Modelpeers
Claude Sonnet 5.5 charactervsClaude Opus 5.5 character

Claude Sonnet 5.5 vs Claude Opus 5.5: Is Opus worth twice the price?

Facts checked Oct 22, 2026. Every number links to its source.

Summary

Derived from sourced facts, not from builder opinions
  • Sonnet 5.5 costs half as much: $2 input and $10 output per million tokens, against $4 and $20 for Opus 5.5.
  • With Claude Code as the agent, Opus 5.5 scored 64.85% on Terminal-Bench 4.0 and Sonnet 5.5 61.82%, within the 95% confidence intervals.
  • On the Vals Index, Sonnet 5.5 ranks first at 70.85%, ahead of Opus 5.5 at 70.18%, and costs $18.70 per test against $28.12. Vals AI counted some Sonnet 5.5 and Opus 5.5 tasks that a fallback model answered after a refusal. Counting those as failures gives 69.03% and 64.16%.
  • Anthropic says Opus 5.5 remains clearly stronger at complex, open-ended work that needs sustained judgment.

Choose Claude Sonnet 5.5 if

  • You run many well-scoped tasks where cost and speed matter.
  • You need to turn off up-front thinking: Sonnet 5.5 accepts between_tools, and Opus 5.5 always thinks.
  • You want lower latency: Anthropic rates Sonnet 5.5 Fast and Opus 5.5 Moderate.
  • You want half the per-token price.

Choose Claude Opus 5.5 if

  • You run long, open-ended agent tasks that need careful judgment.
  • You run hard terminal coding tasks, where Vals AI measured Opus 5.5 cheaper per test ($13.20 against $16.51).
  • You want the slightly higher Terminal-Bench 4.0 score in Claude Code (64.85% against 61.82%, within the confidence intervals).
Claude Sonnet 5.5
Claude Opus 5.5

Follow both to get an email when either one changes its price, limits or behavior.

Best by task

Agentic coding (Terminal-Bench 4.0 official board, both in Claude Code)

Tied

64.85% against 61.82%, both at max effort, 330 trials each. The 95% confidence intervals overlap (plus or minus 3.11 and 2.94).

Agentic coding (Terminal-Bench 4.0, Vals AI, Mini-SWE-agent harness)

Tied

65.15% for Opus 5.5 against 64.14% for Sonnet 5.5, too small to call. Opus 5.5 cost less per test ($13.20 against $16.51).

Real-world work across industries (Vals Index)

Claude Sonnet 5.5

70.85% against 70.18%, at $18.70 per test against $28.12. Vals AI counted some Sonnet 5.5 and Opus 5.5 tasks that a fallback model answered after a refusal. Counting those as failures gives 69.03% and 64.16%.

Overall capability (Epoch Capabilities Index)

Tied

168.83 for Opus 5.5 against 166.18 for Sonnet 5.5. The 95% confidence intervals overlap (165.25 to 173.95 and 163.06 to 171.34).

Coding in Cursor sessions (CursorBench 4.0, reported by Anthropic)

Claude Opus 5.5

57.8% against 55.5%. Maker-reported, not independent.

Cost

One request in an agent loop that rereads a repository, so most input is cached. A full task makes many such requests. Illustrative, not measured.

Claude Sonnet 5.5

$0.062 per request

$62.00 a month for 1,000 requests

Claude Opus 5.5

$0.124 per request

$124.00 a month for 1,000 requests

Claude Sonnet 5.5 costs 50% less for this workload at list prices.

Estimate per request: (input × input rate + cached input × cached rate + output × output rate) ÷ 1,000,000, at list prices on the cited pricing pages. It is not a measured task cost: models use different numbers of tokens, including reasoning tokens, for the same task.

Side by side

Claude Sonnet 5.5Claude Opus 5.5
MakerAnthropic
Anthropic
Model IDclaude-sonnet-5-5 (Bedrock: anthropic.claude-sonnet-5-5)
claude-opus-5-5 (Bedrock: anthropic.claude-opus-5-5)
Released2026-09-28
2026-09-22
StatusActive (latest)
Not published
API price per 1M tokens$2 input / $10 output · 5-minute cache write $2.50, 1-hour cache write $4, cache read $0.10 · Batch 50% off
$4 input / $20 output · 5-minute cache write $5, 1-hour cache write $8, cache read $0.20 · Batch 50% off
Context window1M tokens
1M tokens
Max output128K tokens (300K on Batch API with beta header)
128K tokens (300K on Batch API with beta header)
Knowledge cutoffJune 2026 (reliable knowledge and training data)
June 2026 (reliable knowledge and training data)
Reasoning settingsAdaptive thinking, default effort high · up-front thinking can be turned off with between_tools at high effort or below
medium · adaptive thinking always on
SpeedRated Fast by Anthropic (Opus 5.5 is Moderate)
Not published
Breaking changes from Sonnet 5between_tools replaces disabled thinking · forced tool use returns an error · thinking blocks tied to model and conversation
Not published
Where to use itClaude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS · also Claude Code (sonnet alias)
Claude API, Amazon Bedrock, Google Cloud (Google managed and Anthropic managed), Microsoft Foundry, Claude Platform on AWS · also Claude Code
Retires no sooner than2027-09-28
2027-09-22
Fast mode priceNot published$8 input / $40 output (up to 2.5x speed, Claude Code and Claude Platform)
Quality tracking (MarginLab)Not publishedClaude Code with Opus 5.5/high on a SWE-Bench-Pro subset: 77% latest day (50 tasks), 76% 7-day, 81% 30-day · baseline still being collected, detection paused (last updated 2026-09-29)
Quality tracking (Livenerf)Not publishedWindow 1 (days 11-20) score 0.592 vs baseline 0.584, delta +0.007 plus or minus 0.018: no change
Terminal-Bench 4.0 (official board)61.82%rank 2, Claude Code, max effort, 330 trials64.85% (plus or minus 3.11)rank 1, Claude Code, max effort, 330 trials, $4,718.65 total
Terminal-Bench 4.0 (Vals AI, same harness for all)64.14%rank 2, $16.51 per test65.15%rank 1, $13.20 per test
Vals Index v2.270.85%rank 1, $18.70 per test, includes tasks a fallback model answered after a refusal (69.03% if those count as failures)70.18%rank 2, includes tasks a fallback model answered after a refusal (64.16% if those count as failures)
Epoch Capabilities Index166.18(CI 163.06 to 171.34)168.83(95% CI 165.25 to 173.95), highest in the file

Scores appear side by side only when both were published for the same test. Settings differ between sources, so small gaps can reverse under another harness.

Updates

Both Claude Sonnet 5.5 and Claude Opus 5.5

Now available

Full Claude Platform available on Google Cloud (Anthropic managed)

Claude Models on Google Cloud (Anthropic managed) is generally available with US and Global endpoints. It gives Google Cloud customers the full Claude Platform feature set with Google Cloud IAM, billing and commitment drawdown. Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 are available, and fast mode for Opus 5.5 costs $8 / $40 per million input and output tokens there.

Who is affected:
Teams that buy Claude through Google Cloud and want features that the Google-managed offering lacks.
What to do:
No action needed. Google Cloud customers who need same-day Claude API features can switch to the Anthropic-managed offering.

Claude Sonnet 5.5

3 recorded incidents in the last 90 days.

Claude Opus 5.5

3 recorded incidents in the last 90 days.

FAQ

Is Claude Sonnet 5.5 cheaper than Opus 5.5?

Yes. Sonnet 5.5 costs $2 input and $10 output per million tokens, half of Opus 5.5's $4 and $20.

Is Opus 5.5 better than Sonnet 5.5 at coding?

Not clearly. On Terminal-Bench 4.0 in Claude Code, Opus 5.5 scored 64.85% and Sonnet 5.5 61.82%, within the 95% confidence intervals. Anthropic reports Opus 5.5 ahead on CursorBench 4.0 (57.8% against 55.5%).

Which is faster, Sonnet 5.5 or Opus 5.5?

Anthropic rates Sonnet 5.5's latency as Fast and Opus 5.5's as Moderate.

Can I turn off thinking on Sonnet 5.5?

Partly. Setting thinking to between_tools turns off up-front thinking at high effort or below. Opus 5.5 does not allow turning thinking off.

Do Sonnet 5.5 and Opus 5.5 have the same context window?

Yes. Both take 1M tokens with 128K of output.

Which ranks higher on the Vals Index?

Sonnet 5.5, at 70.85% (rank 1) against Opus 5.5 at 70.18% (rank 2).

Which is cheaper per test on Terminal-Bench?

Opus 5.5, at $13.20 per test against $16.51 for Sonnet 5.5 in Vals AI's run.

Related comparisons

Claude Opus 5 vs Claude Opus 5.5Claude Opus 5.5 vs GPT-6 AstraClaude Sonnet 5.5 pageClaude Opus 5.5 page

Sources and method

Prices, specs and limits come from each maker's own pages. Benchmark scores come from independent benchmark publishers and the makers' announcements, with the setting each was measured in. Changes come from changelogs, deprecation pages and status pages. The summary is written from these facts only.