Atlas

Benchmarks

← All benchmarks

FrontierChallenge (Pass Rate)

Science · 2026-08-25

FrontierChallenge, from Apodex, releases 97 of its 300 end-to-end scientific workflows (74 Hard, 23 Medium) in six reporting domains: quantum chemistry, molecular dynamics, materials characterization, analytical chemistry, life science and electrochemistry/environment. Each task gives an agent fixed scientific inputs and an output contract of required deliverables, such as reports, structured tables, figures, executable code and simulation products, and a task-specific grader scores the submitted bundle from 0 to 100 with deterministic checks; 77 tasks also judge the report with GPT-5.6 Sol, averaging three judge passes. The score is the Pass Rate, the share of the 97 tasks fully completed in one run of a model in an agent scaffold (Codex, Claude Code or Apodex's Frontier Agent), in percent. The published table is historical: it counted a task as complete when its native score was at least 99.9, while the repository's scoring policy since September 23, 2026 requires a completed task_score strictly above 0.999, and Apodex has not recomputed the historical numbers.

Benchmark source: arXiv

Top models (higher is better)

ModelScoreSource
GPT-5.6 Sol20.6FrontierChallenge paper Table 2 (GPT-5.6 Sol / Codex, Overall) · Apodex
Grok 4.620.6FrontierChallenge paper Table 2 (Grok 4.6 / Claude Code, Overall) · Apodex
Claude Opus 517.5FrontierChallenge paper Table 2 (Claude Opus 5 / Claude Code, Overall) · Apodex
Kimi K317.5FrontierChallenge paper Table 2 (Kimi K3 / Claude Code, Overall) · Apodex
GPT-5.6 Terra15.5FrontierChallenge paper Table 2 (GPT-5.6 Terra (max) / Codex, Overall) · Apodex
Qwen3.8-Max15.5FrontierChallenge paper Table 2 (Qwen 3.8 Max / Claude Code, Overall) · Apodex
DeepSeek V4 Pro 081313.4FrontierChallenge paper Table 2 (DeepSeek V4 Pro-0813 / Claude Code, Overall) · Apodex
Apodex 1.112.4FrontierChallenge paper Table 2 (Apodex 1.1 / Frontier Agent (Agent Team), Overall) · Apodex
DeepSeek-V4-Flash-073112.4FrontierChallenge paper Table 2 (DeepSeek V4 Flash-0731 / Claude Code, Overall) · Apodex
Gemini 3.7 Flash10.3FrontierChallenge paper Table 2 (Gemini 3.7 Flash / Claude Code, Overall) · Apodex
Qwen3.5 397B A17B4.1FrontierChallenge paper Table 2 (Qwen3.5-397B-A17B / Claude Code, Overall) · Apodex
GLM-5.23.1FrontierChallenge paper Table 2 (GLM-5.2 / Claude Code, Overall) · Apodex
Loading Atlas data…