Opus 5.5 vs Sonnet 5.5

The pairing most teams now weigh

These are the two current models most production work chooses between. Anthropic positions Opus 5.5 as the default for most work, including complex agentic coding, and Sonnet 5.5 as the faster option for everyday coding, agent and enterprise work — with an Opus model the better choice for the hardest long-horizon tasks. Opus 5.5 costs more per token for input and output, as the table shows.

The detail that narrows the gap: cache reads

Read the cache row before the headline rows. Opus 5.5 discounts its cache reads more deeply than Sonnet 5.5 does, by enough that the two cache-read rates in the table come out the same. For an agentic workload where most input tokens are re-read from cache, the input side of the price gap largely disappears, and what remains is output plus fresh input. The effective difference on that kind of workload is much smaller than the headline suggests — run your own token mix through the token & cost estimator rather than trusting either impression.

Thinking controls differ

Opus 5.5 cannot switch thinking off at all; effort is the only dial, and it defaults to medium. Sonnet 5.5 defaults to high, and while it also rejects disabled, it offers between_tools as a thinking-free setting at effort high or below. For latency-critical routes that must not think, Sonnet 5.5 is the only one of the two with that option.

Choosing

For routine, well-scoped work, start on Sonnet 5.5 at a low or medium effort and move a route up only when an evaluation shows the gap. For long, multi-step work in a large codebase, start on Opus 5.5 at its default and test the neighbouring levels. Both reject forced tool choice, so the same request code serves either.

Related

See the model picker and Sonnet 5.5 vs Haiku 4.5.