Claude Sonnet 5.5 Ships Faster and Breaks Five Things

On September 28, Anthropic released Claude Sonnet 5.5, the current Sonnet model on the Claude API. It's priced exactly like the model it replaces, but it doesn't behave exactly like it: five things that worked on Sonnet 5 either error out or change shape on the new model, and a sixth goes quiet without telling you. I read Anthropic's own announcement and the migration guide so I could work out which ones actually matter.

What Changed in Claude Sonnet 5.5?

The model ID is claude-sonnet-5-5, and it went live the same day across the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. That's the same five-platform spread Opus 5.5 launched with on September 22, and it means nobody building on a specific cloud has to wait for a second wave this time.

The spec sheet is mostly unchanged from Sonnet 5. Context window stays at 1M tokens, max output stays at 128K, and the Batch API still supports up to 300K output tokens with a beta header.

Pricing didn't move either, which is the detail that surprised me most given Opus 5.5 got a straight price cut a week earlier. Sonnet 5.5 costs the same per token as Sonnet 5, which works out to half of what Opus 5.5 charges and a fifth of Fable 5.1's rate on both input and output. It's still the cheapest model in the current lineup that isn't Haiku.

What did move is the reliable knowledge cutoff, from January 2026 on Sonnet 5 to June 2026 on Sonnet 5.5, matching the cutoff Opus 5.5 and Fable 5.1 already carry. Five months is a real gap to close, and it means Sonnet 5.5 has a materially different picture of anything that happened in the first half of this year than the model it's replacing. Default effort stayed put at high on both versions, so that's one lever you don't have to re-check if you're migrating.

Retirement got the standard runway too: Anthropic committed to keeping Sonnet 5.5 active for at least a year, not sooner than September 28, 2027.

Sonnet 5 isn't going anywhere immediately either. It's marked legacy rather than deprecated, so existing calls against claude-sonnet-5 keep working while you migrate on your own schedule. Nobody's forcing your hand this week.

Which Five Things Break When You Upgrade?

Anthropic's model page lists five breaking changes against Sonnet 5, plus a sixth that changes behavior without failing the request outright. This is the part worth reading closely if anything you run already calls Sonnet on the API.

Five and one. Not four. That's a wider blast radius than Opus 5.5 shipped with a week earlier.

The first: disabling thinking with "type": "disabled" no longer works. To turn off up-front thinking on Sonnet 5.5, you send thinking: {"type": "between_tools"} instead, and that only works at high effort or below. If your code sends the old disabled flag, expect it to be rejected rather than silently ignored.

The second is forced tool use. tool_choice set to any or to a named tool used to guarantee the model called something on that turn. On Sonnet 5.5, both return a 400 error. Only auto and none still work, so anything relying on a forced call for structured output needs a different pattern now.

Third, thinking blocks are tied to the specific model and conversation that produced them. Passing one into a resumed conversation or a different model isn't something to lean on anymore, and this is the kind of thing that passes every test you write and then breaks quietly in production months later.

Fourth: the older computer_20251124 computer use tool stops being accepted on the Claude API and Google Cloud with this model.

Fifth, and new compared to what I saw on Opus 5.5's release: the advisor tool now rejects Claude Opus 4.8, Opus 4.7, and Sonnet 5 as advisors outright. If a multi-agent setup pairs Sonnet 5.5 with one of those as an advisor model, that pairing stops working. I hadn't seen an advisor-specific restriction on a release before this one.

The sixth change doesn't error at all. Text that used to stream between tool calls now arrives inside thinking blocks, and at the default display setting that text doesn't reach the client. Any UI that showed that text as a live progress indicator goes quiet between tool calls until you set a display value that returns it, or switch to between_tools and lose the up-front thinking instead.

Is Sonnet 5.5 Actually Faster, or Just Cheaper?

Speed is the headline Anthropic is leading with, and I'd normally discount a vendor's own framing, but this one's specific enough to check: Sonnet 5.5 generates output more than 30 percent faster than Sonnet 5, and Anthropic says real workloads run up to 30 percent cheaper because the model needs fewer tokens to reach the same answer, not because the per-token price changed.

On the Vals Index, an independent benchmark tracker, Sonnet 5.5 placed second of 66 models tested, sitting well behind Opus 5.5 but at roughly 63 percent of the cost per test. That's the trade the whole release is making: close most of the intelligence gap to the flagship while staying priced like the budget option.

CursorBench 4.0, a coding-agent benchmark, backs that up. Sonnet 5.5 scored 55.5 percent there, a sizeable jump from Sonnet 5's 34.1 percent, and close enough to Opus 5.5's own score that picking Opus for a coding agent now needs a better reason than "it's the bigger model." A 21-point jump on one benchmark in one release is the kind of number that usually gets rounded down by real-world use, so I'm treating it as a ceiling rather than a promise until I've run enough of my own tasks through it to trust the gap.

None of this makes Sonnet 5.5 the best model Anthropic ships. Fable 5.1 and Opus 5.5 both still lead on raw capability.

What it does is move the "good enough for most of my agent loop" line further down the price ladder than it sat a week ago. That's a more useful shift for a one-person studio than a new state of the art nobody outside a lab can afford to run constantly. Cheaper and nearly as good beats better and unaffordable most days of the week.

Should You Upgrade Today? What I'd Check First

If you're running Sonnet 5 today, I wouldn't call this a drop-in swap, even though the price tag says otherwise. Grep your codebase for three things first: any request that sends thinking: {"type": "disabled"}, any tool_choice set to any or a named tool, and any code that streams the text between tool calls to a user interface. Those three cover four of the six changes and they're the ones most likely to fail silently instead of loudly.

The tool_choice one is the easiest to miss because it only shows up when the model actually needs to make that exact call, which might not happen on every test run. If your test suite doesn't specifically exercise the forced-tool-use path, a passing CI run tells you nothing about whether that code still works. I'd add a test that deliberately hits it before flipping the model ID in anything that touches production traffic.

The advisor tool restriction only matters if you're running a multi-agent setup that pairs models as advisors, which most single-agent projects aren't doing, so I'd check that one last, and skip it entirely if you don't have one.

Thinking blocks tied to model and conversation mostly bite long-running agents that persist state across sessions. If your agent starts fresh each run, you probably won't notice it either. What I'd actually wait on: anything that pins a specific computer use tool version in production, since computer_20251124 support ends outright on this model rather than just being discouraged, and there's no grace window mentioned anywhere in the migration guide.

I'd move the knowledge cutoff jump higher on my own list than the price tag suggests you should. A five-month gap means Sonnet 5.5 knows about things Sonnet 5 flatly doesn't, and if your prompts lean on the model's own world knowledge rather than retrieval, that's the change most likely to shift output quality in ways a migration checklist won't catch. Test it on your actual prompts today, not just the API surface, before you decide whether to wait.

I wrote about Opus 5.5's own breaking changes and price cut the week before this one in Claude Opus 5.5 Ships Cheaper and With Four Breaking Changes, and about the last Sonnet-adjacent release in Claude Fable 5.1 and Mythos 5.1: What Changed on September 1. The pattern across three releases in a month is the same: prices hold or drop, and the breaking changes keep clustering around the same handful of API surfaces.

Bottom Line

Claude Sonnet 5.5 is a genuine upgrade at the same price, not a routine bump dressed up as one. The speed claim checks out against independent benchmarks, the cutoff jump is real, and the price held flat while Opus 5.5 got cheaper the week before. None of that makes it a safe blind swap.

Five breaking changes and one silent behavior shift are enough to fail a request or go quiet in production if you don't check for them first. My read: grep for the three patterns above before you touch the model ID, budget an afternoon for the migration guide either way, and don't skip testing on your own prompts just because the API surface looks compatible. The five-platform simultaneous launch means there's no reason to wait on a specific cloud catching up. The reason to wait, if you have one, is your own code, not Anthropic's rollout.