If you read our post on Opus 5.5's pricing cut, the instinct here is to expect the same shape of story: new model, lower rate, update the spreadsheet. Sonnet 5.5 doesn't work that way. Pull up Anthropic's pricing page and the numbers for Sonnet 5.5 are identical to Sonnet 5's, down to the decimal: $2 per million input tokens, $10 per million output, $0.20 for cache reads, $2.50 for a five-minute cache write. The sticker didn't move at all.
So what actually changed?
The model got faster and more token-efficient, and Anthropic is specific about both. Output generates more than 30% faster than Sonnet 5 — the company calls it "our fastest Sonnet model to date." And because it typically needs fewer tokens to reach the same answer, most tasks end up costing up to 30% less overall, even though every token still costs exactly what it did before.
The price per token didn't move. The number of tokens it takes to finish the job did.
That distinction matters more than it sounds like it should. A flat price cut saves you the same percentage no matter what you're running. A token-efficiency gain saves you more on some workflows than others, depending on how chatty the model used to be on that specific task. You can't just apply a blanket discount to last month's bill and call it done.
What the benchmarks actually show
Anthropic's own numbers back up the token story, not just the marketing line. On Terminal-Bench 4.0, Sonnet 5.5 at Medium effort beats Sonnet 5's best score for under a tenth of the cost per task. On FrontierCode at High effort, it scores ten points above Sonnet 5 at roughly a fifteenth of the cost. A finance-evaluation workload one early tester ran needed about 121,000 tokens per answer on Sonnet 5.5, against 497,000 on Sonnet 5 — a 76% drop for the same answer quality.
Other early numbers Anthropic published: Slackbot saw about 14% fewer output tokens on its workloads. Zendesk reported tickets processed 20% faster. Box measured 2.4x faster runs with 12% fewer total tokens. Atlassian clocked agents running up to 30% faster. None of these are universal multipliers — they're evidence that the token savings are real and show up differently depending on what the agent is actually doing.
Running the math on a workflow that fits this pattern
Take a support-ticket triage agent handling 200 tickets a day — reads the ticket, checks it against a policy doc, drafts a routing decision. Call it 3,000 input tokens and 500 output tokens per ticket on Sonnet 5. At $2/$10 per million tokens: (200 × 3,000 ÷ 1,000,000 × $2) + (200 × 500 ÷ 1,000,000 × $10) = $1.20 + $1.00 = $2.20 a day, about $66 a month.
Same rate on Sonnet 5.5, but say the output side shrinks by a third — closer to the finance-evaluator's token drop than Slackbot's milder one, since triage decisions are exactly the kind of structured output that benefits from fewer wasted tokens. Output falls to roughly 330 tokens a ticket: (200 × 3,000 ÷ 1,000,000 × $2) + (200 × 330 ÷ 1,000,000 × $10) = $1.20 + $0.66 = $1.86 a day, about $56 a month. Ten dollars a month on a light workflow like this one. Run something output-heavy — a longer draft, a multi-step research summary — and the same percentage token drop moves a much bigger number, because output tokens cost five times what input tokens do on this model.
What to actually do about it
Swapping in Sonnet 5.5 is a model-ID change, not a rebuild — claude-sonnet-5-5 is live on the Claude API, AWS, Google Cloud, and Azure, with a zero-data-retention option if your workflow needs it. The one setting worth checking before you flip the switch: Sonnet 5.5 defaults to Medium effort in Claude Code and the Claude apps, High on the Claude Platform. If your integration pinned an effort level for Sonnet 5, confirm it still matches what you want — effort level affects both quality and the token count this whole post is about.
The honest caveat: none of the percentages above are guaranteed on your specific workflow. Token savings depend on how verbose the model used to be on your exact prompts, and the only way to know your number is to run your own before-and-after. If you're sitting on a Sonnet-tier automation and want someone to run that comparison for you instead of guessing, that's what a scoping call is for — and if you want the fuller context on how Sonnet 5 fits against Claude's other tiers first, our model guide is still accurate on the parts that haven't changed.
— Cole
Sources
- Anthropic — Claude Sonnet 5.5 announcement
- Anthropic — Claude Platform pricing docs (model pricing table, effort levels)