Claude Opus 5.5 Pricing, Caching and Agent Costs

Quick answer

Anthropic released Claude Opus 5.5 on September 22, 2026. The model is aimed at long-running agentic coding and knowledge work, with a 1M-token context window, up to 128K output tokens, and adaptive thinking that is always on. Anthropic prices it at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million tokens.

Anthropic says typical workloads cost about 40% less to run than Opus 5, while output is more than 30% faster. Those are Anthropic's own workload measurements, not a guarantee for every application.

The practical TowCue takeaway is that model selection for persistent agents is becoming a systems-economics problem. Teams should measure cost per accepted, reviewed outcome, including repeated context, cache behavior, latency, repair loops and human review—not only the list price of a fresh prompt.

What changed with Claude Opus 5.5?

Opus 5.5 is the first model in Anthropic's Claude 5.5 family. Anthropic positions it for difficult coding, computer-use, agentic and professional knowledge-work tasks.

The API model ID is claude-opus-5-5. Anthropic lists availability through the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. Claude's product page also says Opus 5.5 is available to Pro, Max, Team and Enterprise users.

The model has a 1M-token context window and a standard maximum output of 128K tokens. Message Batches can reach 300K output tokens with Anthropic's documented beta header.

How much does Claude Opus 5.5 cost?

Anthropic lists standard API pricing at:

$4 / MTok input;
$20 / MTok output;
$5 / MTok for a 5-minute cache write;
$8 / MTok for a 1-hour cache write;
$0.20 / MTok for a cache read.

Batch API input and output receive a 50% discount according to Anthropic's model documentation.

Compared with Opus 5, Anthropic says input and output token prices are 20% lower and cache reads are 60% cheaper. Its launch post says these changes, together with efficiency improvements, make typical workloads about 40% cheaper to run.

That 40% figure should be treated as an Anthropic-reported workload estimate. Your actual saving depends on prompt shape, output length, cache reuse, tool calls and how many repair turns the task needs.

Why do cache reads matter so much for agents?

A long-running coding agent repeatedly carries stable information: repository instructions, tool schemas, architecture notes, policies and prior task context. Paying full fresh-input price for the same stable prefix on every request can become a meaningful part of the total bill.

Anthropic explicitly says cache reads make up a large share of the cost of long-running agentic work. Cutting the Opus cache-read price to $0.20 / MTok therefore matters more for persistent workflows than it does for a one-off question.

The useful design question is not simply, "How cheap is this model per token?" It is, "How much of the expensive context can this workflow safely reuse?"

Is Opus 5.5 now the default choice for every coding task?

No. Lower Opus pricing does not remove the need for routing.

A bounded refactor with deterministic tests may still be better served by a cheaper model. A difficult production bug, architecture migration or long autonomous task may justify Opus 5.5 if stronger reasoning reduces failed attempts and reviewer effort.

Anthropic itself offers multiple model tiers, and Opus 5.5's default effort is medium. Adaptive thinking is always enabled, while the effort parameter controls how deeply the model reasons. That creates another routing dimension: teams can vary both the model and the reasoning effort according to task consequence and uncertainty.

TowCue take: optimize the full agent loop

For production AI work, the useful cost equation is broader than token price:

task → model + effort → context → tools → execution → tests → review → repair → accepted outcome

Track at least:

fresh input and cached-input cost;
output cost;
time to a review-ready result;
failed tests and repair turns;
reviewer time;
rework after review;
defects that escape into production.

A model can be expensive per token and still be cheaper per accepted change if it needs fewer retries. A cheaper model can win when the task is narrow and verification is strong.

What does faster output change?

Anthropic says Opus 5.5 generates output more than 30% faster than Opus 5. For interactive work, lower latency is immediately noticeable. For agents, the effect can compound because one task may contain many sequential model calls separated by tool execution.

Anthropic also offers a separate Fast mode for Opus 5.5 on Claude Code and Claude Platform, advertised at up to 2.5x faster speed. Fast mode uses separate pricing of $8 / MTok input and $40 / MTok output, so teams should treat it as a latency-versus-cost decision rather than a free acceleration.

Who should evaluate Opus 5.5 first?

The clearest candidates are teams already using Opus-class models for long coding sessions, computer-use agents, research workflows or other tasks where repeated context and reviewer time dominate the economics.

Teams using lighter models successfully for short, deterministic tasks have less reason to migrate everything. Instead, create a small evaluation set of real work and compare Opus 5.5 with your current default under the same tools, instructions and tests.

Measure not just task completion, but total elapsed time, retries, review effort and cache hit behavior.

What should teams test before changing defaults?

Use representative tasks rather than public benchmarks alone. Include one mechanical change, one ambiguous bug, one multi-file feature, one research-heavy task and one long-running agent job.

For each run, record the model, effort setting, fresh input, cache reads, output, elapsed time, tool failures, repair turns and reviewer time. Then compare the cost of accepted outcomes.

If your application relies heavily on prompt caching, test cache architecture separately from model quality. A lower cache-read price helps only when stable context is structured so it can actually be reused.

TowCue take: model capability is becoming only one line item

Opus 5.5 reinforces a broader pattern across frontier AI products. Model quality still matters, but production economics increasingly depend on what surrounds the model:

routing policy → reusable context → model + effort → tools → verification → human review

That changes the buying question. Instead of asking which model wins the benchmark, teams should ask which configuration produces the best reviewed outcome for the workload they actually run.

The strongest model should earn its place where ambiguity and failure costs justify it. The cheapest model should earn its place where verification is easy. And repeated context should be treated as infrastructure that can be designed, measured and optimized.

Claude Opus 5.5 cuts agent costs: why cache economics now shape model choice