Claude Sonnet 5.5 chega ao preço de 2/10 USD do Sonnet 5, mas quem desativa o thinking tem de passar antes para between_tools
- O que aconteceu
- A Anthropic lançou o Claude Sonnet 5.5 (claude-sonnet-5-5) a 28 de setembro ao preço do Sonnet 5, 2 USD/10 USD por 1 milhão de tokens, e é agora o Sonnet predefinido na API da Anthropic.
- Porque é importante
- Na tabela da própria Anthropic, aproxima-se do Opus 5.5 a metade do preço, mas as integrações com o thinking desativado têm de passar para between_tools antes de migrar.
- O que fazer
- Passe para between_tools as integrações com o thinking desativado e depois execute as suas próprias avaliações face ao Sonnet 5 e ao Opus 5.5 antes de mudar a produção.
Our verdict: test Claude Sonnet 5.5 on anything you run on Claude Sonnet 5 today, but Sonnet 5 stays recommended and Sonnet 5.5 stays conditional until someone other than Anthropic measures it. Input and output prices are unchanged, and Anthropic's table puts Sonnet 5.5 ahead of Sonnet 5 on every row and level with Claude Opus 5.5 at half the price. We have not run Sonnet 5.5 ourselves. If you run Sonnet with thinking off, you have one step to take before you switch.
What happened
Anthropic released Claude Sonnet 5.5 on 28 September 2026 under the API id claude-sonnet-5-5. Claude Code 2.1.284, released the same day, added it as "now the default Sonnet model on the Anthropic API", with a 1M-token context.
| Per 1M tokens | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| Input | $2 | $4 |
| Output | $10 | $20 |
| Cache reads (current) | $0.10 | $0.20 |
| Cache writes | $2.50 | $5 |
- Price: $2 input and $10 output per 1M tokens, the same as Sonnet 5.
- Cache reads: $0.20 per 1M tokens at launch, per the Claude Code 2.1.284 release notes. Anthropic's page now lists $0.10.
- Cost per task: Anthropic says the model "costs up to 30% less per task than its predecessor" because it needs fewer tokens.
- Speed: Anthropic says it "generates outputs 30%+ faster than Sonnet 5".
- Availability: Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure.
- Default effort: Medium in Claude Code and the Claude apps, High on the Claude Platform.
The cost and speed figures are Anthropic's own measurements.
Why it matters
Anthropic's benchmark table:
| Benchmark | Sonnet 5.5 | Sonnet 5 | Opus 5.5 |
|---|---|---|---|
| Terminal-Bench 4.0 | 70.6% | 10.3% | 66.4% |
| CursorBench 4.0 | 55.5% | 34.1% | 57.8% |
| GDPval-AA v2.1 | 1844 | 1449 | 1846 |
| OSWorld 2.1 | 80.1% (partial) | 57.0% | 81.8% |
On Anthropic's numbers, Sonnet 5.5 comes within about two points of Opus 5.5 at half the input and output price, and beats it on Terminal-Bench 4.0. That makes the Opus comparison, not the Sonnet 5 one, where the money is.
Treat the gap with care. Sonnet 5's Terminal-Bench score of 10.3% is far below anything else in the table, and Anthropic's post does not explain it. Until an independent run reproduces that 60-point gap, it is Anthropic's result, not a fact.
Two safeguards change behaviour:
- Cyber fallback: "higher-risk cybersecurity tasks will visibly fall back to Sonnet 5". If you use Sonnet for security testing, expect some requests to be answered by the older model.
- Preserved thinking: Anthropic says "Claude's thinking cannot be decoupled from the account that created it".
What changes for you
The migration step. In Anthropic's words: "If you run Sonnet with thinking off, you'll need to switch to the new between_tools setting" before moving to Sonnet 5.5. Change that setting first and the model id second. It is the same pattern as Claude Opus 5.5, where our directory notes thinking cannot be turned off.
Our directory entry lists five breaking changes for code written for Sonnet 5. Hold off if you rely on forced tool choice or temperature control.
Your default may already have moved. After updating Claude Code to 2.1.284 or later, check which Sonnet your sessions and pinned model settings actually resolve to.
Where our directory stands. claude-sonnet-5 stays recommended: nothing we read retires it, and the new model's lead rests only on Anthropic's own tests. claude-sonnet-5-5 is conditional for the same reason. For new Sonnet workloads, start your testing on Sonnet 5.5. It costs nothing more per token, and Anthropic's per-task claim, if it holds, makes it cheaper.
O que fazer
- 1 Se alguma integração executa o Sonnet com o thinking desativado, passe-a para a definição between_tools antes de mudar o ID do modelo para claude-sonnet-5-5.
- 2 Execute as suas próprias avaliações do claude-sonnet-5-5 face ao claude-sonnet-5 e ao Opus 5.5. A 2 USD/10 USD contra 4 USD/20 USD, é na comparação com o Opus que estão as poupanças.
- 3 Depois de atualizar o Claude Code para a 2.1.284 ou posterior, verifique que versão do Sonnet estão realmente a usar as suas sessões e as definições de modelo fixadas.
- 4 Para equipas de segurança: é de esperar que alguns prompts de cibersegurança de maior risco sejam respondidos pelo Sonnet 5 e não pelo Sonnet 5.5.
Ferramentas e modelos afetados
Nunca mais precisas de te pôr a par
O resumo semanal: apenas mudanças de veredicto e ações urgentes. Sem enchimento.