Gavin Baker reads Codex's jump to 25M users as OpenAI taking share

Gavin Baker's current read of the AI market: OpenAI is still taking share, Anthropic is "likely reaccelerating," and open source is growing faster than either. The Atreides Management CIO posted that three-line summary with a single chart attached — Codex active users going from 6 million to 25 million in roughly seven weeks.

Reported Codex active users, February to August 2026; the dotted line marks the July 9 ChatGPT Work launch. Credit: chart shared by Gavin Baker, sourced to OpenAI, Sam Altman and Tibo.

#The curve behind the claim

The chart's labelled milestones run 1m+ (Feb 5), 2m+ (Mar 5), 3m (Apr 8), 4m (Apr 21), 5m (May 31) — then a near-vertical section after the July 9 launch of ChatGPT Work, which folded Codex into the redesigned ChatGPT desktop app: 6m on Jul 12, 7m on Jul 13, 8m on Jul 14, 9m on Jul 15, 10m on Jul 21, about 15m on Aug 9, 20m on Aug 19 and 25m on Aug 29. Codex lead Tibo announced the 25-million mark and OpenAI reset usage quotas for all paid ChatGPT Work and Codex subscribers to mark it — the second such reset in days.

Two caveats travel with the line. The post-July figures are widely tracked as combined weekly actives across Codex and ChatGPT Work rather than Codex alone, and "active users" is a product-usage number, not revenue or token volume. It is evidence that OpenAI's coding surface is growing fast; it is not, by itself, a measurement of share.

#Open source is the faster mover

Baker's third claim has its own dataset. A week earlier he posted Vercel AI Gateway data from Guillermo Rauch showing open-weight models going from 28% to 62% of tokens in two months.

Open- versus closed-weight token share on Vercel's AI Gateway, Jun 24–Aug 22, 2026. Credit: chart from Guillermo Rauch, shared by Gavin Baker.

He called that "super impressive given that the sum of OpenAI and Anthropic accelerated in July," concluding that total token demand accelerated more than the frontier did, and that Grok was probably growing faster still. His read for infrastructure investors: "an open-source token costs just as much compute to produce as a frontier token. Nothing about open-source AI inference is 'free.'" His guess at the end state is that closed frontier tokens end up 60–90% of economic value but only 15–25% of tokens.

#The Anthropic half is a forecast, not a number

"Likely reaccelerating" is deliberately hedged, and Baker's earlier posts show what it is anchored to: a reported $65 billion annualized run rate for Anthropic at the end of July, plus capacity he expects to come online through July, August and September ahead of a possible autumn IPO. His own framing on the All-In podcast was that "on the margin, Anthropic is losing share to OpenAI, to open source, and to Grok, and they're still growing so fast" — and he predicted duelling leaks from both labs as the roadshow approaches.

Third-party data does not settle it, because the available series measure different things. Ramp's August index found that in July 43.5% of US businesses paid for Anthropic subscriptions or tokens, up 1.1 points month over month, versus 39.7% for OpenAI, up 0.23 points; xAI rose 0.94 points to 4%, its fastest gain since July 2025, and open-source model-serving platforms reached 6.1% of AI-using businesses. Inside those accounts, GPT-5.6 Sol was 25% of OpenAI tokens and 23% of OpenAI spend, while Fable 5 was 6% of Anthropic tokens but 11.4% of Anthropic dollars.

#Why it matters

Baker is one of the loudest public bulls on AI infrastructure, and his thesis — that share is sloshing between labs while the total token pie grows faster than anyone's slice shrinks — is the case the market will be asked to price at Anthropic's IPO. Worth holding onto: the post itself defines no market, and each supporting series answers a different question. Product users, gateway tokens and corporate card spend can all move in opposite directions at once, and here they do.