Four leaderboards, four denominators: what you buy when you standardize on a coding agent

Each instrument reports in a unit that does not convert into the others. Model-routing platforms count token traffic and developer surveys count what respondents say they use, while enterprise surveys count which foundation models reached production and benchmarks count one model inside one harness with a fixed set of tools, instructions, compute and time limits ^[6]. The devops.com write-up of the Techstrong report is direct about the consequence: none of that evidence, on its own, yields a market-share table for coding agents ^[7]. The report treats that as its starting point rather than a footnote, separating the races before handicapping anyone ^[4].

Run the arithmetic the leaderboards skip. Claude Code was reported used at work by 39% of JetBrains respondents, but 90% of that sample used an agent at least weekly ^[1][8], so Claude Code reaches roughly 43% of the developers who use anything at all ^[15]. Growth reads differently again: Claude Code is about 2.2 times its January figure, Codex about 5.3 times its own, off a far smaller base ^[14]. The same report describes a single organisation running one agent in the IDE, another in the terminal and an internal system for background work ^[9], which is why these percentages should not be read as slices of a fixed pool. Two vendors can both add reported reach without either losing anything.

The sharper tension is between layers. Futurum found OpenAI, Azure OpenAI and Google Gemini widely used as production model providers, with Anthropic's foundation-model share considerably below Claude Code's standing in developer surveys ^[2]. That reconciles cleanly once you notice Claude can arrive through GitHub Copilot or Cursor rather than Anthropic's own front door ^[9], and that model, agent, development environment and enterprise platform are no longer bought from one vendor ^[10]. A standardization decision names one of those layers. The leaderboard cited to justify it usually names a different one.

Then there is what the seats are doing. Individual developer assistance accounts for 47.20% of AI use in the software lifecycle against 5.84% for autonomous end-to-end agents ^[11], a ratio of about eight to one ^[16]. The four modes sum to 84.99%, so roughly 15% of the reported picture sits outside them ^[17]. Across the lifecycle, code generation runs at 40.17% and deployment decisions at 6.20%, a drop of nearly 34 points ^[12][18]. Nearly everything on the market is sold as agentic ^[13], but the observed behaviour is a review-and-reject loop, and caution tightens as an agent nears infrastructure, credentials and production, where the cost of a failure rises with each increment of autonomy ^[3].

So the capability being standardized on is assistance, which is also the least expensive thing to switch. The autonomy that justifies a multi-year commitment is the part almost nobody has put into production yet ^[11]. A buyer who cannot state which unit their own decision is denominated in has quietly adopted a vendor's.