← Back to blog
Sonnet 5.5 vs Sonnet 5: Benchmarks, Cost per Task & Migration

Sonnet 5.5 vs Sonnet 5: Benchmarks, Cost per Task & Migration

Emma Reed

By Emma Reed

MyClaw Editorial

Compare in MyClaw

Compare Sonnet 5.5 and Sonnet 5 on your own work

Run the same task in both models to compare quality, tool behavior, speed, and cost per accepted result before migrating.

AI Takeaway

  • Is Sonnet 5.5 better than Sonnet 5? Yes. In the Sonnet 5.5 vs Sonnet 5 matchup, Anthropic's numbers show CursorBench 4.0 rising from 34.1% to 55.5%, with replies arriving 30%+ faster.
  • Does Sonnet 5.5 cost more than Sonnet 5? No, the token price is identical: $2 in and $10 out per million tokens, per Anthropic. The bill per task can still differ, depending on effort.
  • What breaks when you move from Sonnet 5? Anthropic flags five breaking changes. The two most common are that disabled thinking becomes between_tools and forced tool calls now fail.
  • When is it smarter to stay on Sonnet 5? Stay when your code relies on forced tool use, thinking off at top effort levels, or edited chat history. Sonnet 5 remains supported until at least June 30, 2027.
  • Can you compare Sonnet 5.5 and Sonnet 5 on MyClaw? Yes. MyClaw runs both models in managed OpenClaw and Hermes agents, so you can test them on the same job, and Opus 5.5 is available for tougher work.

Sonnet 5.5 vs Sonnet 5: The Spec Sheet Side by Side

Same price and window, fresher knowledge, stricter API. From Anthropic's model docs:

SpecSonnet 5.5Sonnet 5
ReleasedSeptember 28, 2026June 30, 2026
StatusLatestLegacy (Anthropic suggests migrating)
RetirementNot before September 28, 2027Not before June 30, 2027
Model IDclaude-sonnet-5-5claude-sonnet-5
Context / max output1M / 128K tokens1M / 128K tokens
Knowledge cutoffJune 2026January 2026
API price (per Anthropic)$2 in / $10 out per 1M$2 in / $10 out per 1M
Lowest thinking settingbetween_toolsdisabled
Forced tool useRejectedAllowed
Smallest cacheable prompt512 tokens1,024 tokens
Per-message effort and mid-conversation system messagesYesNo

They share a tokenizer, so any change in your bill comes from how many tokens each model spends on the job. Our original Sonnet 5 review covers the older model in depth.

Introducing Claude Sonnet 5.5 \ Anthropic

Where Sonnet 5.5 Pulls Ahead, Benchmark by Benchmark

Anthropic published these scores at launch. The Gain column is our own arithmetic:

BenchmarkSonnet 5.5Sonnet 5Gain
Terminal-Bench 4.070.6%10.3%+60.3 pts
FrontierCode 1.152.1% (Xhigh)42.4%+9.7 pts
CursorBench 4.055.5%34.1%+21.4 pts
GDPval-AA v2.118441449+395
AA-Briefcase v1.118111359+452
Humanity's Last Exam (tools)64.5%54.9%+9.6 pts
OSWorld 2.180.1%57.0%+23.1 pts
Chartography61.6%15.6%+46.0 pts

Agentic coding and computer use improve most, knowledge-work scores climb roughly 400 points, and chart reading turns from weak spot to strength.

Grains of salt: the vendor ran these tests on a days-old model. A 10.3% Terminal-Bench baseline is oddly low for Sonnet 5, so don't over-read that jump. Its Max run on FrontierCode (46.2%) trailed Xhigh, which Anthropic blames on timeouts and off-task edits. GDPval-AA and AA-Briefcase used a pre-release build with a bug that's now fixed. For the Opus 5.5 comparison, head to our Sonnet 5.5 review.

Chat with Claude Sonnet 5.5

Identical Rate Card, Different Receipts

Per Anthropic, the prices match line by line:

Rate (per Anthropic, per 1M tokens)Sonnet 5.5Sonnet 5
Input / output$2 / $10$2 / $10
Cache hits$0.20$0.20
Cache writes (5 min / 1 hour)$2.50 / $4$2.50 / $4
Batch processingHalf priceHalf price

Sonnet 5's $2/$10 was an introductory rate that Anthropic kept instead of moving to the planned $3/$15.

The difference is token count. Anthropic says Sonnet 5.5 finishes work with fewer tokens and fewer tool-call rounds, cutting cost per task by up to 30%. On several benchmarks, its Low or Medium setting tops Sonnet 5's best result at roughly one-tenth the cost per task. One early tester's finance workload dropped from about 497k tokens per answer on Sonnet 5 to about 121k on Sonnet 5.5.

Don't bank on it blindly: some independent tests suggest it can spend more on open-ended prompts at high effort. The API defaults to High, so choose effort on purpose and track cost per finished task. For a similar upgrade decision on the flagship line, see Opus 5.5 vs Opus 5.

Quiet Changes You'll Spot Without Touching Code

Even with zero errors, the switch can feel different:

  • Effort levels mean something new. Anthropic recalibrated them, so "Medium" on Sonnet 5.5 isn't "Medium" on Sonnet 5. Its advice: High in general, Medium for well-defined agent work, Medium or Low for chat.
  • Progress notes can vanish. Longer notes written between tool calls now come back as thinking blocks, blank at the default display setting.
  • Refusals get labels. Declines carry one of five categories: cyber, bio, frontier_llm, reasoning_extraction or general_harms. Riskier cyber requests visibly hand off to Sonnet 5.
  • Reasoning is bound. Sonnet 5.5 thinking blocks only work in the conversation and account that created them.

Introducing Claude Sonnet 5 \ Anthropic

Moving Off Sonnet 5: A Six-Point Checklist

Swap in claude-sonnet-5-5, then work through Anthropic's list:

  1. Thinking off. Send between_tools instead of disabled. It's only accepted at Low, Medium or High effort, so Xhigh and Max need adaptive thinking.
  2. Tool choice. Replace any or tool with auto plus strict tools; on Amazon Bedrock, drop strict and validate in code.
  3. Chat history. Append rather than edit. Change instructions with a mid-conversation system message.
  4. Computer use. Claude API and Google Cloud users move to computer_toolset_20260801.
  5. Advisors. Pick one Sonnet 5.5 accepts, such as Opus 5.5 or Sonnet 5.5 itself. Sonnet 5 no longer qualifies.
  6. UI updates. Set thinking.display if users watch progress notes between tool calls.

Finish by re-testing effort, re-baselining spend and shifting a small share of traffic first. Bonus: Sonnet 5.5 can read Sonnet 5's thinking blocks, so live conversations keep their reasoning when they move up.

Good Reasons to Keep Sonnet 5 a Little Longer

Anthropic wants you to migrate, but the clock isn't tight. Holding off is sensible if you:

  • Rely on forced tool calls you can't rewrite yet
  • Need thinking off at Xhigh or Max effort
  • Edit earlier turns instead of appending
  • Use Opus 4.8 or Opus 4.7 as your advisor model

Sonnet 5 stays supported until at least June 30, 2027. The real trade-off is its January 2026 knowledge cutoff.

Compare in MyClaw
Compare Sonnet 5.5 and Sonnet 5 on your own work

Run the same task in both models to compare quality, tool behavior, speed, and cost per accepted result before migrating.

Chat with Claude Sonnet 5

Sonnet 5.5 vs Sonnet 5 Verdict: Who Should Upgrade Today

If none of those blockers apply, upgrade. For most teams the Sonnet 5.5 vs Sonnet 5 call is some migration work in exchange for a clearly stronger model at the same list price.

Pros:

  • Much higher Anthropic-reported scores at an unchanged $2/$10 (per Anthropic)
  • Faster replies and leaner token use per task
  • Fresher knowledge plus new mid-conversation controls
  • Smooth hand-off, since it reads Sonnet 5 thinking blocks

Cons:

  • Five breaking API changes to work through
  • Savings hinge on effort; high settings can cost more
  • More refusal categories, with some cyber tasks routed to Sonnet 5
  • Scores are vendor-run and very fresh
WorkloadBetter Fit
Everyday coding and bug fixesSonnet 5.5 at Medium
Reports, slides and chartsSonnet 5.5
Pipelines that force tool callsSonnet 5 until migrated
Complex, open-ended projectsConsider Opus 5.5

Choosing for agents? See our top model picks for OpenClaw.

Let Sonnet 5 and Sonnet 5.5 Race on the Same Job

Benchmarks only go so far; the fastest way to settle this is running both on real work. MyClaw lets you pick either model inside a managed OpenClaw or Hermes Agent, so you can give each the same recurring task and see which delivers better, faster, with fewer retries.

Blocked workflows stay on Sonnet 5 while the rest move; bring in Opus 5.5 for complex, open-ended jobs. Agents run 24/7 with no server to look after. For the flagship, read our Opus 5.5 deep dive.

Set Up a Sonnet 5 vs Sonnet 5.5 Test on MyClaw in 3 Steps

  1. Start an agent. Create a MyClaw account and launch OpenClaw or Hermes Agent.
  2. Add both models. Select Claude Sonnet 5.5 and Claude Sonnet 5, optionally keep Opus 5.5 for heavy lifting, and link your chat app.
  3. Compare on real work. Assign both the same recurring task, then check quality, speed and retries.

Our Hermes Agent hosting guide explains the managed setup.

Compare in MyClaw

Compare Sonnet 5.5 and Sonnet 5 on your own work

Run the same task in both models to compare quality, tool behavior, speed, and cost per accepted result before migrating.

Sonnet 5.5 vs Sonnet 5: Benchmarks, Cost per Task & Migration | MyClaw.ai