ProgramBench
We haven't written a description for this benchmark yet; see the card notes below for how each maker ran it.
- ModelScoreYour cost / mo
- 1Kimi K3 · Moonshot AIbest value77.8%
- 2Muse Wonder 1.0 · Meta50.0%
“Best value”: no other model here reports a higher score at a lower cost for your workload. Scores are self-reported by each maker with different settings, so treat close calls as ties.
How each maker reported it
- Kimi K3 model card §3
Kimi K3 at max effort. Moonshot's own table; some cells (CritPt, AA-LCR, SciCode, GDPval-AA, AA-Briefcase, APEX-Agents, Harvey) are cited from Artificial Analysis or other leaderboards, and several harnesses differ by model. HLE and MMMU-Pro shown without tools.
- Muse Wonder 1.0 p.3–6
All models at max reasoning effort, except Muse Wonder Full on SWE-Marathon (xhigh). Meta uses third-party or provider-reported results where they exist (Artificial Analysis, Vals AI, Surge, Datacurve, Abundant AI); otherwise the highest of its own validated run, self-report or official leaderboard. Personal Superintelligence Index and MemoryBench are Meta-internal. SWE-Marathon uses each vendor's own harness (Codex, Claude Code), so that row is not a common-harness comparison. No OSWorld binary figure exists for Opus 5.5.
Only each model's own maker's number is shown; scores a maker reports for rival models are left out.