The Claude 5.5 choice now includes two released models: Opus 5.5, announced September 22, and Sonnet 5.5, announced September 28. Anthropic positions Opus for complex work requiring sustained judgment and Sonnet for well-scoped daily work. Haiku 5.5 remains a forthcoming release in the latest announcement. Opus release, Sonnet release.
For coding, the useful comparison is the cost of a result you can accept. Sonnet's lower token rates create a good reason to test it. They do not prove it is the cheaper route for every task after repairs and review.
Verified September 29, 2026. This article provides official rates and an illustrative calculation. Agent.Space has not measured either model's speed, pass rate, or cost on the example workload.
Opus 5.5 vs Sonnet 5.5 API prices
Standard first-party API prices, in US dollars per million tokens, from Claude Platform pricing:
One easy-to-miss detail is the equal cache-read price. Opus is twice Sonnet's rate in the input, output, and cache-write columns, but that does not make a mixed bill exactly twice as large. Do not combine a cache-read rate with a cache-write estimate.
Other processing modes, cloud-provider terms, paid tools, and applicable residency modifiers need their own checks. Opus Fast mode has a separate rate; a Standard-versus-Fast comparison changes more than the model. Subscription allowances also do not convert into a fixed token promise. See Claude Code billing paths when comparing plans rather than API requests.
A request-budget example you can recalculate
Suppose a request contains 100,000 uncached input tokens, 500,000 cache-read tokens, and 10,000 billable output tokens. Assume no cache creation, paid tools, special processing, or other charges. This is a pricing example, not an observed run.
Opus is 75% more expensive in this example, rather than 100%, because the cache reads cost the same. If a later request misses the cache or has more output, the ratio changes. If Opus solves the task in fewer calls, that changes the whole-task result again.
To use this method on a real project, replace each assumed quantity with the provider's reported usage. Count all calls, failed attempts, and repairs. Keep human review effort alongside the token bill. The coding-agent API cost guide explains the full accounting method.
Which tasks should you try first?
The following is our evaluation plan, not a benchmark ranking:
A model should earn a default through these results. If Sonnet passes the task reliably with less total effort, use that evidence. If Opus avoids repeated repairs or finds a problem the cheaper route misses, include that evidence. Do not grant more permissions simply because one model is described as stronger.
Do not turn a vendor percentage into your savings forecast
API compatibility also belongs in the comparison. The Opus 5.5 specification documents always-on adaptive thinking and breaking changes to forced tool use and thinking-block handling. A lower price is not a saving if your integration cannot complete its tool loop. Audit those changes before switching an existing application.
Anthropic reports typical task-cost improvements against each model's predecessor. Those are vendor test results, not a Sonnet-versus-Opus guarantee for your workload. A comparison needs the same starting state, compatible tools, recorded effort, and a fixed acceptance rule.
Use independent repository copies, score the outputs, and repeat on several tasks before adopting a default. The regression-evaluation guide helps separate one bad run from a repeatable difference. For CLI selection and provider alias checks, follow Sonnet 5.5 in Claude Code.
Compare available routes in Agent.Space
Agent.Space can provide a shared project boundary for supported harnesses and models. Check live model pricing and the compatible model selector before assuming either Claude 5.5 model is available. These Anthropic API rates are not a quote for Agent.Space usage.
When a supported route fits, start an Agent.Space Workspace, save the task brief and acceptance evidence, and have another Session or teammate inspect the result. Use independent copies for a controlled comparison; Sessions editing one shared file tree do not provide automatic isolation. The first-project guide shows how to start with a bounded task.