This post argues that the pricing of AI coding assistance has stopped being a simple “so much per user per month” subscription. Drawing on vendors’ public documentation from 2 April to 29 May 2026, it describes a shift toward models that combine seat fees, included usage allowances, credits, token consumption and organization-level budgets and caps, and contends that coding agents in particular have taken on the character of metered API billing.
Its explanation is structural. Autocompletion returns short outputs for short contexts, whereas a coding agent reads a repository, explores several files, runs tests, fixes failures and explains its diff, sometimes calling the model for tens of minutes or hours on one task. A single request from the user’s point of view can therefore consume large and highly variable amounts of input, output and cached tokens, which the author argues makes unlimited flat-rate plans hard to sustain across light and heavy users alike.
The post then compares three products and turns to what organizations should check when adopting them, concluding that coding agents must be managed both as developer tools and as consumers of the organization’s compute — a FinOps concern.
Key Points
- For OpenAI Codex, the post reports that ChatGPT Business distinguishes a fixed-price standard seat from a usage-based Codex seat that draws on workspace credits, and that from 2 April 2026 Codex pricing changed from per-message charges to API-token usage, with credit rates defined per million input, cached-input and output tokens. It also notes that ChatGPT plans do not include API access, which is billed separately.
- For Claude Code, it cites Anthropic’s documentation as stating that usage is billed by API token consumption: Pro and Max plans include an allowance, with standard API rates applying once API credits are used beyond it; Team plans combine seat fees, usage limits and optional usage credits; and Enterprise seat fees cover platform access only, with usage billed separately at standard API rates.
- For GitHub Copilot, it reports that Copilot Business and Enterprise move to usage-based billing from 1 June 2026 in a unit called GitHub AI Credits (one credit being US$0.01), converted from token usage by model. Monthly allowances of 1,900 and 3,900 credits per user respectively are pooled across the organization rather than held per user, and code completion and Next Edit Suggestions are stated not to consume credits.
- Set side by side, the post argues, the three differ in detail but share one direction: SaaS seat pricing for the base product, combined with credit or token metering for heavy use.
- It proposes five questions for enterprise adoption: what a fixed fee actually includes; what happens when a limit is reached (blocking, a cheaper model, or paid continuation); whether allowances are pooled across the organization or held per person; whether usage can be audited, made visible and capped by user, organization or project; and whether API-based use (custom tools, agents in CI, or third-party agents driven by an API key) is budgeted separately from SaaS seats.
- The author reads the move to API-like pricing as a sign of pricing power rather than only of high cost: because coding agents enter the daily work of well-paid engineers, companies keep paying near API-equivalent costs for them. This interpretation is presented as the author’s own, prompted by a third party’s comment that these vendors have found product–market fit.
- It suggests questions organizations will need to answer, such as AI cost per pull request, time saved compared with manual work, which tasks justify expensive models, and how to set per-user caps, and compares the shift to the earlier move from owning servers to metered cloud billing.
Context
The post states that pricing, plan names and included allowances change frequently and tells readers to confirm current terms on each vendor’s official pages; its figures describe the situation as of late May 2026. It explicitly leaves Google’s agent offering out of the comparison as less directly comparable. Its argument concerns enterprise purchasing and budget control rather than individual developers, whose experience of paying for a personal subscription it acknowledges may still feel close to the old flat-rate model.