Which model leads PaperBench?

Q: Which model leads PaperBench?

Qwen3.8 Max Preview tops the board at 93

PaperBench Leaderboard — 10 Models Ranked

Pro Mode — raw scores, one benchmark under the lensthird-party eval harness

The field · 10 measured

47.193

10 models measured, most on the third-party eval harness. Qwen3.8 Max Preview tops the board at 93.

10 measured·Lifecycle Updated Oct 8, 2026Score basis

#◆Vendor$/M in$/M out

Leader — ignites whiteField — fill cools with scoreIndependently verified only0 of 10Bars show 47.1–93 · ◆ solid = first-party or better · outlined = arm's-length

01Qwen3.8 Max Preview+1 altAlibaba93VENDOR$3.3/M$9.9/M$0.071

02GPT-5.6 SolOpenAI90.5$4.0/M$20.0/M$0.133

03Claude Fable 5Anthropic88.8$10.0/M$50.0/M$0.338

04Claude Opus 4.8Anthropic80.3$5.0/M$25.0/M$0.187

05Claude Opus 4.5Anthropic72.9$5.0/M$25.0/M$0.206

06Qwen3.7 MaxAlibaba64.8$2.5/M$7.5/M$0.077

07GPT-5.2OpenAI63.7$1.8/M$14.0/M$0.124

08Kimi K2.5+1 altMoonshot63.5VENDOR$0.45/M$2.3/M$0.021

09MiniMax M3MiniMax52.6VENDOR$0.28/M$1.1/M$0.013

10DeepSeek-V3.2DeepSeek47.1$0.28/M$0.42/M$0.007

◆ solid = independently verified · aggregator attested · vendor attributed — outlined = cross-reference · source-attributed. Dimmed ◆ = confidence inferred. “+N alts” = alternate disclosures in the drawer with why-demoted. Prices per million tokens; $/point = input price ÷ score. Every row opens its model dossier.

Verification: 10 models scored · 0 independently verified · 6 vendor cross-reference · 4 vendor-reported · 0 with source disagreement. How these tiers are assigned