price-performance verdict
Anthropic reports a 66.4% score for Opus 5.5 on Terminal-Bench 4.0, compared with 52.3% for Opus 5 and 57.9% for GPT-6 Astra. The numbers may suggest a stronger coding agent, but they are vendor-reported and the compared settings are not identical: Anthropic lists Opus 5.5 at xhigh and GPT-6 Astra at high.
The company also reports a standard error of plus or minus 2.6 percentage points for Opus 5.5 and roughly 1.6 to 2 points for the other Claude results. Anthropic says its Opus 5 result reproduces the public leaderboard within noise. It further cautions that benchmark margins at this capability level are a less reliable guide to real-world differences, and that its observed gap between Opus 5.5 and Fable 5.1 is narrower than the scores imply.
Those caveats rule out treating the table as a controlled head-to-head or proof of a universal price-performance advantage. The benchmark and 40% task-cost claim come from Anthropic, not an independent public replication.
Availability and safeguards can affect deployment
Anthropic’s Claude Opus page lists Opus 5.5 for Claude Pro, Max, Team, and Enterprise, with API access through the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Foundry. The API identifier is claude-opus-5-5.
Lower prices do not remove deployment controls. Anthropic says Opus 5.5 has safeguards similar to Fable 5.1 for cybersecurity, biology, and anti-distillation. Most cybersecurity tasks are routed to Opus 4.8. Vetted organizations can apply to the Life Sciences Verification Program, while Anthropic says an expansion of its Cyber Verification Program is planned.
Routing matters because the model named in a price comparison may not handle every task. Buyers in affected domains should test the applicable path rather than assume the public Opus 5.5 rate and performance profile maps directly to all requests.
The buyer test is cost per accepted result
A useful evaluation should replay representative jobs with the buyer’s real cache policy, prompt sizes, output requirements, effort setting, retries, and acceptance criteria. Teams should record total billed input, output, cache reads, and cache writes for each accepted result, then compare those totals with the same workflow on Opus 5.
They should also separate standard and fast-mode runs, document any safeguard routing, and examine failed or rejected outputs. Averages alone can hide expensive outliers, especially in agentic jobs that loop through tools or retry steps.
The published prices make Opus 5.5 cheaper at every listed standard rate, with cache reads receiving the deepest cut. Anthropic’s 40% figure is plausible only as the company describes it: an estimate for typical token-billed workloads at default settings that combines the rate reductions with token efficiency. The number is a useful hypothesis for a procurement test, not the answer to one.