What the specifications actually tell you
The following standard Claude API rates and settings were checked on 2 October 2026. Prices are in US dollars per million tokens. They are not Claude subscription charges or the full cost of completing a task. [3] [4]
| Specification | Opus 5.5 | Sonnet 5.5 |
|---|---|---|
| Uncached input / million tokens | $4 | $2 |
| Output / million tokens | $20 | $10 |
| Cache read / million tokens | $0.20 | $0.20 |
| Context window | 1M tokens | 1M tokens |
| Default API effort | medium | high |
Cache reads are a separate billing category from ordinary input. Writing a cache has its own charges; the table also excludes batch discounts and separately priced Opus fast mode. Do not apply the input-price ratio to every part of a workload. [3] [4]
The shared context limit tells you how much material can fit, not how accurately each model will use it. A large document collection can contain irrelevant passages, incompatible definitions and outdated instructions. Test whether the answer identifies those conflicts, rather than treating successful upload as successful understanding.
Route by ambiguity, not by job title
Consider three hypothetical assignments. These examples describe evaluation choices; neither model was run on them for this article.
- A localized software fix: a reproducible bug, a small affected component and a clear test. Start by evaluating Sonnet. Escalate if the repair reaches into shared behavior or the original explanation stops fitting the evidence.
- A system-wide change: several services disagree about how to interpret the same field. Include Opus immediately, with checks for compatibility and missing dependencies. The hard part is deciding what must remain consistent.
- A management brief: a summary of an approved source pack is a Sonnet candidate. A recommendation that must reconcile contradictory reports deserves a comparison with Opus. Require both to expose unresolved assumptions.
One team can therefore have several sensible defaults. “Coding” or “writing” is too broad a routing rule. The useful questions are whether the task is bounded, whether errors are easy to detect, and how much judgment the acceptance decision requires.
Make effort part of the decision
Anthropic says Sonnet at higher effort can approach Opus on some evaluations, while still positioning Opus as stronger on complex, open-ended work. That is a qualified vendor claim, not evidence of a universal winner. [2]
API defaults differ, so record effort explicitly. Opus uses adaptive thinking that cannot be switched off; Sonnet permits disabling up-front thinking with between_tools at high effort or below. That setting does not mean all thinking is disabled. [3] [4]
Compare the configurations you would actually deploy. Keep source material, tool permissions and acceptance requirements fixed, then inspect correctness, omitted constraints and time spent repairing the result. Treat increasing effort and switching models as separate experiments; changing both together makes the cause of an improvement harder to identify.
A useful escalation rule names an observable failure: an unsupported conclusion, a broken dependency or an unresolved contradiction. Move difficult cases with that evidence attached. Keep specialist review where the consequences require it. The practical choice is a tested route for each kind of work, with a clear reason to change routes.
Read the underlying guidance.
- Introducing Claude Opus 5.5Anthropic · checked 2026-10-02. 22 September 2026 launch and vendor positioning for demanding, extended coding and knowledge work. No independently reproduced performance result.
- Introducing Claude Sonnet 5.5Anthropic · checked 2026-10-02. 28 September launch; bounded everyday-task positioning; qualified vendor comparison with Opus and dependence on effort. Not proof of performance on the hypothetical tasks.
- Claude Opus 5.5 — Claude Platform DocsAnthropic · checked 2026-10-02. Standard input/output/cache-read rates, separate cache-write and fast-mode pricing, batch discount, 1M context, medium default API effort and always-on adaptive thinking.
- Claude Sonnet 5.5 — Claude Platform DocsAnthropic · checked 2026-10-02. Standard input/output/cache-read rates, separate cache-write pricing, batch discount, 1M context, high default API effort, between_tools behavior and supported effort range.
Researched and written with AI. The producing AI checked the English, source scope, model settings and price categories. A separate AI reviewer independently read the official sources and reviewed the English text on 2 October 2026. Human English and specialist review have not been completed. human English and specialist review; controlled same-workload Opus 5.5 versus Sonnet 5.5 evaluation with explicit effort settings, actual timing and billing evidence.
