LLM landscape
LLM Insights
Compare model capabilities, understand costs, and see which models appear in real enterprise deployments.
Show metrics
- LLM cases
- 1,903
- Tracked releases
- 420
- AI labs
- 5
Data coverage and freshness
Refresh timestamps mark successful pipeline runs. Capability coverage follows Epoch ECI, whose publication lag is not fixed, so new releases can reach the timeline before they receive a score. Each analysis includes its source coverage and assumptions.
Highlights
Training compute vs capability
Compare published training compute with model capability; estimates and missing data are identified.
Cost vs capability
What frontier capability costs to train — and which models deliver the most capability per dollar, traced by a Pareto efficiency frontier.
Capability per inference cost
Compare model capability and running cost using published API prices.
Coverage gap: GPT-Rosalind is tracked as released but is not in Epoch ECI yet; no fixed publication lag. New releases enter this chart after Epoch publishes an ECI score.
Open-model hardware requirements
What it takes to run downloadable models across local memory, single GPUs, multi-GPU servers, quantization formats, and workload profiles.
Compression and efficient inference
A source-backed engineering guide to quantization, pruning, MoE, storage-backed execution, attention, KV-cache optimization, and speculative decoding.
Benchmark performance
How models score on individual public evaluations — GPQA Diamond, Humanity's Last Exam, SciCode, AA-LCR and more — ranked as a per-benchmark bar chart.
Coverage is partial: 1 of 10 tracked evaluations has scores; 9 await published leaderboard scores.
Enterprise LLM adoption
Which LLM ecosystems show up in real, source-linked enterprise deployments — quarterly shares for Azure OpenAI, Bedrock, Gemini, Claude, and open weights.
Google Gemini & Vertex AI leads on discovery rate with 10 cases per 100 exploration runs in 2026-Q2.
Model release timeline
When and how often the large AI labs ship new models — release cadence per lab, kept current by the weekly catalogue refresh and the announcement-watcher agent.