REPORT · BENCHMARK v0.4 · 2026-06-28

The state of enterprise model performance.

Twelve frontier models scored against 1,400 real enterprise tasks. What holds up in production, what collapses under load, and where the frontier actually is: measured, not claimed.

SIX CHAPTERS · 12 MODELS · 1,400 TASKS · UPDATED 2026-07-03

KEY FINDINGS

41%

of tasks show a different leader in production than on public leaderboards.

3.2×

cost spread between models within two points of each other on quality.

94.2

top composite score on Benchmark v0.4, set by a model half the price of the runner-up.

WHAT'S INSIDE

Six chapters. Start anywhere.

  1. 01

    How we measure

    How business workflows become a fixed, blind-scored test packet, with hard gates on top.

    BENCHMARK
  2. 02

    The dumb zone

    Recency and primacy bias: why the middle of a long context goes dark, and what it costs.

    NOTE
  3. 03

    How agents remember

    Context, a compacted working summary, and the long-term store, and how agents forget.

    NOTE
  4. 04

    Three kinds of RAG

    RAG, Graph RAG, and Agentic RAG: how each retrieves, where each fails, when each wins.

    NOTE
  5. 05

    The frontier, priced

    Every model worth deploying, plotted by measured intelligence against tokens per dollar.

    MODELS
  6. 06

    Big model, small model

    When the cheaper or open model wins your workflow, and how to route between them.

    NOTE

Read it before your vendor does.

Start reading