Which answer did voters prefer?
Arena's ranking method aggregates head-to-head votes. Its results reflect the prompts, model versions and voters represented in those comparisons. We do not reproduce Arena's rankings here.
Tool 09 · A companion to the IC 100 2026 edition
Who builds the frontier and who brings it into use?
Twenty-five organizations, three distinct roles, and an inspectable evidence trail. This editorial set looks at model builders, scale enablers and AI-centered deployers. It does not compress them into one rank.
12builders
05enablers
08deployers
25source-led profiles
01 / THE FRONTIER MAP
Three kinds of contribution.
Each organization occupies one primary role. Open a profile for its dated evidence, source limits and shareable cards.
Cutoff September 27, 2026
Edition fixed 2026 selection
Loading the frozen edition.
02 / COMPARABLE LENS
One measure, twelve models.
Epoch AI's Capabilities Index estimates model capability with uncertainty. This view is limited to one selected model for each builder. It does not score an organization, product adoption or financial success.
Bars show Epoch's reported estimate and interval from the September 27, 2026 extract. We selected twelve rows and redrew them on a 135–175 visual scale; the values were not changed. The index combines evaluations; it is not a direct measurement of a single task. Intervals overlap for several models. Compare the same snapshot only. Epoch source and methodology ↗
03 / READ THE MEASURES
A model rank is one kind of evidence.
Different measures answer different questions. None by itself establishes a firm's reach or economic advantage.
How did a model perform across tests?
Epoch's Capabilities Index combines benchmark results and reports uncertainty. The chart above is a dated comparison of selected models, not a score for their companies.
Who turns capability into use?
The IC25 profiles follow documented products, infrastructure and distribution. Customer claims and usage figures retain their original definitions. No single rank captures these different roles.
This distinction runs through Intelligence Capitalism: access to capable models is only one part of building a firm around them.
04 / DOCUMENTED CONNECTIONS
Follow the hand-offs.
These dated relationships show how a builder, distribution provider or deployer can connect. They do not imply ownership or exclusive supply.
OpenAI → Microsoft Azure
OpenAI named Azure as an API distribution channel for GPT-6 Astra.
Release evidence ↗Anthropic → Microsoft Azure
Microsoft documented general availability of Claude in Microsoft Foundry.
Hosting evidence ↗CoreWeave → Perplexity
CoreWeave announced an agreement to power Perplexity inference workloads.
Agreement evidence ↗05 / EDITORIAL METHOD
The selection is open for inspection.
Seats and cutoff
The fixed 2026 edition allocates 12 seats to frontier model builders, five to scale enablers and eight to AI-native deployers. The evidence cutoff is September 27, 2026. The builder model must be released within the preceding 18 months. Each organization has one primary role and at least two linked source points; builders require an independent capability observation. Source dates are shown when confirmed.
These are editorial seats, not a top-25 performance ranking. The published edition is fixed at its cutoff; later evidence belongs in a later edition or a dated update.
What the evidence can show
Epoch's licensed CC BY data support model-level comparisons among the selected builders. Company releases document products, customer relationships and reported adoption. ARR, recognized revenue, cumulative users, monthly users and device distribution are different measures. No common growth or productivity rank is calculated where comparable inputs are absent.
Read alongside IC 100
The IC 100 is a fixed cohort of large public companies read through annual filings and reported financial figures. Frontier IC25 selects organizations for documented roles in AI development and use, including private organizations. Where an organization also appears in IC 100, its profile links back to the filing view.
The two tools answer different questions. A model capability estimate is never treated as a company investment or filing-language measure.
Inspect the wider candidate review
The discovery register contains 55 names: the 25 selected profiles above and 30 watchlist candidates. The watchlist has not completed the same two-source review, so it is not a rejected or ranked list.
Builder watchlist: Amazon, 01.AI, Baichuan, Cohere, Databricks, Microsoft Research, Stability AI, Technology Innovation Institute.
Enabler watchlist: AMD, Amazon Web Services, Google Cloud, Oracle Cloud, Cerebras, Groq, Samsung Electronics, ASML, SK Hynix, Huawei Ascend.
Deployer watchlist: Cursor, Canva, Jasper, Notion, Replit, Cognition, Sierra, Character.AI, Runway, Midjourney, Writer, Hebbia.
Cursor is a specific parent-seat close call: Cursor and xAI both joined SpaceX in 2026. They do not occupy separate seats in this edition.
09 / Frontier IC25 · fixed 2026 edition
Return to IC 100