AI
“A Anthropic lançou o Claude Sonnet 5.5 em 28/09/2026 com o mesmo preço do Sonnet 5 (US$ 2/milhão de tokens de entrada e US$ 10 de saída), afirmando que ele é 30%+ mais rápido e custa até 30% menos por tarefa que o Sonnet 5”
Plain restatementAnthropic released Claude Sonnet 5.5 on 28 September 2026 at the same list API price as Claude Sonnet 5 (USD 2 per million input tokens, USD 10 per million output tokens), and Anthropic states the model generates output more than 30% faster and can reduce cost per completed task by up to 30% relative to Sonnet 5.
This post checks out. Anthropic did release Claude Sonnet 5.5 on 28 September 2026, and its own pricing documentation confirms $2 per million input tokens and $10 per million output tokens, the same list price as Sonnet 5. Anthropic's announcement page does state the model generates output more than 30% faster and can cost up to 30% less per completed task, and the post correctly presents those as Anthropic's claims rather than as independent test results. The post also correctly labels the benchmark table as the manufacturer's own and keeps the effort-level footnote on the Opus comparison that a lot of coverage dropped. One important piece of context the post does not mention: the independent evaluator Artificial Analysis measured Sonnet 5.5 at about $7.60 per task on its own test suite at maximum effort, roughly 50% more than Sonnet 5, because the model produces far more tokens per task. That does not contradict Anthropic's "up to 30% less" best-case claim, but it does mean the cost savings are not guaranteed and depend heavily on how the model is configured. The API migration details in the post, including the settings that now return errors, match Anthropic's official migration guide.
[drifted from the evidence:] A Anthropic [drifted from the evidence:] lançou o Claude Sonnet 5.5 [drifted from the evidence:] em 28/09/2026 com o mesmo preço do Sonnet 5 [drifted from the evidence:] (US$ 2/milhão de tokens [drifted from the evidence:] de entrada e US$ 10 [drifted from the evidence:] de saída), afirmando que ele é 30%+ [drifted from the evidence:] mais rápido e custa até 30% [drifted from the evidence:] menos por tarefa que o Sonnet 5
Anthropic [added by the neutral restatement:] released Claude Sonnet 5.5 [added by the neutral restatement:] on 28 September 2026 at the same list API price as Claude Sonnet 5 [added by the neutral restatement:] (USD 2 per million input tokens, [added by the neutral restatement:] USD 10 [added by the neutral restatement:] per million output tokens), and Anthropic states the model generates output more than 30% [added by the neutral restatement:] faster and can reduce cost per completed task by up to 30% [added by the neutral restatement:] relative to Sonnet 5.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- Anthropic released Claude Sonnet 5.5 on 28 September 2026. The date appears on Anthropic's own announcement page.
- The list price is $2 per million input tokens and $10 per million output tokens, confirmed on Anthropic's platform documentation, and it is identical to the price Anthropic set for Sonnet 5 in June 2026.
- The post's cache-read figure of $0.20 per million tokens matches Anthropic's published pricing table.
- Anthropic does claim 30%+ faster output generation and up to 30% lower cost per task versus Sonnet 5. The post attributes both figures to Anthropic rather than presenting them as independently measured.
- The benchmark figures the post cites match Anthropic's published table: Terminal-Bench 4.0 at 70.6% for Sonnet 5.5 against 10.3% for Sonnet 5 and 66.4% for Opus 5.5, and CursorBench 4.0 at 55.5% for Sonnet 5.5 against 34.1% for Sonnet 5 and 57.8% for Opus 5.5.
- The post preserves the effort-level footnote on the Opus 5.5 Terminal-Bench figure. Anthropic's page marks that 66.4% with a footnote, and press analysis confirms Anthropic says the Opus 5.5 result is reported at Xhigh effort which represents the model's highest score, while the Sonnet 5.5 figure carries no effort label, so this is not a like-for-like win.
- The post labels the benchmarks as the manufacturer's table rather than independent testing. That label is correct.
- The API migration details check out against Anthropic's migration guide. The documentation states that on Claude Sonnet 5.5, disabled returns a 400 invalid_request_error and lists accepted thinking types for Sonnet 5.5 as "adaptive" and "between_tools". The guide also states that the working request leaves out five settings that return a 400 error: thinking budgets, sampling parameters, assistant prefill, forced tool choice, and thinking: {"type": "disabled"}.
- The post's claim that between_tools works only at low, medium or high effort is supported by secondary documentation coverage, which states that the replacement, between_tools, only works up to high effort, and at xhigh or max it returns a 400.
- The post's note that migrants from Sonnet 4.6 or earlier should expect about 30% more tokens for the same text matches the migration guide, which states that Claude Sonnet 5.5 uses Claude Sonnet 5's tokenizer, and against Claude Sonnet 4.6, Claude Sonnet 4.5, and Claude Haiku 4.5, the same text produces about 30% more tokens, depending on the content. The guide's checklist for those models includes removing non-default temperature, top_p, and top_k values.
- The Lovable quote is reproduced accurately. Anthropic's page carries the line that its coding evals showed a third fewer tool calls and roughly half the shell runs to finish a task.
- The post's positioning summary matches Anthropic's own framing: where Opus 5.5 is built for complex work requiring careful judgment, Sonnet 5.5 is strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets.
- Omitted qualifier, minor and running against the post's own framing: on CursorBench 4.0 the post gives 55.5% for Sonnet 5.5 without noting the effort level. Coverage of Anthropic's per-effort data records CursorBench climbing with effort, reported as 39.2% at medium, 47.8% at high, 53.1% at xhigh and 55.5% at max, so the 55.5% is the model's top setting. Omitting this makes the gap to Opus 5.5 look narrower than a like-for-like comparison might show, which understates rather than overstates the post's own subject.
- The post does not mention that an independent evaluator published a contrary cost measurement the same day. Artificial Analysis measured higher, not lower, cost per task in its max-effort configuration. The post's attribution of the 30% figure to Anthropic is accurate, so this is an omission of available counter-evidence rather than a misstatement.
- Whether the "up to 30% less per task" outcome holds in ordinary use is unresolved. Anthropic's figure is a best-case vendor measurement under undisclosed task conditions. Artificial Analysis, running its own Intelligence Index, measured $7.60 per task at max effort for Sonnet 5.5, about 50% higher than Sonnet 5's cost per task on the same suite, alongside the highest output tokens per task it has recorded. These two results are not strictly contradictory, since one is a best-case claim across a task mix Anthropic has not published and the other is one evaluator's fixed suite at one effort setting, but they point in opposite directions and the question is open.
- The 30%+ speed figure has not been independently confirmed as a like-for-like measurement. Anthropic has not published the task set, effort setting, or measurement method behind it.
- Whether Terminal-Bench 4.0's 10.3% for Sonnet 5 reflects a genuine capability gap or a harness or scaffolding difference on a newer benchmark version is not established from the published material. A 60-point jump between consecutive models in one family is unusual, and Anthropic has not published the harness for either run.
- Early-tester figures quoted by Anthropic, including the Lovable tool-call and shell-run numbers, come from companies with a commercial relationship to the vendor and were run on their own internal evals with no published methodology.
Anthropic's own announcement page, dated September 28, 2026, introduces Claude Sonnet 5.5 as the second model in the Claude 5.5 family, a clear upgrade over Claude Sonnet 5 that runs 30%+ faster and costs up to 30% less for most work. The same page states that Sonnet 5.5 generates outputs 30%+ faster than Sonnet 5, making it Anthropic's fastest Sonnet model to date. The Claude Platform documentation lists the model's specifications directly: input pricing of $2 per MTok and output pricing of $10 per MTok, with a 1M token context window and 128K max output, and 5-minute cache write at $2.50/MTok, 1-hour cache write at $4/MTok, cache read at $0.20/MTok, and a 50% Batch API discount. The Sonnet 5 announcement from June 2026 confirms the baseline: Sonnet 5 was priced at $2 per million input tokens and $10 per million output tokens. The prices are therefore identical. On the cost-per-task mechanism, VentureBeat reports that Anthropic says Sonnet 5.5 generates output more than 30% faster than Claude Sonnet 5 and can reduce the total cost of completing a task by as much as 30%, primarily because it uses fewer tokens and fewer tool calls. An independent evaluator measured the cost direction differently in one configuration. Artificial Analysis reports that Anthropic priced Sonnet 5.5 identically to Sonnet 5 at $0.2/$2/$10 per 1M cache input/input/output tokens, however it outputs a higher number of Output Tokens per Task and costs $7.60 per task, around 50% higher than Sonnet 5's cost per task. With max effort, Sonnet 5.5 gains 18 points over Sonnet 5 and moves to #2 on the Intelligence Index, behind only Opus 5.5 at max effort. Artificial Analysis also states that this is the highest token use it has measured, around 60% higher than Opus 5.5 at max or Sonnet 5 at max, and that at this pricing Claude Sonnet 5.5 sits off the Intelligence vs. Cost per Task Pareto Frontier.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Compact share page: ai.trueseeker.com/s/7727efd7b0b8/yjAG60cP6XQLZoDOVPT0OviC8iN