Polylog
← Trends

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 74 · +38 7d · Short term (next 30 days) · tracking since June 28, 2026 · updated July 7, 2026

Score history

Daily conviction score, 0 to 100. Higher means the thesis is more strongly corroborated.

Jul 6 · 68Jul 7 · 74

Now 74 · +6 since Jul 6 · ranged 68 to 74

Showing the last few days. Unlock full score history.

Why the conviction moved

  • Jul 7
    Strengthened +2

    LingBot-Vision, pretrained for spatial perception, reports beating models roughly seven times larger, an added data point that architecture and training focus can substitute for raw scale. This reinforces improving capability per unit of parameters/compute.

  • Jul 7
    Strengthened +4

    Claude Sonnet 5 delivers Opus-class agent performance at a fraction of the price, tightening the capability-per-dollar curve. This is a fresh data point that a smaller model can match prior frontier capability far more cheaply, loosening the scale-capability link.

  • Jul 6
    Strengthened +5

    Claude Sonnet 5 delivers higher capability (92.4% SWE-bench Verified) at lower cost than prior flagship models, continuing the trend of greater capability at lower price.

  • Jul 6
    Strengthened +6

    Meta's Muse Spark, the first model from Meta Superintelligence Labs, claims frontier capability using more than an order of magnitude less compute than Llama 4 Maverick.

Showing the last 2 days. Unlock the full record.

Source trail

  • Supporting · July 7, 2026

    Anthropic's Claude Sonnet 5 Pushes Agentic Coding Down the Cost Curve

    Claude Sonnet 5 delivers Opus-class agent performance at a fraction of the price, tightening the capability-per-dollar curve. This is a fresh data point that a smaller model can match prior frontier capability far more cheaply, loosening the scale-capability link.

    Anthropic
  • Supporting · July 7, 2026

    LingBot-Vision Claims a Spatial-Perception-Native ViT That Beats Far Larger Models

    LingBot-Vision, pretrained for spatial perception, reports beating models roughly seven times larger, an added data point that architecture and training focus can substitute for raw scale. This reinforces improving capability per unit of parameters/compute.

    AI with Papers (Telegram)

Unlock full source trail, score history, and daily updates.

8 more sources in the full trail.

Unlock Trends

Affected regions & assets

RegionsGlobal