Claude Sonnet 5.5 vs Opus 5.5: price, benchmarks

Sonnet vs Opus: how Sonnet 5.5 compares on LLM benchmarks

Sonnet 5.5 tops Sonnet 5 on every row of Anthropic's launch table and beats Opus 5.5 on just one. Anthropic still points to Opus 5.5 for open-ended work that needs sustained judgment, and it adds that benchmark scores show only one side of a model.

Sonnet 5.5 is meant to work alongside Opus 5.5, not replace it. Anthropic pitches it as the faster, cheaper choice for well-scoped everyday tasks and bug fixes, and for documents, slides and spreadsheets.

Terminal Bench and OSWorld scores jump over Sonnet 5

Terminal-Bench 4.0 shows the biggest jump, with Sonnet 5.5 climbing from 10.3% to 70.6%. That beats the 66.4% Opus 5.5 scored at xhigh effort, which Anthropic notes is Opus's best result on the test.

OSWorld 2.1 measures how well a model runs a desktop through computer use, and there Sonnet 5.5 rises from 57.0% to 80.1%. Opus 5.5 still scores 1.7 points higher.

Anthropic's launch benchmarks

Terminal-Bench 4.0 (%)

Agentic coding in a command-line interface. Opus 5.5 at xhigh, its best score.

0204060

70.6Sonnet 5.5

10.3Sonnet 5

66.4Opus 5.5

OSWorld 2.1 (%)

Operating a computer through its screen.

0255075

80.1Sonnet 5.5

57.0Sonnet 5

81.8Opus 5.5

CursorBench 4.0 (%)

Ambiguous, multi-file coding tasks taken from real Cursor sessions.

02040

55.5Sonnet 5.5

34.1Sonnet 5

57.8Opus 5.5

GDPval-AA v2.1 (Elo)

Knowledge work across 44 occupations. Bar height is relative to Opus 5.5 = 100.

05001,0001,500

1844Sonnet 5.5

1449Sonnet 5

1846Opus 5.5

Opus 5.5 leads on GDPval and Cursor Bench by under 3 points

Every charted row besides Terminal-Bench goes to Opus 5.5, each by less than three points. On CursorBench 4.0 it is 57.8% to 55.5%. On GDPval-AA, a knowledge-work test rated in Elo, Opus gets 1846 and Sonnet 1844.

Those gaps are small for a model that costs half as much per token. Benchmarks don't show which jobs suit each model, though, and the task-by-task split below covers that.

Agentic coding and AI slides: which tasks go to Sonnet 5.5

In agentic coding and document work, give Sonnet 5.5 tasks with a clear scope and save Opus 5.5 for work where the plan itself is the hard part. The split below follows what Anthropic and its early customers report.

For slides, Anthropic ran an internal test using a public company's quarterly earnings materials, earnings call transcripts and a slide template. Sonnet 5.5 built a 10-slide operating review, and two experts rated the first draft ready to send without changes.

The two models also pair well. Kevin Ngo of Creator said that once Opus 5.5 sets out a game's architecture and overall framework, he would trust Sonnet 5.5 to build it.

API pricing and token cost: half per token, not always per task

API price per million tokens (USD)

Input

Lower is cheaper. Bar height is relative to Opus 5.5 = 100.

01234

2Sonnet 5.5

2Sonnet 5

4Opus 5.5

Output

Same ratio as the input chart: Sonnet 5.5 is half of Opus 5.5.

05101520

10Sonnet 5.5

10Sonnet 5

20Opus 5.5

Token prices are unchanged from Sonnet 5, at $2 input and $10 output per million. Opus 5.5 lists $4 and $20, so each Sonnet token costs half as much as an Opus token.

The cost of a task also depends on how many tokens the model uses. Sonnet 5.5 keeps Sonnet 5's tokenizer, so the same prompt comes to the same number of tokens, and the savings Anthropic reports come from using fewer tokens on the same job. For full Opus 5.5 pricing, see Claude Opus 5.5 pricing, benchmarks, API changes.

Cost per task falls most at low and medium reasoning effort

Anthropic's tests put Sonnet 5.5 at up to 30% less per task than Sonnet 5, with the gap to Opus widest at lower effort. At low or medium effort it beat Sonnet 5's best scores on several benchmarks while spending about a tenth as much per task. It also returns output more than 30% faster than Sonnet 5.

At higher effort, Anthropic says, Sonnet 5.5 can match Opus 5.5, but at a similar cost per task. So at the top settings the half-price edge can disappear.

The default effort depends on where you use the model. Claude Code and the Claude apps run Sonnet 5.5 at medium, while the Claude API defaults to high, so an API call without the setting usually uses more tokens than the same request in the app.

The effort levels were also recalibrated, so a setting tuned for Sonnet 5 will not carry over. For agentic coding, the Docs suggest medium effort for well-specified tasks and high for harder or longer ones.

Cache pricing: prompt caching reads are $0.20 per million

On Sonnet 5.5, cache reads cost $0.20 per million tokens, the same as on Opus 5.5. Cache writes cost $2.50 per million tokens for five-minute storage and $4 for one hour.

For work that can wait, the Batch API halves input and output prices. All of these rates are the same as Sonnet 5's.

Five breaking changes to check before the API migration

Code written for Sonnet 5 needs more than a new model ID. Sonnet 5.5 rejects five kinds of request that Sonnet 5 accepted.

Accounts opened on or after August 31, 2026 (UTC) get the history check without opting in. One more change alters responses without making them fail: text the model writes between tool calls now comes in thinking blocks, so an interface that streams that text may fall silent between calls.

Adaptive thinking and tool choice: two settings now rejected

To skip up-front thinking on Sonnet 5.5, send between_tools rather than disabled. Sending disabled gets the request rejected, with an error message pointing you to between_tools. The between_tools setting works only at low, medium and high effort; at xhigh or max, leave out the thinking field so adaptive thinking stays on.

Forcing a tool call with tool_choice set to any or tool is also rejected, though auto and none still work. When you need a tool call instead of a text reply, say in the prompt when that tool should be used.

Beyond the model ID, the old computer use tool now fails

With Sonnet 5.5, the Claude API and Google Cloud no longer accept computer_20251124, so computer use agents need computer_toolset_20260801. Amazon Bedrock still accepts the older tool.

The advisor tool no longer accepts Opus 4.8, Opus 4.7 or Sonnet 5 as advisors to a Sonnet 5.5 executor. Opus 5.5 and Opus 5 remain on the supported list.

Summary: Sonnet 5.5 nears Opus at half the token price

Claude Sonnet 5.5 comes within three points of Opus 5.5 on most launch benchmarks at half the token price, and the savings hold best at low or medium effort. Give it bug fixes, scoped coding tasks and document or slide work, and leave the open-ended, judgment-heavy work to Opus 5.5.

In the Claude apps and Claude Code, choosing the new model is all you need to do. On the API, check the five changes above and set effort explicitly, because a request without it runs at high.

The next Haiku came out on October 7, 2026, and is covered in Claude Haiku 5.5 API pricing and benchmarks. Claude models compared: Fable, Opus, Sonnet, Haiku shows how the whole lineup fits together.