models.fyi

Every score below is at the effort you pick — not the press-release number.

Public leaderboards quote each model at maximum reasoning effort. Nobody runs xhigh for everything. Choose a tier and see what the models are actually worth there — models that don't support the tier resolve to their nearest one (shown per row).

ProprietaryOpen weights
↑ better · ← cheaper

Cost per typical task, log scale (1K in + 600 out + reasoning tokens)

#ModelEffort usedIndexGPQAAIMESWE-benchCost / task$/M in · outtok/sCtx
1Claude Fable 5
Anthropic
High82.287.892.280.8$0.54$10 · $50651M
2GPT-5.5
OpenAI
High80.786.893.176.8$0.097$1.25 · $10140400K
3Claude Opus 4.8
Anthropic
High79.685.889.378.8$0.26$5 · $25551M
4Gemini 3 Pro
Google
High79.087.490.274.7$0.12$2 · $121301M
5Claude Sonnet 5
Anthropic
High77.284.088.076.0$0.12$3 · $15801M
6Grok 4.1
xAI
High76.283.587.471.8$0.18$3 · $1590256K
7GLM-5.2
Zhipu AI
High75.881.685.674.7$0.021$0.60 · $2.2095200K
8DeepSeek-V3.2
DeepSeek
High75.583.090.072.0$0.0039$0.28 · $0.4260128K
9Kimi K2 Thinking
Moonshot AI
High75.384.089.071.0$0.032$0.60 · $2.5050256K
10Qwen3-235B-A22B
Alibaba
High72.581.086.068.0$0.023$0.70 · $2.8070128K
11Llama 4 Maverick
Meta
Standard52.470.040.044.0$0.0008$0.27 · $0.851601M

Index = weighted composite of 8 benchmarks (weights on /benchmarks) · cost/task assumes 1K in + 600 out + the tier's typical reasoning tokens · preview data.