Methodology

How we know what we say

Public data, reproducible pipelines and a decision log that records the comparisons we reject as carefully as the ones we make.

Levels of evidence

E1

Published by the source and re-read in the original document or database.

E2

Recomputed by us from microdata and matching a figure the source itself publishes.

E3

Our own tabulation, reproducible and tested, with no published counterpart to check against.

What “open-weight” means here

A model counts as open-weight only when automated checks found its weight files in a public repository that can be read without signing in, and only from the first documented date of that availability; where the first archived copy of a repository shows a page without the weight files, the date is the day the files appeared. The licence is a separate question: a verified permissive licence, a non-commercial or unverified licence, and gated access with unread terms are kept apart. No model is confirmed as closed-weight merely because its weights could not be found, on OpenRouter or on Hugging Face; the rule for confirming closed weights is being revalidated, and the nine decisions recorded so far are suspended until it is. Openness is never inferred from how a model is served.

Data as of

The report is titled 2026, but every source has its own date. The chapters give each figure’s date beside it; this table puts them in one place.

SourceWhat it measuresPeriod coveredRead or retrievedChapters
Hugging Face Hub counters30-day downloads of the tracked models; downloads of quantised copiesAugust 2024 to October 2026, read on the 1st of each month22 October 2026 (snapshots)01, 03, 06
ATOM Project tracked-models listWhich models are tracked, and their developers’ regionsList at commit d4b170920 October 202601
OpenAlexWorks naming a model family2023 to 202526 October 202601
Epoch AI benchmark runsScores on three testsModels released by 10 November 2026Archive of 10 November 202602
Artificial Analysis, VIALS-AAScientific images interpreted, all three attempts right (permission requested)Models released by 10 November 202611 November 2026 (as displayed)02
Vals AI, Terminal-Bench 4.0Long agentic tasks completed, with cost and time per task, the same agent for every modelModels released by 10 November 2026Leaderboard of 9 November 202602, 04
DeepSWE (Datacurve), via Epoch AIProgramming tasks solved, the same agent for every modelModels released to early September 2026Archive of 7 November 2026; Epoch AI’s review of DeepSWE 1.1, 7 September 202602
LM ArenaHow it labels models (cited, not scored)Snapshots to 15 October 202619 October 202602
OpenRouter rankingsDaily tokens by model version1 January 2025 to 14 October 202615 October 202601, 03
OpenRouter, requests sorted by subjectTokens by subject, the most used modelsWeeks of 21 September to 2 November 20269 November 202603
OpenRouter providers and endpointsMonthly tokens by provider; endpoints’ prices and declared precisionThe month to 9 November 20269 November 2026 (as displayed)06
Cloudflare Radar, Workers AIShares of requests by model and by taskJune 2025 to October 20264 November 202601, 03
Providers’ price pages and GPU list pricesAPI prices per token; GPU rental prices per hourNovember 2026November 202606
SemiAnalysis InferenceXMeasured throughput per GPU on given configurationsConfigurations as published9 November 202606
Menlo Ventures, Ramp AI Index, Vercel AI Gateway, ETR/Futurum, FuturumPublished figures on companies’ use and spending2025 to September 2026Cited as published; dates in chapter 0505
JetBrains and Stack Overflow developer surveysMicrodata on developers’ use of models2025 survey wavesDates in chapters 05 and 0605, 06
Eurostat, isoc_eb_aiEnterprises using AISurveys of 2021 to 202519 October 202605
Developers’ model cards and licencesWhat each named release isReleases to November 20267 November 202602, 06

Figures held back

Some figures exist in the project but are not placed in this edition:

FigureSubjectWhy not shown
5.4Open versus closed models reaching production, by organisation size (developers)Not placed: no web chart was ever built.
6.3Where surveyed organisations host generative AI: LF 2024 and CNCF 2025, side by sideNot placed in this edition: held as an internal placeholder for its side-by-side design (Sprint 09 brief). The datasets state CDLA-Permissive-2.0 in their introductions and on their Zenodo records.
3.4Open-weight share of observed tokens, lower bound (2026-Q3)Internal until the high-impact classification review is completed and the classification-date sensitivity is documented. The chart is held back; chapter 03 cites its third-quarter lower bound in the text, as DEC-076 allows.

Our interest, and how we handle it

Helmcode, which publishes this report, builds and operates private inference infrastructure for open-weight models. It therefore has a commercial interest in the subject of the report, and most directly in what running open models costs. These safeguards apply throughout:

  • Every number is linked to its source and to the file that produces it; the figures are drawn by the same code that writes these pages, so each can be rebuilt.
  • Independent observations (benchmark runs, surveys, public price lists, usage rankings) are kept apart from claims made by companies, which are cited as claims.
  • No Helmcode figure is used in any calculation.
  • Where a result depends on assumptions, the assumptions are shown across their full range and the cases that cut against us are reported alongside the others.
  • Any comparison of dedicated infrastructure with API prices was to be read by a reviewer outside Helmcode before it was published (decision DEC-110). That review has not taken place. On 9 November 2026 the editor lifted the requirement for chapter 06, which publishes the comparison as a modelled scenario, not as an observed cost.

This report has not been peer-reviewed.

Sources and rights

No partner, sponsor or institutional endorsement is implied.

Data: Epoch AI, OpenRouter, Eurostat, JetBrains, LM Arena (CC BY 4.0); Linux Foundation Research and CNCF microdata (CDLA-Permissive-2.0); Hugging Face Hub counters (used with Hugging Face’s permission); Cloudflare Radar AI Insights, Workers AI model data (used with Cloudflare’s permission); OpenAlex (CC0); OpenHands Index (Apache-2.0). Each figure states its own source, base and evidence level. Type: Newsreader, IBM Plex Sans, IBM Plex Sans Condensed, Roboto Mono (SIL OFL 1.1).

  • Epoch AI, Capabilities & benchmarking and Data on AI models (CC BY 4.0).
  • OpenRouter rankings-daily (CC BY 4.0): source line “OpenRouter (openrouter.ai/rankings), as of …” on every figure.
  • Eurostat, isoc_eb_ai (CC BY 4.0).
  • JetBrains Developer Ecosystem Survey 2025 raw data (CC BY 4.0).
  • LM Arena leaderboard dataset (CC BY 4.0).
  • Linux Foundation Research and CNCF microdata on Zenodo: CDLA-Permissive-2.0, as stated in the dataset introductions and on the Zenodo records.
  • Stack Overflow Developer Survey 2025 (Open Database License, ODbL 1.0).
  • Commercial surveys marked all rights reserved are cited as text only; none of their charts is reproduced.

Licences: CC BY 4.0 · CDLA-Permissive-2.0 · CC0 1.0 · Apache-2.0 · MIT · ODbL 1.0.

Reproducibility

Every number on these pages is read from the project's analysis files by the same code that builds the page; the charts are drawn in the browser from the embedded data, which can be inspected under each figure.