BenchAtlas

BenchAtlas Index · snapshot 2026-07-13

The benchmark landscape, in one instrument panel

146 ranked systems · 10 live sources · every number linked to its source

compare two models →

Top of the index

full rankings →
#SystemIndexΔ
1gpt-5.5-pre-releasexhigh effort97.5·
2gemini-3-1-pro-preview94.4·
3gpt-5.6-solmax effort94.4·
4muse-spark93.7·
5claude-opus-4-632K budget88.7·
6gpt-5.4-2026-03-05xhigh effort88.5·
7claude-opus-4-664K budget87.5·
8kimi-k2.684.5·

Category leaders

as of 2026-07-13
Agentic Coding ×0.2
gpt-5.6-terramax effort
98.6
1f
Reasoning & Math ×0.2
openai-gpt-5-6-solmax effort
99.7
1f
Coding ×0.15
claude-fable-5max effort
98.6
1f
Human Preference ×0.15
claude-fable-5
99.9
1f
Knowledge & Science ×0.15
gemini-3-1-pro-previewhigh effort
99.3
2f
Long Context & Instruction ×0.15
gemini-3-1-pro-previewhigh effort
98.6
1f

Best category percentile among all scored systems (×weight in the Index); “Nf” = benchmark families behind the score.

BenchAtlas Index, top 10

bands are leave-one-family-out ranges
78848995100
1. gpt-5.5-pre-release (xhigh effort)
97.5
2. gemini-3-1-pro-preview
94.4
3. gpt-5.6-sol (max effort)
94.4
4. muse-spark
93.7
5. claude-opus-4-6 (32K budget)
88.7
6. gpt-5.4-2026-03-05 (xhigh effort)
88.5
7. claude-opus-4-6 (64K budget)
87.5
8. kimi-k2.6
84.5
9. gpt-5.6-sol (xhigh effort)
83.1
10. claude-opus-4-8 (xhigh effort)
82.4

A band shows how far a system’s score moves when any single benchmark family is left out — wide bands mean the score leans on few families.

Biggest moves

vs previous snapshot
First snapshot (2026-07-13) — movement appears after the next one.

Recent releases

full timeline →
gpt-5.6-luna_unknown
OpenAI
2026-07-09
license unknown
gpt-5.6-terra_unknown
OpenAI
2026-07-09
license unknown
gpt-5.6-sol_unknown
OpenAI
2026-07-09
license unknown
gpt-5.6-sol_prounknown
OpenAI
2026-07-09
license unknown
gpt-5.6-sol_promax
OpenAI
2026-07-09
license unknown
gpt-5.6-sol
OpenAI
2026-07-09
license unknown
gpt-5.6-luna
OpenAI
2026-07-09
license unknown
gpt-5.6-terra
OpenAI
2026-07-09
license unknown

Newly observed results

append-only fact table, newest first
SystemBenchmarkScore
DeepCoder-14B-PreviewLiveCodeBench v560.6%
Llama-3_1-Nemotron-Nano-8B-v1LiveCodeBench v551.3%
llama-3.1-nemotron-ultra-253b-v1LiveCodeBench v564.5%
grok-3-mini_highLiveCodeBench v569.9%
MetaStone-L1-7BLiveCodeBench v539.3%
QwQ-Max-PreviewLiveCodeBench v565.6%
Kimi-k1.6-IOILiveCodeBench v565.9%
Kimi-k1.6-IOI-highLiveCodeBench v573.8%

Source freshness

source transparency →
ARC Prize Leaderboard3h ago
Artificial Analysisnever run
Epoch AI Benchmarking Hub3h ago
LiveBench3h ago
LiveCodeBench Leaderboard3h ago
LMArena Leaderboard Dataset3h ago
Scale Labs3h ago
SWE-bench Leaderboard3h ago
Terminal-Bench Leaderboard3h ago
Vals AI3h ago

Health derived from ingestion runs: a streak of 3+ failures = failing, 1–2 = degraded. “Never run” sources are registered adapters awaiting their first scheduled ingestion.

The BenchAtlas Index is a percentile-based composite across independent benchmark families (methodology). Missing coverage is shown as absence, never as a low score, and every published number traces back to a raw source snapshot.