Models & LLMsComparison

Claude Sonnet 5.5 vs Opus 5.5: Which Model to Use

Claude Sonnet 5.5 costs half as much per token as Opus 5.5 and runs faster. Use it for well-scoped work, keep Opus 5.5 for long, open-ended judgment, and fix the settings that return 400 when you move from Sonnet 5.

Toolbit AI - Team
6 min read
Claude Sonnet 5.5 vs Opus 5.5: Which Model to Use

Use Claude Sonnet 5.5 as your default for well-scoped work. Send long, open-ended judgment to Claude Opus 5.5. Sonnet 5.5 costs $2 / $10 per million input / output tokens; Opus 5.5 costs $4 / $20. Anthropic rates Sonnet "Fast" and Opus "Moderate". Moving from Sonnet 5? Four settings now return a 400 error, listed below.

Quick verdict

  • Pick Sonnet 5.5 if the task is well specified: bug fixes, scoped coding agents, docs, high-volume loops.
  • Pick Opus 5.5 if the work is long-running and open-ended, you have no evals yet (Anthropic's docs say start there), or your evals show Sonnet falls short.
AxisSonnet 5.5Opus 5.5Winner
Price per token$2 / $10 per MTok$4 / $20 per MTokSonnet
Speed (Anthropic's class)FastModerateSonnet
ThinkingAdaptive; can switch to between_toolsAdaptive, always on (disabled returns 400)Depends
API default efforthighmediumDepends
Open-ended judgment"well-scoped everyday tasks""complex work requiring careful judgment"Opus

How we compared

We read Anthropic's models overview, pricing page, effort docs, migration guide and both launch pages on October 21, 2026. We compared list prices, defaults and documented API behavior. We did not run either model hands-on, so speed and benchmark figures are Anthropic's own (vendor).

Pricing: Sonnet 5.5 costs half per token, but per-task cost depends on tokens used

Sonnet 5.5 is exactly half the price of Opus 5.5 per token. Prices checked October 21, 2026, on the Claude API:

Sonnet 5.5Opus 5.5
Model IDclaude-sonnet-5-5claude-opus-5-5
Input / output$2 / $10 per MTok$4 / $20 per MTok
Cache read$0.10 per MTok$0.20 per MTok
Cache write$2.50 per MTok$5 per MTok
Context / max output1M / 128K tokens1M / 128K tokens
RetirementNot sooner than September 28, 2027Not sooner than September 22, 2027

Correction: an earlier version of this post listed a $0.20 cache read for both models. Sonnet 5.5 is $0.10.

A worked example: 1M input plus 100K output tokens costs $3.00 on Sonnet 5.5 and $6.00 on Opus 5.5 at list price. That's before caching, and batch processing saves 50%. Real per-task cost depends on how many tokens each model spends. Measure that on your own tasks. Our guide to what an AI agent run costs shows how to budget it.

Other price notes. US-only inference on Sonnet 5.5 is priced at 1.1x. Opus 5.5 fast mode costs $8 / $40 per million tokens. Anthropic says Sonnet 5.5 costs "up to 30% less per task" and runs "30%+ faster" (vendor). That comparison is with Sonnet 5, not Opus 5.5. Bedrock IDs add an anthropic. prefix; Google Cloud and Foundry use the Claude API IDs.

Which should you pick? Route well-scoped work to Sonnet, open-ended work to Opus

Route by how well the task is specified, not by a benchmark headline. Anthropic calls Sonnet 5.5 "a faster, lower-cost complement to Claude Opus 5.5" (vendor). Its docs describe Opus 5.5 as built "for long-running agentic coding and knowledge work".

Routing guide: well-scoped coding and everyday tasks go to Sonnet 5.5; long-running open-ended judgment goes to Opus 5.5; the Free plan includes Sonnet but not Opus
  1. Well-specified coding, fixes, docs and cost-sensitive loops go to Sonnet 5.5.
  2. Long-running, open-ended work that needs careful judgment goes to Opus 5.5.
  3. On the Free plan, Sonnet is included and Opus is not. Pro, Max, Team and Enterprise get both.
  4. No evals yet and unsure? Anthropic's docs say to start with Opus 5.5 for most workloads.

Rule 4 is the vendor's safe default for a workload you haven't measured. Rule 1 is the cheaper one, so prove it first with the bake-off below. For cross-vendor routing, see how to pick the right AI model for every task and our look at model routers and their cache math.

Benchmarks are vendor-reported and not like-for-like

Anthropic's Terminal-Bench 4.0 table puts Sonnet 5.5 at 70.6% and Opus 5.5 at 66.4% (vendor). The rows aren't comparable. Opus ran "at Xhigh effort"; the effort behind Sonnet's 70.6% is unconfirmed. Anthropic also reports 55.5% for Sonnet 5.5 on CursorBench 4.0, "second only to Opus 5.5" (vendor). Treat both as orientation, not proof Sonnet replaces Opus. Run your own eval set.

Moving from Sonnet 5 to Sonnet 5.5 breaks four settings

Swap the ID to claude-sonnet-5-5, then fix these before you move traffic. The Sonnet 5.5 migration guide has the full list.

  1. Thinking off. thinking: {"type": "disabled"} returns 400. Send {"type": "between_tools"} instead. It means the model skips thinking except between tool calls. It works at low, medium and high only.
  2. Forced tools. tool_choice of any or tool returns 400. Send tool_choice: {"type": "auto"} and mark the tool strict: true.
  3. Computer use. On the Claude API and Google Cloud, use computer_toolset_20260801. There, computer_20251124 returns 400. On Amazon Bedrock, the checklist says to move to computer_20251124.
  4. Thinking blocks. Sonnet 5.5 doesn't read blocks from Opus 5, Opus 5.5, Fable or Mythos models. The API drops them, returns 200 and doesn't bill them. Replaying a block after editing earlier history returns 400 on accounts created on or after August 31, 2026.
Sonnet 5 to 5.5 thinking swap: disabled returns 400; migrate to between_tools at low, medium or high; between_tools is rejected at xhigh and max

Two more checks. Remove non-default temperature, top_p and top_k values, which return 400. Text between tool calls now comes back in thinking blocks, so a UI that streams only text can go quiet. Set display: "updates" (beta, header thinking-display-updates-2026-08-18) to stream those progress notes. Then re-run your effort sweep, because levels are recalibrated.

Treat migration day as a bake-off: run the same tasks on both model IDs, log effort, latency, tokens and tool-call success, then promote the winner per workload class.

Start Sonnet 5.5 at medium effort, then move to high

Effort sets how much the model thinks before it answers. Start Sonnet 5.5 at medium for well-specified agentic coding. Anthropic's effort docs say to move to high for harder or longer tasks.

LevelSonnet 5.5 guidanceWatch-outs
mediumWell-specified agentic coding; default in Claude Code and the appsRaise for harder tasks
highClaude API default; harder or longer tasksMore thinking tokens
xhigh / max"only where your evals show a quality gain"between_tools returns 400

Opus 5.5 defaults to medium on the API, and its adaptive thinking is always on. Sending disabled to Opus 5.5 returns 400 at every effort level. Its default in the apps is unconfirmed.

Set max_tokens to 128,000 for agentic coding and stream the response. Thinking counts toward max_tokens even when it isn't returned. Per-message effort (beta) can't change mid-conversation when you use between_tools.

Pick neither if your workload sits outside this pair

Pick neither if your evals show both models miss the bar, or a cheaper tier fits your volume. Anthropic's lineup also includes Claude Fable 5.1 and Claude Mythos 5.1; check the vendor page for current prices. If price across vendors matters more, compare our Opus 5.5 API price roundup and the Claude overview.

Decision rule: no evals yet? Anthropic's docs say start on Opus 5.5. Once a small eval shows Sonnet 5.5 at medium meets the bar for a workload, move it to Sonnet and try high before going back to Opus.

Share this article

Related articles

Continue exploring similar guides and insights