Skip to content
Bellwether26 October 2026

AI Coding Agents · Issue Nº 9

Codex catches Claude Code in installs, not in talk

Codex has closed the npm-download gap with Claude Code to 5% and is gaining GitHub stars 3.7× faster, yet still trails in developer conversation by 1.6×. The adoption gap is collapsing; the discourse gap is not.

February 2026 to August 2026 (August partial; weekly npm data through Aug 31) · Published 2026-09-09

The install gap collapses

Codex 20.4M, Claude Code 21.5M (ratio 1.05×, Aug 31)

Weekly npm downloads

Codex +103.9%, Claude Code +93.9%

Growth since mid-June

13.7M (Aug 17) → 20.4M (Aug 31), +49%

Codex 14-day sprint

Codex 90.7k → 120.2k (+32.5%); Claude Code 132k → 143.5k (+8.7%)

GitHub stars, full window

Codex +16.2%, Claude Code +2.5%

Star growth, last 28 days

Codex growing 3.7× faster

Star growth ratio

Adoption convergence

On August 31, Codex recorded 20.4 million weekly npm downloads. Claude Code recorded 21.5 million. The ratio between them was 1.05×, the narrowest it has been since either tool appeared on the board.

Both tools roughly doubled their install bases since mid-June. Codex went from 10.0 million to 20.4 million, a gain of 103.9%. Claude Code went from 11.1 million to 21.5 million, a gain of 93.9%. The absolute growth was identical at 10.4 million downloads each. What changed is that Codex started from a lower base and closed the percentage gap.

Codex's sharpest movement came in its final two weeks. Weekly downloads jumped from 13.7 million on August 17 to 20.4 million on August 31, a 49% increase in 14 days. The proximate triggers are visible in the conversation data. On August 25, OpenAI restored five-hour Codex and Work limits for ChatGPT Plus users (103 points on HN). On August 21, a developer posted quick impressions from a week of using Codex more than Claude (59 points). An AWS Bedrock billing bug specific to Codex drew 144 points the same day. The install spike and the discussion spike share a window, but they are not the same shape: installs went up and stayed up; conversation spiked and then fell.

GitHub stars tell the same story at a different velocity. Codex grew from 90.7k to 120.2k stars over the full window, a 32.5% increase. Claude Code grew from 132k to 143.5k, an 8.7% increase. Codex is gaining stars 3.7 times faster. In the last 28 days alone, Codex added stars at 16.2% while Claude Code added at 2.5%. At those rates, Codex would overtake Claude Code's star count in roughly five months, though star growth is not linear and the brief does not project it.

The conversation floor

Developer conversation has not followed adoption. In August, Claude Code appeared in 212 tagged stories across HN and Lobsters, a 6.9% share of all developer-tool conversation. Codex appeared in 132, a 4.3% share. The gap is 1.6×.

That gap has held for seven months. Codex's conversation share has moved between 2.8% and 5.1% since February. Claude Code's has moved between 6.9% and 12.1%. Both are declining: Claude Code from 12.1% in March to 6.9% in August, Codex from a July peak of 5.1% back to 4.3% in August. The two lines are converging, but from the top down rather than from the bottom up.

Codex's August conversation was concentrated in hn_new (129 stories) with only 3 reaching hn_top. The highest-scoring Codex story was a bug report, not a endorsement: the AWS Bedrock billing issue drew 144 points. The highest-scoring Claude Code story was a usage guide: "Maximizing the value of your Claude Code sessions" at 302 points. Developers are installing Codex at scale but writing about Claude Code with more enthusiasm.

The weekly data confirms the pattern. Since June 29, Codex has held the number-two slot among AI coding tools every single week. Its weekly mentions ranged from 27 to 52. Cursor, the third-place tool, ranged from 7 to 20. The closest week was August 17, when Codex had 27 and Cursor had 17. Cursor has not closed the gap in any week. Codex's floor is above Cursor's ceiling.

What the gap means

The divergence between installs and conversation is the story. Developers are adding Codex to their toolchains at nearly the same rate as Claude Code. They are not talking about it at nearly the same rate.

Two explanations fit the data. First, Codex's late-August install surge was triggered by a limit restoration and a billing bug, both operational events that drive downloads without generating enthusiasm. Developers who hit a restored rate limit install the tool and go back to work. They do not write blog posts about it. Second, Claude Code benefits from a longer accumulation of community content: guides, tips, session-optimization stories. That library of practical writing sustains conversation even as the tool's own share of new discussion declines.

The monthly shares for both tools are falling. Claude Code went from 10.9% in February to 6.9% in August. Codex went from 3.8% to 4.3%. The total AI-coding-tool conversation pie is shrinking, or at least not growing as fast as the overall developer discourse. Both leaders are losing share to the long tail of other tools and topics. But within the shrinking pie, Claude Code still commands 1.6× the attention of Codex.

Monthly conversation share, February–August

Feb 10.9% → Mar 12.1% → Apr 11.7% → May 8.7% → Jun 7.9% → Jul 7.7% → Aug 6.9%

Claude Code

Feb 3.8% → Mar 2.9% → Apr 2.8% → May 4.3% → Jun 4.2% → Jul 5.1% → Aug 4.3%

Codex

Feb 1.3% → Mar 2.1% → Apr 1.9% → May 0.9% → Jun 2.5% → Jul 1.3% → Aug 2.1%

Cursor

Feb 1.0% → Mar 1.5% → Apr 1.9% → May 1.7% → Jun 2.0% → Jul 1.2% → Aug 1.0%

GitHub Copilot

The dispatch

The developer economy, by email