Open-source model matches closed frontier performance

The gap between open and closed models has been narrowing.

RESOLUTION CRITERIA

True if an open-weights model scores within 2% of the leading closed model on MMLU, HumanEval, and GPQA.

▲ FOR

151 days remaining provides runway

Prior cycles: DeepSeek V4 within percentage points of Claude Opus 4.7 on SWE-bench Verified — near-parity on code benchmark documented

Claude Opus 5.5 at 58 on Artificial Analysis index is an open-accessible model within 1 point of closed frontier

▼ AGAINST

Gemini 4 Carbon (closed) scores 59, topping the Artificial Analysis Intelligence Index — closed frontier maintains composite leaderboard lead over open-weight models

Carbon's cybersecurity benchmark dominance introduces a new capability dimension where open-weight models have not demonstrated equivalent performance

SWE-bench and HumanEval+ parity does not translate to aggregate leaderboard parity — resolution criterion unverified

Cumulative +52pt drift from 28% initial is extreme — ceiling discipline requires downward pressure absent resolution confirmation

HIGH time pressure: 151 days with aggregate parity criterion still unverified on composite measures

RECENT SIGNALS (4)

↑

Mistral opens weights of 1-trillion-parameter Large 4 model, with dual-use cyber concerns

TechRepublic

↕

Artificial Analysis moves most of its Intelligence Index v5 out of public view, shifting how frontier AI models are ranked

X (Daily AI Recap, Neo)

↕

Mistral's Large 4 claims frontier-level results among open models, tying China's GLM-5.3 on a coding agent test

Fortune (via ripr)

↓

Gap between closed and open-weight frontier models widens to 12 points, with Kimi K3.1 Pro leading open weights

X (Daily AI Recap, Neo)