Language model provider shares in 2026: why there is no single answer

"Language model market share" is not one quantity but at least four, and they rank differently:

OpenAI leads on visits, Google on crawling, Anthropic among developers. None of these answers is wrong, and none of them answers "who is bigger" in general.

What follows is each quantity with its limits.

By Similarweb panel data covering June 2025 to May 2026, total visits to generative service websites run to 9.5 billion a month, up 70% year on year. Within that volume the ordering shifted sharply:

An important detail: ChatGPT’s share fell while its absolute visit count stayed roughly flat. The market grew around it rather than at its expense. Presenting "a fall from 76% to 53%" as user attrition is an error.

What this metric does not cover, by the source’s own statement: API usage, desktop applications and AI features embedded in other products. That is to say, all enterprise load. For judging where money goes, this quantity is of little use.

By AI crawler traffic for 20–26 August 2026 the leader is Googlebot at 24.8%, then Meta-ExternalAgent 14.7%, ClaudeBot 10.4%, Amazonbot 8.2%, GPTBot 7.4%. Here OpenAI sits fifth.

This is a share of AI crawler traffic rather than of the whole web, and it speaks to who collects data most actively — a quantity only loosely related to user counts.

Among developers the ordering is a third one. Per a JetBrains survey fielded from May to July 2026 among more than fifteen thousand professional developers with reweighted results: Claude Code 39% (47% in the US), GitHub Copilot 21%, Codex 16%, Cursor 12%.

A year earlier Copilot led with 29%. So this segment turned over within a year, and it matches neither the visit figures nor the crawling ones.

A caveat on the survey: JetBrains’ full 2026 report is not published and the figures appear piecemeal on its blog. And it is self-selected, albeit reweighted for composition.

On 27 August 2026 we queried client library download counters directly from the registries:

The same pair of providers, and the ratio differs twofold. In JavaScript the clients are essentially level; in Python one trails the other by half.

The usual caveats apply: downloads are registry requests dominated by build pipelines rather than people, and a week and a month are not directly comparable. But within one ecosystem the ratio between two comparable packages is measured correctly, and that ratio is what carries meaning here.

Nobody publishes market share by revenue. All the providers but one are private and file no audited accounts. Figures of the form "annual revenue of N billion" are ARR — a short recent period multiplied out to a year, not money earned.

Nobody publishes share by API call volume either. And that is the quantity closest to "market share" in the ordinary sense. Only the providers themselves know it.

No Russian cut exists at all — for any of the four quantities. What is known first-hand about running models in the Russian web we collected in our operational report.

The practical conclusion: before accepting somebody’s figure, ask what exactly it counts. Website visits, crawler traffic, a survey answer and a package download are four different things, and they get passed off for one another routinely.

Sources: Cloudflare Radar AI insights, the JetBrains 2026 survey, the npm counter, PyPI statistics.