Not established from the available sources.
ANTHROPIC · MODEL RELEASE TRACKER
Claude Sonnet 4.5
Claude Sonnet 4.5 is an Anthropic model version referenced in reporting about a second wave of text-watermark retrofits; Anthropic told API customers Sonnet 4.5 will receive watermarking in the November 16, 2026 rollout (Bedrock rollout may take a few extra days).CURRENT SNAPSHOT1/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Not established from the available sources.
Not established from the available sources.
Pass@1 improvement with DeReAct (GAIA and SWE-bench Verified)DeReAct improved Pass@1 by 4.2–5.2 points for Claude Sonnet 4.5 on GAIA and SWE-bench Verified.
HumanEval+ held-out score (coding benchmark)86.6% on the held-out HumanEval+ instances (as reported in the cited paper).
Benchmark inclusionClaude Sonnet 4.5 was included among the models evaluated in the Ai2 "Old Ideas, Novel Problems" benchmark for novelty judges.
Not established from the available sources.
VERIFIABLE FACTS
Every value stays attached to a source and date.
BENCHMARKS · Pass@1 improvement with DeReAct (GAIA and SWE-bench Verified)INDEPENDENTLY SUPPORTED
DeReAct improved Pass@1 by 4.2–5.2 points for Claude Sonnet 4.5 on GAIA and SWE-bench Verified.
BENCHMARKS · HumanEval+ held-out score (coding benchmark)INDEPENDENTLY SUPPORTED
86.6% on the held-out HumanEval+ instances (as reported in the cited paper).
CAPABILITIES · Benchmarks evaluated (implying coding capability)INDEPENDENTLY SUPPORTED
Reportedly evaluated on coding benchmarks HumanEval+ and Polyglot (paper reports held-out scoring on those splits).
BENCHMARKS · Benchmark inclusionDEVELOPER CLAIM
Claude Sonnet 4.5 was included among the models evaluated in the Ai2 "Old Ideas, Novel Problems" benchmark for novelty judges.
CAPABILITIES · Capability — novelty evaluationDEVELOPER CLAIM
Claude Sonnet 4.5 was used as a novelty judge to assess whether research ideas are novel relative to prior work.
LIMITATIONS · Limitation — reliability and prompt sensitivityDEVELOPER CLAIM
The study reports that novelty evaluation across the evaluated LLM judges, including Claude Sonnet 4.5, was far from reliable: small changes to prompts produced dramatically different judgments (in one experiment, changing a few words shifted a model's accuracy by more than 50 percentage points), and retrieval or additional reasoning did not consistently improve judgments.
SAFETY · Text watermark rollout dateDEVELOPER CLAIM
Anthropic emailed API customers that Claude Sonnet 4.5 will receive a text watermark starting November 16, 2026 (the rollout on Amazon Bedrock may take a few extra days).
LIMITATIONS · Watermark sampling restrictionDEVELOPER CLAIM
The watermark on Claude Sonnet 4.5 applies only when sampling parameters temperature, top_p, and top_k are left at their default values; changing any one of these parameters can avoid the watermark, per Anthropic's email to customers.
SAFETY · Anthropic's stated properties of the watermarkDEVELOPER CLAIM
Anthropic describes the watermark as an "imperceptible pattern in word choice" that it says does not change meaning, quality, or readability; adds no extra characters or tokens; does not affect pricing or request/response formats; barely slows output speed; and "contains no information about the user, their organization, or their conversations with Claude."
WHAT CHANGED
Stored passport versions, without reconstructed history.
SAFETY · Anthropic's stated properties of the watermark+ Anthropic describes the watermark as an "imperceptible pattern in word choice" that it says does not change meaning, quality, or readability; adds no extra characters or tokens; does not affect pricing or request/response formats; barely slows output speed; and "contains no information about the user, their organization, or their conversations with Claude."
SAFETY · Text watermark rollout date+ Anthropic emailed API customers that Claude Sonnet 4.5 will receive a text watermark starting November 16, 2026 (the rollout on Amazon Bedrock may take a few extra days).
LIMITATIONS · Watermark sampling restriction+ The watermark on Claude Sonnet 4.5 applies only when sampling parameters temperature, top_p, and top_k are left at their default values; changing any one of these parameters can avoid the watermark, per Anthropic's email to customers.
BENCHMARKS · Benchmark inclusion+ Claude Sonnet 4.5 was included among the models evaluated in the Ai2 "Old Ideas, Novel Problems" benchmark for novelty judges.
LIMITATIONS · Limitation — reliability and prompt sensitivity+ The study reports that novelty evaluation across the evaluated LLM judges, including Claude Sonnet 4.5, was far from reliable: small changes to prompts produced dramatically different judgments (in one experiment, changing a few words shifted a model's accuracy by more than 50 percentage points), and retrieval or additional reasoning did not consistently improve judgments.
CAPABILITIES · Capability — novelty evaluation+ Claude Sonnet 4.5 was used as a novelty judge to assess whether research ideas are novel relative to prior work.
CAPABILITIES · Benchmarks evaluated (implying coding capability)+ Reportedly evaluated on coding benchmarks HumanEval+ and Polyglot (paper reports held-out scoring on those splits).
BENCHMARKS · HumanEval+ held-out score (coding benchmark)+ 86.6% on the held-out HumanEval+ instances (as reported in the cited paper).