Benchmark
CoBench
Benchmark by Anthropic
2.1
53.2%
Claude Opus 5
frontier lab
Anthropic
CoBench and operational automation included. Internal tasks are not public; lab disclosure is not independent replication.
Published evidence
Benchmark
CoBench
Benchmark by Anthropic
2.1
53.4%
Claude Mythos 5.1
Benchmark
CoBench
Benchmark by Anthropic
2.1
55.8%
Claude Opus 5.5
Benchmark
CoBench
Benchmark by Anthropic
Prior version (label unspecified)
59.6%
Claude Opus 5
Benchmark
CoBench
Benchmark by Anthropic
Prior version (label unspecified)
57.6%
Claude Mythos 5.1
Benchmark
PaperBench
Benchmark by OpenAI
Original paper v1
21.0 ± 0.8%
Claude 3.5 Sonnet (New) / BasicAgent
Benchmark
PaperBench
Benchmark by OpenAI
Original paper v1
16.1 ± 0.1%
Claude 3.5 Sonnet (New) / IterativeAgent
Benchmark
MLE-bench
Benchmark by OpenAI
Original paper v1 (2024)
7.6 ± 1.8%
Claude 3.5 Sonnet / AIDE
Benchmark
Task-completion time horizons
Benchmark by METR
TH1 (historical estimates)
289 minutes
Claude Opus 4.5 / METR horizon evaluation
Benchmark
Task-completion time horizons
Benchmark by METR
TH1.1
320 minutes
Claude Opus 4.5 / METR horizon evaluation
Benchmark
Task-completion time horizons
Benchmark by METR
TH1 (historical estimates)
86 minutes
Claude Opus 4 / METR horizon evaluation
Benchmark
Task-completion time horizons
Benchmark by METR
TH1.1
101 minutes
Claude Opus 4 / METR horizon evaluation
Benchmark
Task-completion time horizons
Benchmark by METR
TH1 (historical estimates)
56 minutes
Claude 3.7 Sonnet / METR horizon evaluation
Benchmark
Task-completion time horizons
Benchmark by METR
TH1.1
60 minutes
Claude 3.7 Sonnet / METR horizon evaluation
Benchmark
AI4AI-Bench
Benchmark by AI4AI-Bench authors
Paper v1
0.250 normalized score
Claude Opus 5 / Claude Code effort aggregate
Benchmark
AI4AI-Bench
Benchmark by AI4AI-Bench authors
Paper v1
0.145 normalized score
Claude Sonnet 5 / Claude Code effort aggregate
Indicator
Anthropic R&D Automation Index
Measure by Anthropic
August 2026 snapshot
26%
Anthropic
Indicator
Anthropic R&D Automation Index
Measure by Anthropic
August 2026 snapshot
No measured subset fully autonomous
Anthropic
Benchmark
Task-completion time horizons
Benchmark by METR
TH1 (historical estimates)
122 minutes
Claude Sonnet 4.5 / METR horizon evaluation
Benchmark
Task-completion time horizons
Benchmark by METR
TH1 (historical estimates)
114 minutes
Claude Opus 4.1 / METR horizon evaluation
Benchmark
Task-completion time horizons
Benchmark by METR
TH1 (historical estimates)
75 minutes
Claude Sonnet 4 / METR horizon evaluation
Benchmark
AI4AI-Bench
Benchmark by AI4AI-Bench authors
Paper v1
0.288 normalized score
Claude Opus 5 / Claude Code, medium effort
Indicator
Automated weak-to-strong research study
Measure by Anthropic
April 2026 chat-preference study
0.97 PGR PGR
Nine Claude Opus 4.6 research agents
Indicator
Automated weak-to-strong research study
Measure by Anthropic
April 2026 chat-preference study
0.23 PGR PGR
Human-tuned weak-to-strong method
Benchmark
PostTrainBench
Benchmark by PostTrainBench team
PostTrainBench v1.1 · September 2026 results
41.79%
Fable 5 + Opus 4.8 Max GPQA fallback
Benchmark
PostTrainBench
Benchmark by PostTrainBench team
PostTrainBench v1.1 · September 2026 results
28.56%
Claude Opus 4.7, xHigh reasoning / Claude Code
Benchmark
PostTrainBench
Benchmark by PostTrainBench team
PostTrainBench v1.1 · September 2026 results
33.84%
Claude Opus 4.8, High reasoning / Claude Code
Benchmark
PostTrainBench
Benchmark by PostTrainBench team
PostTrainBench v1.1 · September 2026 results
32.90%
Claude Opus 4.8, Max reasoning / Claude Code
Benchmark
PostTrainBench
Benchmark by PostTrainBench team
PostTrainBench v1.1 · September 2026 results
35.04%
Claude Opus 5 / Claude Code
Benchmark
PostTrainBench
Benchmark by PostTrainBench team
PostTrainBench v1.2 · September 30 release
44.65%
Claude Fable 5.1 + Opus 5 GPQA fallback / Claude Code, Max
Benchmark
PostTrainBench
Benchmark by PostTrainBench team
PostTrainBench v1.2 · September 30 release
43.80%
Claude Opus 5.5 / Claude Code, Max
Benchmark
PostTrainBench
Benchmark by PostTrainBench team
PostTrainBench v1.2 · September 30 release
37.27%
Fable 5 + Opus 4.8 Max GPQA fallback
Benchmark
PostTrainBench
Benchmark by PostTrainBench team
PostTrainBench v1.2 · September 30 release
35.66%
Claude Opus 5 / Claude Code
Benchmark
PostTrainBench
Benchmark by PostTrainBench team
PostTrainBench v1.2 · September 30 release
31.00%
Claude Opus 4.8, Max reasoning / Claude Code
Benchmark
PostTrainBench
Benchmark by PostTrainBench team
PostTrainBench v1.2 · September 30 release
30.66%
Claude Opus 4.8, High reasoning / Claude Code
Benchmark
PostTrainBench
Benchmark by PostTrainBench team
PostTrainBench v1.2 · September 30 release
24.34%
Claude Opus 4.7, xHigh reasoning / Claude Code
Benchmark
PostTrainBench
Benchmark by PostTrainBench team
PostTrainBench v1.1 / Google OpenCode evaluation, October 2026
40.2%
Claude Fable 5.1 / Google OpenCode evaluation
Benchmark
PostTrainBench
Benchmark by PostTrainBench team
PostTrainBench v1.1 / Google OpenCode evaluation, October 2026
49.3%
Claude Opus 5.5 / Google OpenCode evaluation
Benchmark
TasteVal
Benchmark by P-Zero Research
October 2026 paper v1 · 20-model main cohort
2.30× ×
Opus 5.5 (max) / TasteVal
Benchmark
TasteVal
Benchmark by P-Zero Research
October 2026 paper v1 · 20-model main cohort
1.14× ×
Opus 5.5 (max) / TasteVal
Benchmark
TasteVal
Benchmark by P-Zero Research
October 2026 paper v1 · 20-model main cohort
1.72× ×
Fable 5.1 (max) / TasteVal
Benchmark
TasteVal
Benchmark by P-Zero Research
October 2026 paper v1 · 20-model main cohort
1.09× ×
Fable 5.1 (max) / TasteVal
Benchmark
TasteVal
Benchmark by P-Zero Research
October 2026 paper v1 · 20-model main cohort
1.29× ×
Opus 5.0 (max) / TasteVal
Benchmark
TasteVal
Benchmark by P-Zero Research
October 2026 paper v1 · 20-model main cohort
1.11× ×
Opus 5.0 (max) / TasteVal
Benchmark
TasteVal
Benchmark by P-Zero Research
October 2026 paper v1 · 20-model main cohort
0.84× ×
Opus 4.8 (max) / TasteVal
Benchmark
TasteVal
Benchmark by P-Zero Research
October 2026 paper v1 · 20-model main cohort
1.01× ×
Opus 4.8 (max) / TasteVal
Benchmark
TasteVal
Benchmark by P-Zero Research
October 2026 paper v1 · 20-model main cohort
0.50× ×
Opus 4.7 (max) / TasteVal
Benchmark
TasteVal
Benchmark by P-Zero Research
October 2026 paper v1 · 20-model main cohort
0.75× ×
Opus 4.7 (max) / TasteVal
Benchmark
TasteVal
Benchmark by P-Zero Research
October 2026 paper v1 · 20-model main cohort
0.40× ×
Opus 4.6 (max) / TasteVal
Benchmark
TasteVal
Benchmark by P-Zero Research
October 2026 paper v1 · 20-model main cohort
0.78× ×
Opus 4.6 (max) / TasteVal
Benchmark
TasteVal
Benchmark by P-Zero Research
October 2026 paper v1 · 20-model main cohort
0.12× ×
Opus 4.5 (24k thinking budget) / TasteVal
Benchmark
TasteVal
Benchmark by P-Zero Research
October 2026 paper v1 · 20-model main cohort
0.37× ×
Opus 4.5 (24k thinking budget) / TasteVal
Benchmark
InnovationEval
Benchmark by Epoch AI
Initial SDPO task · October 7, 2026
2%
Claude Fable 5 · InnovationEval
Relevant sources
Claude Opus 5.5 System Card ↗
system card · 2026-09-22 · available
Measurements for understanding the pace of AI development inside frontier labs ↗
organizational research article · 2026-09-17 · available
Automated Weak-to-Strong Researcher ↗
organizational research article · 2026-04 · available
Automated Researchers Can Reliably Mitigate Alignment Failures ↗
original paper · 2026-08-28 · available
Automated Researchers Can Mitigate Well-Characterized Alignment Failures ↗
organizational research article · 2026-08-28 · available
Automated researchers can reliably mitigate alignment failures ↗
organizational research article · 2026-08-28 · available
When AI builds itself ↗
organizational research article · 2026 · available
Claude-shaped science ↗
organizational research article · 2026-10-01 · available
AI models may be capable of self-replicating without access to a datacenter ↗
organizational research article · 2026-10-12 · available
De novo design of protein binders against scorpion toxins with Claude ↗
original paper · 2026-10-20 · available
System Card: Claude Fable 5.2 & Claude Mythos 5.2 ↗
system card · 2026-11-09 · available