Anthropic unveils Sonnet 5.5, a model that comes close to Opus

Anthropic announced Claude Sonnet 5.5 in a post from the Claude account on X (@claudeai): “Introducing Claude Sonnet 5.5, the second model in the Claude 5.5 family. It’s a clear upgrade over Sonnet 5, runs more than 30% faster, and costs up to 30% less for most work.”

Sonnet 5.5 beat Opus 5.5 on Terminal-Bench 4.0 with 70.6% and stays close to it on every other test. On the other two coding tests, FrontierCode and CursorBench, Opus 5.5 still leads.

Thinking harder doesn’t always help. On FrontierCode, Sonnet 5.5 scored lower at Max, its highest effort setting, than at the next setting down. Anthropic says that at Max it more often splits code review across subagents; in two cases examined, that led to a timeout or to edits beyond the task’s scope.

FrontierCode 1.1 score

• Opus 5.5: 54.4%
• Sonnet 5.5, Xhigh effort: 52.1%
• Sonnet 5.5, Max effort (highest): 46.2%
• Sonnet 5: 42.4%

Claude now holds 3 of the top 5 spots on the Artificial Analysis Intelligence Index. At 56, Sonnet 5.5 even scores higher than Fable 5.1, behind only Opus 5.5 at 58.

On knowledge, it’s a trade-off. In Artificial Analysis’s AA-Omniscience test, Sonnet 5.5 answered fewer questions correctly than Opus 5.5. But on the questions it couldn’t answer correctly, it made up a wrong answer less often.

AA-Omniscience, max effort

• Accuracy, Opus 5.5: 66%
• Accuracy, Sonnet 5.5: 54%
• Hallucination rate, Opus 5.5: 59%
• Hallucination rate, Sonnet 5.5: 47% (lower is better)

Anthropic says that on several benchmarks, Sonnet 5.5 at Low or Medium effort beats Sonnet 5’s best score for about a tenth of the cost per task. In Anthropic’s chart for AA-Briefcase, a knowledge-work test, Sonnet 5.5 at Medium already scores higher than Sonnet 5’s best.

But at the top setting, the savings flip. Artificial Analysis says Sonnet 5.5 at Max used more tokens per task than any model it has measured, and cost about 50% more per task than Sonnet 5.

Output tokens per task

• Sonnet 5.5, Max: 193K
• Opus 5.5, Max: 119K
• Sonnet 5, Max: 118K
• GPT-6 Astra, Max: 27K
• Sonnet 5.5, Medium: 19K

Sonnet 5.5 is the first Sonnet to ship with the cyber safeguards used on Opus, and Anthropic says to expect more refusals even on benign security work. It’s also the first Sonnet that blocks attempts to extract its hidden reasoning to train other models.

New safeguards on Sonnet 5.5

• Higher-risk cybersecurity request → Sonnet 5 answers instead
• Benign security work → More refusals expected
• Attempt to extract hidden reasoning → Blocked, no fallback model

Sonnet 5.5 is the first Sonnet to beat Pokémon Red working only from screenshots. And Anthropic says Haiku 5.5, built for high-volume, cost-sensitive work, joins the Claude 5.5 family in the coming weeks.

Two more things

• Pokémon Red → Beaten from screenshots alone
• Claude Haiku 5.5 → Coming in the next few weeks

X user @full_kelly_ replied to Claude’s Sonnet 5.5 announcement: “we should slow down AI / anyway, here’s Opus 5.5 / update: here’s Sonnet 5.5”