LiveAI model-release ecosystemWestern cycleAnchors pending

By end of 2027, the capability gap between the best open-weights model and the best closed frontier model stays below 12 months on standard benchmarks?

As of 6 Oct · 22:19 UTCResolves 31 Dec 20274 movers0 pressure moments
Timeline A · Open-Weights Catches UpReality
50%
Crowd 55%Gap +5

By end of 2027, the best open-weights model—likely a successor to DeepSeek, Llama, or Qwen—performs within 12 months of frontier closed models (GPT-5, Claude 4, or similar) on standard benchmarks like MMLU, GSM8K, and live chatbot arena matchups. Open models close the gap through sustained investment by Meta, Alibaba, Chinese labs, and smaller research teams, with inference costs low enough that capability-per-dollar favors open weights. The release cadence accelerates: major open checkpoints drop every 4–6 months, each time narrowing what frontier labs can claim as exclusive.

Timeline B · Closed Frontier Pulls AheadReality
50%
Crowd 45%Gap −5

By end of 2027, frontier closed models maintain a capability lead of 12 months or longer on standard benchmarks, despite aggressive open-weights releases. Closed labs (OpenAI, Anthropic, Google DeepMind) sustain advantages through proprietary training data, reinforcement learning from human feedback at scale, and compute resources open labs cannot match. Open models plateau in capability gain relative to compute spent, or hit scaling walls. Benchmark performance stalls while frontier labs keep releasing stronger versions, widening the gap to 18+ months ahead of the open frontier.

crowd 55
The quiet paththe one nobody’s watching

A large-scale training accident or dataset breach in 2026-27 scrambles both closed and open capability roadmaps.

6%
The read
Open-Weights Catches Up and Closed Frontier Pulls Ahead are near balance.

Neither path has decisive pull. The levers below are where trajectory bends.

Where reality is heading

A, 80% bandCrowd today
Projected reality support to 31 Dec 2028A rises from 41% to about 52% with an 80% band of 33 to 71. B falls from 53% to about 44%. Quiet path near 4 to 6%. Crowd today: A 55%.100%7550250Crowd today · A 55%A ~52%B ~44%P ~4%NowJan 2027Jan 202831 Dec 2028
Reality crosses toward A in early 2028. Today’s crowd number sits at the top of A’s band, possible but optimistic. View as table

The movers, ranked by leverage

01
China's AI-development investment and the DeepSeek/Qwen/Baidu ecosystemA central driver of open-weights capability
leverage 0
Conditional
02
US / EU open-weights regulatory postureWhether frontier open-weights become legally restricted
leverage 0
Conditional
03
Compute cost curvesIf training-compute cost drops faster than frontier scaling, open-weights maintains
leverage 0
Conditional
04
RL / reasoning training recipesWhether these diffuse quickly or stay proprietary
leverage 0
Conditional

What would change our mind

Signals toward A
  • Open-weights frontier models (Meta Llama series, Mistral, DeepSeek, Qwen, perhaps new entrants) stay within a 12-month capability window of the best closed models on standard benchmarks (MMLU-Pro, HLE, SWE-Bench, GPQA, FrontierMath)
  • The commodity-capability regime consolidates
  • AI value shifts from base-model to agentic / tool-use layer
Signals toward B
  • The closed-weights frontier (OpenAI, Anthropic, Google) pulls ahead, with the open-weights lag growing past 12 months
  • regulatory restriction on open-weight release above a capability threshold, cost of RL + reasoning training beyond open-source budgets, or deliberate frontier-lab moat-building
  • China's AI-development investment and the DeepSeek/Qwen/Baidu ecosystem: movement in the direction that supports Closed Frontier Pulls Ahead

Evidence