Language modelActive

Claude Sonnet 4.5

Anthropic

Released: September 2025
Data date: September 8, 2026
Successor: Claude Sonnet 4.6

Speed: 43.8 tokens/s
Output throughput
Price: $3 · $15
Input · Output / 1M tokens
Input: Text, Image
Output: Text

Claude Sonnet 4.5 stood out for computer use. Anthropic highlighted its 61.4% result on OSWorld-Verified in the launch announcement. As the predecessor to Claude Sonnet 4.6, it targeted complex coding and agent tasks.

Specifications and access

Specification	Value and source
Weights	ProprietarySource
Access	APISource
Context window	200,000 tokensSource
Knowledge cutoff	January 2025Source
Notes	Anthropic does not disclose parameter countsSource
API model ID	claude-sonnet-4-5-20250929Source
Source status	active (legacy)Source
Release	2025-09-29Source
Maximum output	64,000 tokensSource
Input	Text, imagesSource
Output	TextSource
Reasoning mode	extended thinkingSource
Knowledge cutoff	2025-01 reliable / 2025-07 trainingSource
Tool calling	YesSource
Structured output	YesSource

Published benchmarks

Agent tasks

BFCL v4

Overall

Berkeley's function-calling test combines single and multi-turn tool calls with web search, memory, and hallucination measurement. The overall score is an unweighted mix of subcategories and depends on native function calling or prompt workarounds.

1.Claude Opus 4.577.47%

2.Claude Sonnet 4.573.24%

higher is betterAs of September 2, 2026Scale: 0% - 100%

Source: Berkeley BFCL

CC BY 4.0

gradually.ai

Methodology and context

Comparison cohort: bfcl-v4-f7cf735-overall
Metric: unweighted overall accuracy
Scale: 0-100
Participants: 2
Methodology and context: Unweighted overall score across agentic, multi-turn, and single-turn function-calling tests.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Model snapshot: claude-sonnet-4-5-20250929-fc. Evaluator: Berkeley BFCL. Dataset type: public. Leaderboard date: 2026-04-12. Source ID: berkeley-bfcl. Evidence capture ID: curated-berkeley-bfcl-2026-09-02-171020422eb3. Source payload hash: 171020422eb31aa497dbf9dd95b72c153c0db68a3eceec63a4c059bade96f4c2. Parser version: curated-v1. Imported at: 2026-09-29

Source: Berkeley BFCL

Agent Poker Bench

Overall

1.GPT-51,103.18 rating

2.Gemini 3 Flash Preview1,100.21 rating

3.DeepSeek-V3.11,058.23 rating

4.Claude Sonnet 4.51,055.5 rating

5.Claude Opus 4.51,033.38 rating

6.Gemini 2.5 Pro1,032.6 rating

7.GPT OSS 120B1,015.33 rating

8.Kimi K2 Thinking1,011.63 rating

9.GLM-4.6945.76 rating

higher is betterAs of September 29, 2026Scale: 0 rating - 1,103.18 rating

Source: Vals AI

CC BY 4.0

gradually.ai

Methodology and context

Comparison cohort: vals-poker-agent:1:overall
Metric: TrueSkill rating
Scale: No upper bound is documented. The chart spans the recorded values and zero.
Participants: 9
Methodology and context: Source model ID: anthropic/claude-sonnet-4-5-20250929. Source model name: Claude Sonnet 4.5. Model snapshot: claude-sonnet-4-5-20250929. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2025-12-23. Latency in seconds: 16.67560795007542. Standard error: 0. Cost per task in USD: 0.027591309991158296. Source ID: vals-ai. Evidence capture ID: vals-poker_agent-2026-09-29-f29b75137483. Source payload hash: f29b751374830275dd1fba4c8a4f029b69433858ae99420aa90befe3093d0096. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source: Vals AI

Coding

CORE-Bench Hard

Automatic validation

45 public tasks for reproducing research results with code, data, and dependencies. Automatic and manually validated scores are different metrics. The agent scaffold strongly affects the score.

1.Claude Opus 4.577.78%

2.Claude Sonnet 4.562.22%

higher is betterAs of August 6, 2026Scale: 0% - 100%

Source: Princeton HAL

CC BY 4.0

gradually.ai

Methodology and context

Comparison cohort: hal-corebench-hard-claude-code-automatic
Metric: automatic accuracy
Scale: 0-100
Participants: 2
Methodology and context: Claude Code on 45 public paper reproductions. Total cost is divided by 45.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Evaluator: Princeton HAL. Dataset type: public. Harness: Claude Code. Agent: Claude Code. Cost per task in USD: 1.5184444444444445. Source ID: princeton-hal. Evidence capture ID: curated-princeton-hal-2026-08-06-d289e35e619a. Source payload hash: d289e35e619aac7b28355c86830d55c245fc6b287e2327069860ac0522c95e8b. Parser version: curated-v1. Imported at: 2026-09-29

Source: Princeton HAL

SciCode

Accuracy

65 scientific coding problems with 338 subproblems across 16 research fields. Zero Shot, Tool Calling, and HAL Generalist are not directly comparable scaffolds.

1.o39.23%

2.GPT-56.15%

3.Claude Sonnet 4.53.08%

4.GPT-4.11.54%

higher is betterAs of August 5, 2026Scale: 0% - 100%

Source: Princeton HAL

CC BY 4.0

gradually.ai

Methodology and context

Comparison cohort: hal-scicode-scicode-tool-calling-overall
Metric: accuracy
Scale: 0-100
Participants: 4
Methodology and context: SciCode Tool Calling on 65 main scientific problems. Total cost is divided by 65.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Evaluator: Princeton HAL. Dataset type: public. Harness: SciCode Tool Calling. Agent: SciCode Tool Calling. Cost per task in USD: 1.6044615384615386. Source ID: princeton-hal. Evidence capture ID: curated-princeton-hal-2026-08-05-c67a4c4ed842. Source payload hash: c67a4c4ed8428dfb71febdf30c9d21b46b0644189ef4d4de5ce334f3f2edc12c. Parser version: curated-v1. Imported at: 2026-09-29

Source: Princeton HAL

Cybersecurity

MT-AgentRisk

Attack success rate

365 harmful sequences test safety risks in multi-turn tool use. A lower attack success rate is better. Capability and safety are separate dimensions.

Note (Claude Sonnet 4.5): Table 4 reports a 15.43% refusal rate, while Table 9 in the same paper reports 15.34%. This row uses the direct single-turn versus multi-turn comparison table.

1.Claude Sonnet 4.571.78%

2.Gemini 3 Flash Preview78.9%

3.DeepSeek-V3.284.93%

lower is betterAs of August 9, 2026Scale: 0% - 100%

Source: MT-AgentRisk

CC BY 4.0

gradually.ai

Methodology and context

Comparison cohort: mt-agent-risk:multi-turn:asr
Metric: attack success rate
Scale: 0-100
Participants: 3
Methodology and context: OpenHands without a defense layer, judged by GPT-4.1. Single-turn and multi-turn cohorts remain separate.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Evaluator: MT-AgentRisk. Dataset type: public. Leaderboard date: 2026-02-13. Harness: OpenHands. Source ID: mt-agent-risk. Evidence capture ID: curated-mt-agent-risk-2026-08-09-f550ff9bf634. Source payload hash: f550ff9bf6346dc509fb04e302b64acda7943dd7e5863a5ca3206286480a11a5. Parser version: curated-v1. Imported at: 2026-09-29

Source: MT-AgentRisk

MT-AgentRisk

Refusal rate

365 harmful sequences test safety risks in multi-turn tool use. A lower attack success rate is better. Capability and safety are separate dimensions.

Note (Claude Sonnet 4.5): Table 4 reports a 15.43% refusal rate, while Table 9 in the same paper reports 15.34%. This row uses the direct single-turn versus multi-turn comparison table.

1.Claude Sonnet 4.515.43%

2.Gemini 3 Flash Preview6.03%

3.DeepSeek-V3.21.1%

higher is betterAs of August 9, 2026Scale: 0% - 100%

Source: MT-AgentRisk

CC BY 4.0

gradually.ai

Methodology and context

Comparison cohort: mt-agent-risk:multi-turn:refusal
Metric: refusal rate
Scale: 0-100
Participants: 3
Methodology and context: OpenHands without a defense layer, judged by GPT-4.1. Single-turn and multi-turn cohorts remain separate.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Evaluator: MT-AgentRisk. Dataset type: public. Leaderboard date: 2026-02-13. Harness: OpenHands. Source ID: mt-agent-risk. Evidence capture ID: curated-mt-agent-risk-2026-08-09-9e7ba7aa3157. Source payload hash: 9e7ba7aa3157bc30470265facc4bb25edf1a9b2f6cf7b733102470330ed4b034. Parser version: curated-v1. Imported at: 2026-09-29

Source: MT-AgentRisk

MultiChallenge

Overall

Multi-turn conversations test memory, instruction retention, versioned editing, and self-coherence. Tasks were selected from shared failures of six older frontier models and may disadvantage those systems.

1.Gemini 3.1 Pro Preview71.37%

2.GPT-5.4 Pro69.23%

3.GPT-563.19%

4.o3-pro62.4%

5.Claude Opus 4.558.97%

6.o356.62%

7.Claude Opus 4.656.02%

8.Claude Sonnet 4.555.32%

higher is betterAs of August 26, 2026Scale: 0% - 100%

Source: Scale AI

CC BY 4.0

gradually.ai

Methodology and context

Comparison cohort: scale-multichallenge-overall
Metric: rubric pass rate
Scale: 0-100
Participants: 8
Methodology and context: Multi-turn conversations with binary task-specific rubrics.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Evaluator: Scale AI. Dataset type: public. Attempts: pass@1. Reasoning effort: thinking. Score display: 55.32% ± 1.47. Source ID: scale-ai. Evidence capture ID: curated-scale-ai-2026-08-26-93c4c6df84ba. Source payload hash: 93c4c6df84baa57ecc302765b454f1d1d402750d6ee8968f2085772a44e192b1. Parser version: curated-v1. Imported at: 2026-09-29

Source: Scale AI

Finance

PRBench Finance

Overall

600 realistic finance questions with weighted criteria, including a 300-question hard subset. An o4-mini judge scores the responses, so domain correctness and operational reliability remain separate questions.

1.Claude Fable 553.86%

2.Claude Opus 4.653.28%

3.GPT-551.32%

4.GPT-5 Pro51.06%

5.GPT-5.6 Sol50.45%

6.o3-pro49.08%

7.o347.69%

8.Claude Opus 4.546.16%

9.GPT-5.445.63%

10.Claude Sonnet 4.543.79%

11.Gemini 3.1 Pro Preview41.87%

12.o4-mini39.22%

13.Gemini 2.5 Pro38.92%

14.Gemini 2.5 Flash38.41%

15.GPT-4.134.32%

16.Llama 4 Maverick22.36%

higher is betterAs of August 26, 2026Scale: 0% - 100%

Source: Scale AI

CC BY 4.0

gradually.ai

Methodology and context

Comparison cohort: scale-prbench-finance-full
Metric: weighted rubric score
Scale: 0-100
Participants: 16
Methodology and context: 600 finance tasks with task-specific rubrics and o4-mini as judge.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Evaluator: Scale AI. Dataset type: public. Attempts: pass@1. Score display: 43.79% ± 0.4. Source ID: scale-ai. Evidence capture ID: curated-scale-ai-2026-08-26-db93e34ac02d. Source payload hash: db93e34ac02d5628c3ec931b3f6d17257e1b7695c21cf8aee846e136474ca9b9. Parser version: curated-v1. Imported at: 2026-09-29

Source: Scale AI

CorpFin v2

Overall

1.Claude Opus 573.19%

2.Claude Fable 571.83%

3.Kimi K371.56%

4.Inkling Small69.62%

5.Inkling68.57%

6.Grok 4.368.53%

7.GPT-5.568.42%

8.Kimi K2.568.26%

9.MiniMax M368.1%

10.Grok 4.567.41%

11.Claude Sonnet 566.98%

12.Kimi K2.666.74%

13.Claude Opus 4.866.71%

14.Gemini 3 Flash Preview66.43%

15.Grok 4.666.16%

16.GLM-5.266.12%

17.Claude Opus 4.766.08%

18.Qwen 3.8 Max 090265.85%

19.Nemotron 3 Ultra65.46%

20.Claude Sonnet 4.665.31%

20.GPT-5.6 Terra65.31%

22.GPT-5.465.27%

23.Gemini 3.5 Flash64.69%

24.Gemini 3.1 Pro Preview64.49%

25.GLM-5.164.45%

26.GPT-5.6 Sol64.38%

27.GPT-5.6 Luna64.22%

28.Qwen 3.7 Max63.71%

29.Grok 4.20 Reasoning63.68%

30.Gemini 3.6 Flash63.33%

31.GPT-4.163.05%

32.GLM-562.9%

33.Qwen 3.6 27B62.32%

34.Ling 3.0 Flash61.93%

35.MiMo-V2.5-Pro61.42%

36.DeepSeek-V4-Pro61.38%

37.Claude Opus 4.561.31%

38.GPT-5.4 nano61.19%

38.MiniMax M2.761.19%

40.GPT-561.07%

41.Mistral Large 361.03%

42.GLM-4.560.96%

43.GPT-5.4 mini60.92%

44.Claude Sonnet 4.560.8%

44.Gemini 2.5 Pro60.8%

46.Gemini 3.5 Flash-Lite60.68%

47.Kimi K2 Thinking60.57%

48.Claude Haiku 4.560.3%

49.GPT-5 mini60.18%

50.MiMo-V2.559.91%

51.o359.71%

52.MiniMax M2.559.6%

53.Gemini 3.1 Flash-Lite59.36%

54.o4-mini58.97%

55.MiniMax M2.158.9%

56.Mistral Medium 3.558.78%

57.GPT OSS 120B58.24%

58.GPT-4.1 mini57.93%

59.GLM-4.656.84%

60.DeepSeek-V3 032454.74%

61.Trinity Large Thinking54.66%

62.Nemotron 3.5 Lightning54.24%

63.DeepSeek-R154.12%

64.GPT OSS 20B53.15%

65.DeepSeek-V352.49%

66.DeepSeek-V3.151.48%

67.Kimi K2 071150.39%

68.DeepSeek-V3.247.94%

69.Llama 4 Scout46.78%

70.GLM-4.746.39%

71.Command A45.96%

72.GPT-4o mini45.46%

73.o3-mini45.3%

74.Mistral Small 3.144.17%

75.GPT-4.1 nano42.08%

higher is betterAs of September 29, 2026Scale: 0% - 100%

Source: Vals AI

CC BY 4.0

gradually.ai

Methodology and context

Comparison cohort: vals-corp-fin-v2:2:overall
Metric: accuracy
Scale: 0-100
Participants: 75
Methodology and context: Source model ID: anthropic/claude-sonnet-4-5-20250929. Source model name: Claude Sonnet 4.5. Model snapshot: claude-sonnet-4-5-20250929. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2026-08-12. Latency in seconds: 13.383. Cost per task in USD: 0.237245. Source ID: vals-ai. Evidence capture ID: vals-corp_fin_v2-2026-09-29-83186cfafe52. Source payload hash: 83186cfafe526f0b8fe2ae86af05e8209fe799c8bfa48292ef1fef5db66650ee. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source: Vals AI

PRBench Legal

Overall

500 realistic legal questions with weighted criteria across jurisdictions and practice areas. The benchmark measures professional response quality but cannot replace legal review or human accountability.

1.Claude Fable 552.56%

2.Claude Opus 4.652.27%

3.GPT-5.6 Sol50.5%

4.GPT-5 Pro49.89%

5.o3-pro49.67%

6.GPT-548.96%

7.o348.57%

8.GPT-5.444.35%

9.Claude Opus 4.544.21%

10.Gemini 3.1 Pro Preview44.02%

11.Gemini 2.5 Pro41.43%

12.Gemini 2.5 Flash41.02%

13.Claude Sonnet 4.540.84%

14.o4-mini38.11%

15.GPT-4.136.48%

16.Llama 4 Maverick24.84%

higher is betterAs of August 26, 2026Scale: 0% - 100%

Source: Scale AI

CC BY 4.0

gradually.ai

Methodology and context

Comparison cohort: scale-prbench-legal-full
Metric: weighted rubric score
Scale: 0-100
Participants: 16
Methodology and context: 500 legal tasks with task-specific rubrics and o4-mini as judge.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Evaluator: Scale AI. Dataset type: public. Attempts: pass@1. Score display: 40.84% ± 0.4. Source ID: scale-ai. Evidence capture ID: curated-scale-ai-2026-08-26-cfd9f8c7d859. Source payload hash: cfd9f8c7d85930ec8d7acf7515a8200c29b116b3913b0affc48e4e68b3103115. Parser version: curated-v1. Imported at: 2026-09-29

Source: Scale AI

Health

MedCode

Overall

1.Claude Opus 563.57%

2.Gemini 3.1 Pro Preview59.06%

3.Claude Fable 556.07%

4.Gemini 3 Flash Preview55.92%

5.Gemini 3.5 Flash55.83%

6.Claude Opus 4.754.86%

7.Claude Fable 5.153.51%

8.Gemini 3.7 Flash53.39%

9.Claude Opus 4.853.22%

10.Gemini 3.6 Flash53.15%

11.Claude Sonnet 5.552.92%

12.Gemini 2.5 Pro50.59%

13.Claude Opus 5.549.8%

14.GPT-549.63%

15.Grok 4.749.55%

16.GPT-5.549.1%

17.Kimi K348.88%

18.GPT-6 Astra48.49%

19.Claude Opus 4.648.24%

20.Gemini 3.8 Flash48.14%

21.Gemini 3.1 Flash-Lite47.6%

22.Claude Sonnet 547.54%

23.o347.29%

24.GPT-6 Sol47.07%

25.MiniMax M346.29%

26.Claude Opus 4.545.17%

27.Grok 4.644.71%

28.GPT-6 Luna44.69%

29.GPT-5.6 Sol43.97%

30.Gemini 3.5 Flash-Lite43.49%

31.GPT-5.6 Terra43.41%

32.Grok 4.543.29%

33.GPT-5 mini43.05%

34.GLM-5.342.86%

35.GPT-5.6 Luna42.39%

36.GLM-5.141.6%

37.GPT-5.441.29%

38.Inkling41.19%

39.DeepSeek-V4.1-Flash41.17%

40.GPT-5.4 nano41.03%

41.GLM-5.240.77%

42.Qwen 3.8 Max 090240.67%

43.Claude Sonnet 4.540.57%

44.DeepSeek-V4-Pro40.46%

45.Kimi K2.640.14%

46.Kimi K2.539.32%

47.Qwen 3.7 Max38.75%

48.Nemotron 3 Ultra38.62%

49.Grok 4.338.07%

50.Inkling Small37.89%

51.MiniMax M2.734.44%

52.MiniMax M2.134.08%

53.o4-mini33.79%

54.Mistral Medium 3.533.75%

55.GLM-4.732.77%

56.MiMo-V2.5-Pro32.48%

57.Ling 3.0 Flash32.28%

58.Grok 4.20 Reasoning32.16%

59.MiMo-V2.531.9%

60.Mercury 2.531.33%

61.GPT-5 nano30.44%

62.Qwen 3.8 27B28.7%

63.Gemini 2.5 Flash-Lite27.12%

64.Llama 4 Scout23.31%

higher is betterAs of September 29, 2026Scale: 0% - 100%

Source: Vals AI

CC BY 4.0

gradually.ai

Methodology and context

Comparison cohort: vals-medcode:1:overall
Metric: accuracy
Scale: 0-100
Participants: 64
Methodology and context: Source model ID: anthropic/claude-sonnet-4-5-20250929. Source model name: Claude Sonnet 4.5. Model snapshot: claude-sonnet-4-5-20250929. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2026-09-26. Latency in seconds: 12.006. Cost per task in USD: 0.042403. Source ID: vals-ai. Evidence capture ID: vals-medcode-2026-09-29-7062229b5ab4. Source payload hash: 7062229b5ab4a5e16f014a78fd550f9f34e7147be6be796b610ab1ca3de055ab. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source: Vals AI

MedScribe

Overall

1.Claude Opus 5.591.43%

2.Claude Fable 5.191.29%

3.Claude Sonnet 5.591.1%

4.Claude Opus 590.99%

5.Grok 4.789.38%

6.GLM-5.3-Flash88.94%

7.GLM-5.388.81%

8.Claude Fable 588.52%

9.Kimi K387.96%

10.GPT-6 Astra87.91%

11.MiniMax M387.25%

12.Grok 4.586.88%

13.GPT-5.586.87%

14.Claude Opus 4.686.74%

15.Grok 4.686.53%

16.Claude Opus 4.885.76%

17.DeepSeek-V4.1-Flash85.5%

18.Inkling85.41%

19.GPT-5.6 Sol85.23%

20.Qwen 3.8 Max 090284.95%

21.Claude Sonnet 4.584.52%

22.Gemini 3.8 Flash84.5%

23.GPT-5.6 Luna84.39%

24.Inkling Small84.11%

25.Gemini 3.7 Flash83.94%

26.Qwen 3.8 27B83.85%

27.MiMo-V2.5-Pro83.73%

28.GPT-6 Luna83.71%

29.GPT-583.65%

30.GLM-5.283.53%

31.Claude Opus 4.583.25%

32.Claude Opus 4.782.95%

33.GPT-5.6 Terra82.87%

34.GPT-6 Sol82.03%

35.Ling 3.0 Flash80.9%

36.MiniMax M2.180.78%

37.GPT-5 mini80.58%

38.MiniMax M2.779.87%

39.Gemini 3.6 Flash79.66%

40.Qwen 3.7 Max79.4%

41.Kimi K2.678.15%

42.GPT-5.477.55%

43.GPT-5.4 nano77.09%

44.o376.65%

45.Gemini 3.5 Flash76.57%

46.Kimi K2.576.44%

47.Gemini 3.1 Pro Preview76.11%

48.Claude Sonnet 576.05%

49.DeepSeek-V4-Pro75.14%

50.Grok 4.374.4%

51.Gemini 2.5 Pro73.55%

52.GPT-5 nano72.87%

53.Gemini 2.5 Flash-Lite72.83%

54.GLM-5.172.27%

55.MiMo-V2.572.15%

56.Gemini 3.5 Flash-Lite70.89%

57.Gemini 3 Flash Preview69.92%

58.o4-mini69.14%

59.GLM-4.768.63%

60.Mistral Medium 3.567.73%

61.Gemini 3.1 Flash-Lite63.9%

62.Grok 4.20 Reasoning63.41%

63.Mercury 2.555.09%

64.Llama 4 Scout50.59%

65.Nemotron 3.5 Lightning4.27%

higher is betterAs of September 29, 2026Scale: 0% - 100%

Source: Vals AI

CC BY 4.0

gradually.ai

Methodology and context

Comparison cohort: vals-medscribe:1:overall
Metric: accuracy
Scale: 0-100
Participants: 65
Methodology and context: Source model ID: anthropic/claude-sonnet-4-5-20250929. Source model name: Claude Sonnet 4.5. Model snapshot: claude-sonnet-4-5-20250929. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2026-09-26. Latency in seconds: 44.419. Cost per task in USD: 0.054649. Source ID: vals-ai. Evidence capture ID: vals-medscribe-2026-09-29-6533d95c8055. Source payload hash: 6533d95c80553d2dd26927f7269ab6454c589d2181f1858c662a4f493b5685c7. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source: Vals AI

Recorded measurement details

BFCL v4: Overall (Show measurement, test conditions, and source)

Source value: 73.24
Score: 73.24
Metric: unweighted overall accuracy
Unit: %
Benchmark version: v4, commit f7cf735
Category: agentic
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: bfcl-v4-f7cf735-overall
Source type: official-leaderboard
Evaluator: Berkeley BFCL
Status: active
Retrieved at: 2026-09-02
Methodology: Unweighted overall score across agentic, multi-turn, and single-turn function-calling tests.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Model snapshot: claude-sonnet-4-5-20250929-fc. Evaluator: Berkeley BFCL. Dataset type: public. Leaderboard date: 2026-04-12. Source ID: berkeley-bfcl. Evidence capture ID: curated-berkeley-bfcl-2026-09-02-171020422eb3. Source payload hash: 171020422eb31aa497dbf9dd95b72c153c0db68a3eceec63a4c059bade96f4c2. Parser version: curated-v1. Imported at: 2026-09-29
Comparison methodology: Unweighted overall score across agentic, multi-turn, and single-turn function-calling tests.

Source Berkeley BFCL

CORE-Bench Hard: Automatic validation (Show measurement, test conditions, and source)

Source value: 62.22
Score: 62.22
Metric: automatic accuracy
Unit: %
Benchmark version: 45-paper public hard set
Category: coding
Task: automatic
Task label: Automatic validation
Direction: Higher is better
Comparison cohort: hal-corebench-hard-claude-code-automatic
Source type: independent-evaluator
Evaluator: Princeton HAL
Status: active
Retrieved at: 2026-08-06
Methodology: Claude Code on 45 public paper reproductions. Total cost is divided by 45.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Evaluator: Princeton HAL. Dataset type: public. Harness: Claude Code. Agent: Claude Code. Cost per task in USD: 1.5184444444444445. Source ID: princeton-hal. Evidence capture ID: curated-princeton-hal-2026-08-06-d289e35e619a. Source payload hash: d289e35e619aac7b28355c86830d55c245fc6b287e2327069860ac0522c95e8b. Parser version: curated-v1. Imported at: 2026-09-29
Comparison methodology: Claude Code on 45 public paper reproductions. Total cost is divided by 45.

Source Princeton HAL

SciCode: Accuracy (Show measurement, test conditions, and source)

Source value: 3.08
Score: 3.08
Metric: accuracy
Unit: %
Benchmark version: 65-problem main set
Category: coding
Task: overall
Task label: Accuracy
Direction: Higher is better
Comparison cohort: hal-scicode-scicode-tool-calling-overall
Source type: independent-evaluator
Evaluator: Princeton HAL
Status: active
Retrieved at: 2026-08-05
Methodology: SciCode Tool Calling on 65 main scientific problems. Total cost is divided by 65.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Evaluator: Princeton HAL. Dataset type: public. Harness: SciCode Tool Calling. Agent: SciCode Tool Calling. Cost per task in USD: 1.6044615384615386. Source ID: princeton-hal. Evidence capture ID: curated-princeton-hal-2026-08-05-c67a4c4ed842. Source payload hash: c67a4c4ed8428dfb71febdf30c9d21b46b0644189ef4d4de5ce334f3f2edc12c. Parser version: curated-v1. Imported at: 2026-09-29
Comparison methodology: SciCode Tool Calling on 65 main scientific problems. Total cost is divided by 65.

Source Princeton HAL

MT-AgentRisk: Attack success rate (Show measurement, test conditions, and source)

Source value: 71.78
Score: 71.78
Metric: attack success rate
Unit: %
Benchmark version: ICML 2026 paper
Category: cybersecurity
Task: asr
Task label: Attack success rate
Direction: Lower is better
Comparison cohort: mt-agent-risk:multi-turn:asr
Source type: independent-evaluator
Evaluator: MT-AgentRisk
Status: active
Retrieved at: 2026-08-09
Methodology: OpenHands without a defense layer, judged by GPT-4.1. Single-turn and multi-turn cohorts remain separate.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Evaluator: MT-AgentRisk. Dataset type: public. Leaderboard date: 2026-02-13. Harness: OpenHands. Source ID: mt-agent-risk. Evidence capture ID: curated-mt-agent-risk-2026-08-09-f550ff9bf634. Source payload hash: f550ff9bf6346dc509fb04e302b64acda7943dd7e5863a5ca3206286480a11a5. Parser version: curated-v1. Imported at: 2026-09-29
Comparison methodology: OpenHands without a defense layer, judged by GPT-4.1. Single-turn and multi-turn cohorts remain separate.
Context: Table 4 reports a 15.43% refusal rate, while Table 9 in the same paper reports 15.34%. This row uses the direct single-turn versus multi-turn comparison table.

Source MT-AgentRisk

MT-AgentRisk: Refusal rate (Show measurement, test conditions, and source)

Source value: 15.43
Score: 15.43
Metric: refusal rate
Unit: %
Benchmark version: ICML 2026 paper
Category: cybersecurity
Task: refusal
Task label: Refusal rate
Direction: Higher is better
Comparison cohort: mt-agent-risk:multi-turn:refusal
Source type: independent-evaluator
Evaluator: MT-AgentRisk
Status: active
Retrieved at: 2026-08-09
Methodology: OpenHands without a defense layer, judged by GPT-4.1. Single-turn and multi-turn cohorts remain separate.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Evaluator: MT-AgentRisk. Dataset type: public. Leaderboard date: 2026-02-13. Harness: OpenHands. Source ID: mt-agent-risk. Evidence capture ID: curated-mt-agent-risk-2026-08-09-9e7ba7aa3157. Source payload hash: 9e7ba7aa3157bc30470265facc4bb25edf1a9b2f6cf7b733102470330ed4b034. Parser version: curated-v1. Imported at: 2026-09-29
Comparison methodology: OpenHands without a defense layer, judged by GPT-4.1. Single-turn and multi-turn cohorts remain separate.
Context: Table 4 reports a 15.43% refusal rate, while Table 9 in the same paper reports 15.34%. This row uses the direct single-turn versus multi-turn comparison table.

Source MT-AgentRisk

MultiChallenge: Overall (Show measurement, test conditions, and source)

Source value: 55.32
Score: 55.32
Metric: rubric pass rate
Unit: %
Benchmark version: four multi-turn challenge types
Category: productivity
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: scale-multichallenge-overall
Confidence interval: 1.47
Source type: official-leaderboard
Evaluator: Scale AI
Status: active
Retrieved at: 2026-08-26
Methodology: Multi-turn conversations with binary task-specific rubrics.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Evaluator: Scale AI. Dataset type: public. Attempts: pass@1. Reasoning effort: thinking. Score display: 55.32% ± 1.47. Source ID: scale-ai. Evidence capture ID: curated-scale-ai-2026-08-26-93c4c6df84ba. Source payload hash: 93c4c6df84baa57ecc302765b454f1d1d402750d6ee8968f2085772a44e192b1. Parser version: curated-v1. Imported at: 2026-09-29
Comparison methodology: Multi-turn conversations with binary task-specific rubrics.

Source Scale AI

PRBench Finance: Overall (Show measurement, test conditions, and source)

Source value: 43.79
Score: 43.79
Metric: weighted rubric score
Unit: %
Benchmark version: full finance set
Category: finance
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: scale-prbench-finance-full
Confidence interval: 0.4
Source type: official-leaderboard
Evaluator: Scale AI
Status: active
Retrieved at: 2026-08-26
Methodology: 600 finance tasks with task-specific rubrics and o4-mini as judge.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Evaluator: Scale AI. Dataset type: public. Attempts: pass@1. Score display: 43.79% ± 0.4. Source ID: scale-ai. Evidence capture ID: curated-scale-ai-2026-08-26-db93e34ac02d. Source payload hash: db93e34ac02d5628c3ec931b3f6d17257e1b7695c21cf8aee846e136474ca9b9. Parser version: curated-v1. Imported at: 2026-09-29
Comparison methodology: 600 finance tasks with task-specific rubrics and o4-mini as judge.

Source Scale AI

PRBench Legal: Overall (Show measurement, test conditions, and source)

Source value: 40.84
Score: 40.84
Metric: weighted rubric score
Unit: %
Benchmark version: full legal set
Category: legal
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: scale-prbench-legal-full
Confidence interval: 0.4
Source type: official-leaderboard
Evaluator: Scale AI
Status: active
Retrieved at: 2026-08-26
Methodology: 500 legal tasks with task-specific rubrics and o4-mini as judge.. Source model ID: curated/claude-sonnet-4.5. Source model name: Claude Sonnet 4.5. Provider reported by source: Anthropic. Evaluator: Scale AI. Dataset type: public. Attempts: pass@1. Score display: 40.84% ± 0.4. Source ID: scale-ai. Evidence capture ID: curated-scale-ai-2026-08-26-cfd9f8c7d859. Source payload hash: cfd9f8c7d85930ec8d7acf7515a8200c29b116b3913b0affc48e4e68b3103115. Parser version: curated-v1. Imported at: 2026-09-29
Comparison methodology: 500 legal tasks with task-specific rubrics and o4-mini as judge.

Source Scale AI

CorpFin v2: Overall (Show measurement, test conditions, and source)

Source value: 60.8
Score: 60.8
Metric: accuracy
Unit: %
Benchmark version: 2
Category: finance
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: vals-corp-fin-v2:2:overall
Source type: independent-evaluator
Evaluator: Vals AI
Status: active
Retrieved at: 2026-09-29
Methodology: Source model ID: anthropic/claude-sonnet-4-5-20250929. Source model name: Claude Sonnet 4.5. Model snapshot: claude-sonnet-4-5-20250929. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2026-08-12. Latency in seconds: 13.383. Cost per task in USD: 0.237245. Source ID: vals-ai. Evidence capture ID: vals-corp_fin_v2-2026-09-29-83186cfafe52. Source payload hash: 83186cfafe526f0b8fe2ae86af05e8209fe799c8bfa48292ef1fef5db66650ee. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source Vals AI

MedCode: Overall (Show measurement, test conditions, and source)

Source value: 40.569
Score: 40.569
Metric: accuracy
Unit: %
Benchmark version: 1
Category: health
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: vals-medcode:1:overall
Source type: independent-evaluator
Evaluator: Vals AI
Status: active
Retrieved at: 2026-09-29
Methodology: Source model ID: anthropic/claude-sonnet-4-5-20250929. Source model name: Claude Sonnet 4.5. Model snapshot: claude-sonnet-4-5-20250929. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2026-09-26. Latency in seconds: 12.006. Cost per task in USD: 0.042403. Source ID: vals-ai. Evidence capture ID: vals-medcode-2026-09-29-7062229b5ab4. Source payload hash: 7062229b5ab4a5e16f014a78fd550f9f34e7147be6be796b610ab1ca3de055ab. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source Vals AI

MedScribe: Overall (Show measurement, test conditions, and source)

Source value: 84.515
Score: 84.515
Metric: accuracy
Unit: %
Benchmark version: 1
Category: health
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: vals-medscribe:1:overall
Source type: independent-evaluator
Evaluator: Vals AI
Status: active
Retrieved at: 2026-09-29
Methodology: Source model ID: anthropic/claude-sonnet-4-5-20250929. Source model name: Claude Sonnet 4.5. Model snapshot: claude-sonnet-4-5-20250929. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2026-09-26. Latency in seconds: 44.419. Cost per task in USD: 0.054649. Source ID: vals-ai. Evidence capture ID: vals-medscribe-2026-09-29-6533d95c8055. Source payload hash: 6533d95c80553d2dd26927f7269ab6454c589d2181f1858c662a4f493b5685c7. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source Vals AI

Agent Poker Bench: Overall (Show measurement, test conditions, and source)

Source value: 1055.504459218733
Score: 1055.504459218733
Metric: TrueSkill rating
Unit: rating
Benchmark version: 1
Category: agentic
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: vals-poker-agent:1:overall
Source type: independent-evaluator
Evaluator: Vals AI
Status: active
Retrieved at: 2026-09-29
Methodology: Source model ID: anthropic/claude-sonnet-4-5-20250929. Source model name: Claude Sonnet 4.5. Model snapshot: claude-sonnet-4-5-20250929. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2025-12-23. Latency in seconds: 16.67560795007542. Standard error: 0. Cost per task in USD: 0.027591309991158296. Source ID: vals-ai. Evidence capture ID: vals-poker_agent-2026-09-29-f29b75137483. Source payload hash: f29b751374830275dd1fba4c8a4f029b69433858ae99420aa90befe3093d0096. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source Vals AI

Page 1 of 4

Predecessor and successor

Sources and data date

Type	Evidence and data date
Speed	Anthropic, Non-reasoning. Output speed after the first streaming chunk, not total response time. Classified as deprecated by Artificial Analysis. Artificial Analysis maintains only the default 10K-input workload there. Retrieved September 8, 2026. Source: Artificial Analysis · Artificial Analysis (retrieved September 8, 2026)Speed measurement methodology
Price	Base API prices without caching or batch discounts. Higher context tiers and other rates are listed under Costs. Source: Pricing · API pricing
Input	Documented modalities. Retrieved September 8, 2026. Source: Model documentation · Model metadata (retrieved September 8, 2026) · Sonnet 4.5 model overview · Editorial description reviewed October 2, 2026
Output	Documented modalities. Retrieved September 8, 2026. Source: Model documentation · Model metadata (retrieved September 8, 2026) · Sonnet 4.5 model overview · Editorial description reviewed October 2, 2026
Research status	Research date September 8, 2026. 3 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not found in the inspected sources	parameters-billions Original wording Anthropic does not publish parameter counts for this model.
Not found in the inspected sources	active-parameters-billions Original wording Anthropic does not publish active parameter counts for this model.
Additional source	Model metadata
Additional source	Context window data
Additional source	Berkeley BFCL (retrieved September 2, 2026)
Additional source	Princeton HAL (retrieved August 5, 2026)
Additional source	Princeton HAL (retrieved August 6, 2026)
Additional source	Princeton HAL (retrieved August 5, 2026)
Additional source	Princeton HAL (retrieved August 6, 2026)
Additional source	Princeton HAL (retrieved August 5, 2026)
Additional source	MCP-Universe (retrieved August 9, 2026)
Additional source	MT-AgentRisk (retrieved August 9, 2026)
Additional source	Scale AI (retrieved August 26, 2026)
Additional source	Scale AI (retrieved August 26, 2026)
Additional source	Scale AI (retrieved August 26, 2026)
Additional source	Scale AI (retrieved August 26, 2026)
Additional source	Scale AI (retrieved August 26, 2026)
Additional source	Scale AI (retrieved August 26, 2026)
Additional source	Scale AI (retrieved August 26, 2026)
Additional source	Vals AI (retrieved September 29, 2026)
Additional source	Vals AI (retrieved September 29, 2026)
Additional source	Vals AI (retrieved September 29, 2026)
Additional source	Vals AI (retrieved September 29, 2026)
Additional source	Vals AI (retrieved September 29, 2026)
Additional source	Anthropic (retrieved September 8, 2026) · Anthropic, Introducing Claude Sonnet 4.5 · Editorial description reviewed October 2, 2026
Additional source	OpenRouter: Popularity index (derived), CC BY 4.0
Additional source	Vercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional source	Creative Commons Attribution 4.0
Additional source	Anthropic "Introducing Claude Sonnet 4.6" announcement (2026-09-19)

Claude Sonnet 4.5

Specifications and access

Published benchmarks

Agent tasks

BFCL v4

Agent Poker Bench

Coding

CORE-Bench Hard

SciCode

Cybersecurity

MT-AgentRisk

MT-AgentRisk

MultiChallenge

Finance

PRBench Finance

CorpFin v2

PRBench Legal

Health

MedCode

MedScribe

Recorded measurement details

Our tests

Pelican on a bicycle

Popularity

Published prices

Compare this model

Predecessor and successor

Sources and data date