Beam 1.0

7 sources, 1 independent of the vendor: centre 18; the sources disagree widely, so shown as a band.

• Artificial Analysis · third party
• 16.7 GPQA Diamond: 90.5 · vendor
• 16.1 Humanity's Last Exam: 36.2 · vendor
• 12.6 SWE-bench Verified: 80.9 · vendor
• 23.6 Terminal-Bench 2.1: 80.1 · vendor
• 17.5 aime 2026: 97.8 · vendor
• 24.7 tau3 banking: 38 · vendor
• Humanity's Last Exam: 51.2 · vendor · not used: a different reasoning mode, not mixed
• automationbench public: 37 · vendor · not used: no reference models for this benchmark
• browsecomp: 77.4 · vendor · not used: no reference models for this benchmark
• deepsearchqa: 80.1 · vendor · not used: no reference models for this benchmark
• deepswe v1 1: 44.4 · vendor · not used: no reference runs on the same harness
• ifbench: 79.7 · vendor · not used: only 0 reference models
• longbench v2: 66.9 · vendor · not used: no reference models for this benchmark
• mcp atlas: 78.7 · vendor · not used: no reference models for this benchmark
• swe atlas codebase qna: 34.6 · vendor · not used: no reference models for this benchmark
• swe bench multilingual: 78 · vendor · not used: correlation 0.328 too low
• swe bench pro v1: 65.5 · vendor · not used: no reference models for this benchmark
• swe bench pro v2 hard: 77.2 · vendor · not used: no reference models for this benchmark

how FTI is computed