Frontier Model Efficiency Gains
Capability per unit of training and inference compute keeps improving, letting newer models match prior frontier performance far more cheaply and gradually loosening the link between raw scale and capability.
strengthening · confidence 100 · +8 7d · +5 30d · Medium term (3-9 months) · tracking since June 28, 2026 · updated October 12, 2026
Score history
Daily conviction score, 0 to 100. Higher means the thesis is more strongly corroborated.
Now 100
Showing the last few days. Unlock full score history.
Why the conviction moved
- Oct 12Strengthened +1
Several small-model papers, including Kairos's linear-attention hybrid and integer-arithmetic inference formats, aim at getting more capability out of limited training and inference compute. They are incremental, research-stage evidence that efficiency gains are loosening the link between scale and capability.
- Oct 12Strengthened +3
Anthropic reports that Haiku 5.5 scores 72.4% on OSWorld while costing one tenth of its predecessor. A small-tier model now delivers agentic computer-use performance near the frontier, more evidence that capability per unit of serving cost keeps improving.
- Oct 11Strengthened +3
Haiku 5.5's OSWorld score rose from 15.7% to 72.4% while its price fell 10x and its context grew to one million tokens. A small-tier model now does computer-use work that was beyond the previous generation, at a fraction of the serving cost.
Showing the last 2 days. Unlock the full record.
Source trail
Supporting · October 12, 2026
Anthropic prices Claude Haiku 5.5 at one tenth of its predecessor and gives it a one-million-token context window
Anthropic reports that Haiku 5.5 scores 72.4% on OSWorld while costing one tenth of its predecessor. A small-tier model now delivers agentic computer-use performance near the frontier, more evidence that capability per unit of serving cost keeps improving.
Anthropic NewsSupporting · October 12, 2026
Small-model papers test linear attention, integer number formats and new sequence models for compute-limited training
Several small-model papers, including Kairos's linear-attention hybrid and integer-arithmetic inference formats, aim at getting more capability out of limited training and inference compute. They are incremental, research-stage evidence that efficiency gains are loosening the link between scale and capability.
arXiv cs.LG (Kairos)
Unlock full source trail, score history, and daily updates.
83 more sources in the full trail.
Unlock TrendsAffected regions & assets
Townsquare
Argue the thesis in Townsquare.