Claude Sonnet 5.5 Review: Great Value Until You Max It

Claude Sonnet 5.5 arrived on 28 September at exactly Sonnet 5’s price, $2 in and $10 out per million tokens, with benchmark scores Anthropic put close to Opus 5.5. The independent numbers confirm most of that. On the Artificial Analysis Intelligence Index it scores 56 at maximum effort, second only to Opus 5.5’s 58, and ahead of Claude Fable 5.1 and GPT-6 Astra at 53. Sonnet 5 scored 38 on the same index, so the gain over the model it replaces is 18 points.

Our Opus 5.5 review found that the effort setting decides the verdict more than the model does. That is even more true of Sonnet 5.5, because here the answer changes sides: at some settings it is the best value Anthropic sells, and at others its own sibling beats it on both score and price.

Second on the index, and what each score costs

Artificial Analysis ran Sonnet 5.5 at all five effort settings and publishes what one task cost to complete. It also times the complete response, which it defines as input processing, reasoning and answer generation together. Here is every Sonnet 5.5 and Opus 5.5 setting, ordered by score, with the models they are usually compared against:

Model and effort	Intelligence	Cost per task	Full response	List price
Claude Opus 5.5, max	58	$5.98	685 s	$4 / $20
Claude Sonnet 5.5, max	56	$7.60	374 s	$2 / $10
Claude Opus 5.5, xhigh	56	$3.46	150 s	$4 / $20
Claude Opus 5.5, high	54	$1.82	62 s	$4 / $20
GPT-6 Astra, max	53	$3.26	314 s	$10 / $50
Claude Sonnet 5.5, xhigh	52	$2.74	58 s	$2 / $10
Claude Opus 5.5, medium	51	$1.34	28.5 s	$4 / $20
GPT-6 Sol, max	48	$1.05	186 s	$2 / $10
Claude Sonnet 5.5, high (API default)	47	$1.08	22.6 s	$2 / $10
Claude Opus 5.5, low	42	$0.55	18.5 s	$4 / $20
Claude Sonnet 5.5, medium (apps default)	41	$0.59	6.0 s	$2 / $10
Claude Sonnet 5, max	38	$5.09	190 s	$2 / $10
Claude Sonnet 5.5, low	36	$0.41	6.8 s	$2 / $10

Three readings come out of those numbers.

Up to High, it is excellent value. Medium, the default in Claude’s apps and Claude Code, scores 41 for $0.59 a task. That already beats Sonnet 5’s best score of 38, which cost $5.09, for about a ninth of the price. High, the default on the API, scores 47 for $1.08. That is a point behind GPT-6 Sol at its maximum at almost exactly the same cost, and Sonnet 5.5 has far more headroom above it.

Above High, the value belongs to Opus 5.5. Sonnet 5.5 at Xhigh scores 52 for $2.74, while Opus 5.5 at High scores 54 for $1.82. At Max it reaches 56 for $7.60, and Opus 5.5 reaches the same 56 at Xhigh for $3.46, less than half as much. The reason is token use. At Max, Sonnet 5.5 generated 410 million output tokens across the index, against a median of 88 million and 260 million for Opus 5.5 at its own maximum. A cheaper price per token buys nothing when the model spends more tokens on the same task.

At the bottom of the range, the two overlap. Opus 5.5 at Low scores 42 for $0.55, a point above Sonnet 5.5 at Medium for four cents less. On score and cost alone, Sonnet 5.5 is the cheapest route to a given score only at its Low and High settings. What settles the overlap is the column the index does not score.

Speed is the reason to choose it

Sonnet 5.5 is much faster. At Medium it returns a complete answer in about 6 seconds, with the first output after 1.2 seconds. Opus 5.5 takes 18.5 seconds at Low and 28.5 at Medium. At High the gap is 22.6 seconds against 62.2, and at Max it is about 6 minutes against 11. Output also streams faster at every setting: 85 to 139 tokens a second across Sonnet 5.5’s range, against 74 to 93 for Opus 5.5.

For anything a person waits on, such as a chat reply, a code suggestion or a support answer, that difference outweighs a few cents per task. It is the case for Sonnet 5.5 even at the points where Opus 5.5 is cheaper. Anthropic’s partners report the same pattern in their own tests, though Anthropic chose which to publish. Balyasny Asset Management measured about 121,000 tokens per answer against Sonnet 5’s 497,000 on 2,441 finance tasks, Zendesk says tickets were processed 20% faster, and Box reports 2.4 times the speed with 12% fewer tokens.

Anthropic’s benchmarks

The launch table is Anthropic’s own, run in its own harness. It puts Sonnet 5.5 a few points under Opus 5.5 almost everywhere, with one exception.

That exception is Terminal-Bench 4.0, where Sonnet 5.5 scores 70.6% against Opus 5.5’s best of 66.4%. Anthropic lists Sonnet 5 at 10.3% on the same test and does not explain that figure. On FrontierCode, Anthropic flags a fault of its own: at Max, Sonnet 5.5 more often ran Claude Code’s multi-agent review skill and changed things outside the task, and it scored 46.2% there against 52.1% at Xhigh. Leaving Terminal-Bench aside, the largest gains over Sonnet 5 are in computer use and chart reading, which rise by 23 and 46 points.

Professional work is where Sonnet 5.5 comes closest to Opus 5.5. On GDPval-AA, which covers real tasks from 44 occupations, it finishes two Elo points behind, and on AA-Briefcase eleven. Artificial Analysis ran both boards on a pre-release deployment with a structured-output bug that has since been fixed. Anthropic expects any effect to be small and to understate Sonnet 5.5.

What we found running it ourselves

We gave Sonnet 5.5 the three build briefs that Opus 5.5, GPT-6 Astra and GPT-6 Sol had already taken: a motorway interchange and a seaside fairground as animated SVGs, and a cinematic rocket launch as one web page, one attempt each at maximum effort with no way to look at the result. It finished second in all three rounds and took 12 points, to 13 for Opus 5.5, 11 for Astra and 5 for Sol.

Its interchange is only the second in the series with nothing wrong in it. It is a cloverleaf whose looping cars change level halfway round the loop without ever being drawn on the wrong one, every car keeps to the correct side, and its ramps open properly out of the road, which Opus 5.5’s do not. Its fairground matches Opus 5.5’s in almost everything: a wheel that stays level, with lights running up every spoke, and a carousel that turns as one assembly. The difference is that its horses turn round in a single frame. Its launch flies the whole mission to a correct orbit and then deploys its satellite, though film grain and a flat band of ground showing above the curved horizon spoil the picture.

It was slow. It thought for 32 to 48 minutes on each brief and filled Claude Code’s reasoning limit every time, which is the verbosity at maximum effort that the independent numbers show. At half Opus 5.5’s price per token it still came within a point of it, the strongest showing a Sonnet has made in our tests.

What changes for developers

The price list is identical to Sonnet 5’s: $2.50 per million tokens for five-minute cache writes, $4 for one-hour writes, $0.20 for cache reads at launch, halved to $0.10 on 7 October, and half price on the Batch API. The tokenizer is also unchanged. The API behaviour is not. Anthropic lists five breaking changes:

• Thinking can be reduced but not disabled. thinking: {"type": "disabled"} returns a 400 error. The floor is a new between_tools mode, which skips up-front thinking and is accepted only up to High effort.
• Forced tool use is gone, as it is on Opus 5.5. Use auto with strict tool use.
• Thinking blocks are tied to the model and the conversation. On accounts created on or after 31 August 2026, replaying one after editing an earlier turn is an error, so keep histories append-only.
• Computer use needs the new toolset on the Claude API and Google Cloud.
• The advisor tool no longer pairs a Sonnet 5.5 executor with Opus 4.8, Opus 4.7 or Sonnet 5.

One further change can make an app go quiet: text written between tool calls now arrives inside thinking blocks. Sonnet 5.5 is also the first Sonnet with the cyber safeguards Anthropic uses on its top models, so higher-risk security tasks visibly fall back to Sonnet 5. For ordinary development none of this shows. For security research it does.

Who should be running it

If you are…	Do this
On Claude Sonnet 5	Move. Same list price, 18 more index points at max, and at medium it beats Sonnet 5’s best for about a ninth of the cost per task. Budget time for the breaking API changes
Building something people wait on	Use it at medium. It returns a complete answer in about 6 seconds, three to five times faster than Opus 5.5 at low or medium
Running long agentic jobs	Use high, the API default. If you need more, switch to Opus 5.5 rather than raising the effort: Opus 5.5 at high scores 54 for $1.82, and at xhigh matches Sonnet 5.5’s max score for less than half the cost
On Claude Opus 5.5	Keep it for anything above high effort. Try Sonnet 5.5 where response time matters more than the last few index points
On GPT-6 Sol	At high, Sonnet 5.5 is a point behind Sol’s best score at the same cost per task, and answers in 22.6 seconds against 186. Above that, it has headroom Sol does not
Running cheap bulk jobs	If a score in the high 30s is enough, GPT-6 Luna at max costs $0.07 a task against $0.41 for Sonnet 5.5 at low

Verdict

Sonnet 5.5 earns 4.5 out of 5. It is the second-best model on the independent index at the same list price as the model it replaces. It is the fastest way to a strong answer from Claude, and from Low to High effort it is a genuine bargain.

It stops short of Opus 5.5’s 4.9 for one reason, and that reason is Anthropic’s own flagship. Above High, Opus 5.5 reaches every score Sonnet 5.5 can for less money, and at Max the gap is more than double. The breaking API changes are a smaller cost, the same kind Opus 5.5 brought, and worth a day of migration work before switching.

The practical rule fits in one line: Medium for anything a person is waiting on, High for agentic work, and when that is not enough, Opus 5.5 rather than Max. The leaderboard now carries it.