Short answer
For well-scoped, everyday coding, evaluate Claude Sonnet 5.5 first: Anthropic positions it for tasks such as building features, fixing bugs, and reviewing code. Try Claude Opus 5.5 when your work involves long-running coding agents, large refactors, or complex systems engineering. That’s a practical starting point—not a guarantee that either model will perform better on your codebase. Anthropic’s descriptions are product guidance; your own task results should decide the production choice. (anthropic.com)
The current Anthropic API model IDs are claude-sonnet-5-5 and claude-opus-5-5. At standard direct API list rates, Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens; Opus 5.5 costs $4 and $20, respectively. So Sonnet’s listed input and output rates are half of Opus’s, but your actual bill depends on token use and any applicable pricing options. (platform.claude.com)
Current models and API prices
| Model | Anthropic API model ID | Standard input price | Standard output price | Anthropic’s stated coding fit |
|---|---|---|---|---|
| Claude Sonnet 5.5 | claude-sonnet-5-5 |
$2 / million tokens | $10 / million tokens | Everyday development, including features, bug fixes, and code review |
| Claude Opus 5.5 | claude-opus-5-5 |
$4 / million tokens | $20 / million tokens | Long-running coding agents, large-scale refactoring, and complex systems engineering |
Prices are in USD per million tokens for the Anthropic API. Anthropic lists both models as active and current; API model IDs may be different on other platforms—for example, Amazon Bedrock uses an anthropic. prefix. If you use a cloud marketplace or another API provider, check that provider’s model identifier, availability, and billable rates rather than assuming Anthropic’s direct API price applies. (platform.claude.com)
A simple cost comparison: for a request using 1 million input tokens and 200,000 output tokens, the listed token charges would be $4 with Sonnet 5.5 ($2 input + $2 output) and $8 with Opus 5.5 ($4 input + $4 output). This illustration excludes caching, batch processing, regional options, and other charges. Actual workloads may use different amounts of input and output; compare usage from your own evaluation. (platform.claude.com)
Anthropic also lists Sonnet 5.5 US-only inference at 1.1× its input and output rates. If that option is required for your workload, use its applicable rate—not the standard figures above—in your estimate. Verify the configured channel and options before projecting production spend. (anthropic.com)
When to test Sonnet—and when to test Opus
Start with Sonnet 5.5 if your tasks are bounded and easy to review: implementing a defined feature, fixing a reproducible bug, or reviewing a modest change. Anthropic describes Sonnet as a faster, lower-cost complement to Opus for well-scoped coding and everyday development. (anthropic.com)
Include Opus 5.5 in your evaluation if the task must span a longer autonomous run, touch a large system, or involve substantial refactoring. Anthropic’s model documentation lists multihour autonomous coding agents, large-scale refactoring, and complex systems engineering among its typical uses. These descriptions are useful for deciding what to test, but they do not establish that Opus will win on every complex task. (platform.claude.com)
What an older head-to-head test can—and cannot—tell you
A third-party comparison dated February 17, 2026 tested Sonnet 4.6 and Opus 4.6 by asking each to generate a tower-defense game. Both received 9/9 against the author’s checklist. The author preferred Sonnet 4.6’s interface and gameplay feel, and claimed it was “almost 50% cheaper,” but supplied no prices, billing units, or calculation. The article’s separate chat-app test covered Sonnet 4.6 only, not Opus. These are author-reported results from specific prompts—not a repeatable benchmark, current price comparison, or direct test of Sonnet 5.5 versus Opus 5.5.
That test can still suggest a useful evaluation habit: turn a broad coding request into observable pass/fail checks. Its results should not decide which current API model you deploy. The versions, task, and pricing have changed, and the article does not show code artifacts, repeated trials, or measured API usage. (platform.claude.com)
A practical comparison for your codebase
- Choose representative tasks. Include work your application actually sends to the model—for example, a routine feature, a bug with a reproducible test, and a multi-file change if that is part of your workload.
- Set acceptance criteria before running either model. Specify what must work, which tests must pass, what files may change, and any constraints such as backwards compatibility. For a game-generation task, criteria might include functioning controls, correct scoring, and a successful fresh launch.
- Keep the comparison fair. Give both models the same task, relevant repository context, tools, permissions, and time or turn limits. Use separate clean branches or equivalent snapshots so one model’s edits do not affect the other’s result.
- Check the outputs, not just the explanations. Run the same tests, inspect the diff, and record whether each result meets the agreed criteria. Include review effort and necessary fixes in your assessment.
- Measure production-relevant usage. Record input and output tokens and the configured API rates for each completed task. If reliability matters, repeat representative tasks and track success across runs; one successful generation is not a dependable estimate.
- Choose by your own quality bar and total cost. If Sonnet passes often enough at lower measured spend, it may be the better default for that workload. If Opus materially improves completion or reduces correction work on harder tasks, reserve it for those tasks—or compare the value of that improvement against the extra cost.
This method adapts the checklist approach used in the older Converge test, but its suggested steps are an evaluation plan, not reported experimental results. Neither that comparison nor Anthropic’s model descriptions can substitute for testing the exact model IDs, API channel, and coding tasks you intend to use. (platform.claude.com)
Bottom line
For coding through Anthropic’s API, Sonnet 5.5 is a sensible first model to evaluate for well-scoped everyday work; test Opus 5.5 for long-running agents, large refactors, and complex systems tasks. The current standard list rates are $2/$10 per million input/output tokens for Sonnet and $4/$20 for Opus. Treat that as a price-rate comparison, not a promise that the total cost of a task will be exactly half. Decide from comparable tests on your own workload, using the exact API channel and rates you plan to run in production. (anthropic.com)
Sources
Claude Sonnet 5.5 — Claude Platform Docs
Claude Opus 5.5 — Claude Platform Docs