The coding-agent race has moved to the usage meter

The new battleground is a progress bar

The most visible rivalry in AI coding usually happens on leaderboards. A new model arrives, a benchmark moves, and developers spend a few days comparing outputs.

This weekend, the competition looked more ordinary and more consequential. It appeared in the usage meter.

Anthropic extended higher weekly limits for Claude Code. OpenAI temporarily removed a five-hour restriction for Codex users on several paid plans and reset current usage. Both moves gave developers something immediately useful: more time before their work was interrupted or moved into paid credits.

The offers may be temporary. The change they reveal is not. Coding agents are moving from occasional assistants to tools that can stay with a task for hours, inspect a repository, run tests and revise their own work. Once that happens, model quality is only part of the experience. The other part is how long the agent can keep going before the meter intervenes.

What Anthropic changed

On 12 July, Anthropic's official Claude account said it was extending Claude Fable 5 access on all paid plans through 19 July. In the same announcement, it said Claude Code's weekly rate limits would remain 50% higher during that period.

That is an increase to the included allowance, not a promise of unlimited use. Claude's paid plans normally combine a session limit that resets every five hours with weekly limits. Usage across Claude's conversational products and Claude Code can draw from the same allowance, depending on the plan. Customers who reach their included limit may wait for a reset, change plan or continue with separately billed usage credits where available.

Anthropic has already made capacity a public part of its product story. In May, the company said new compute agreements allowed it to double Claude Code's five-hour limits for several paid plans and remove a peak-hours reduction for Pro and Max accounts. The current promotion is narrower and explicitly dated, but it follows the same logic: additional compute is being translated into a more generous user experience.

What OpenAI changed

OpenAI's move came after what one of its product leaders described as an intense 48 hours for Codex and ChatGPT Work.

On 12 July, Thibault Sottiaux said OpenAI was temporarily removing the five-hour usage restriction for Plus, Business and Pro plans. He also said the company had reached six million active users and was applying a usage reset. These figures and actions come from OpenAI; Model Current has not independently audited the user count or individual account resets.

The following day, Sottiaux added more detail. He said inference optimisations for GPT-5.6 Sol should provide roughly 10% more usage. He also said OpenAI had reduced the product's context-size limit from 372,000 to 272,000 tokens after finding that the larger setting caused more usage to be charged than intended, and that experiments involving reasoning effort had been reverted. OpenAI also identified higher-than-intended multi-agent use at demanding reasoning settings and said it was working on a fix.

Those explanations matter because a usage bar is not a simple count of prompts. A long repository, a large context window, multiple agent runs and more intensive reasoning can consume very different amounts of capacity. OpenAI's current rate card reflects that complexity by tying credits to input, cached-input and output tokens.

The temporary exception is also moving faster than the documentation. OpenAI's public Codex pricing page still describes a five-hour window and additional weekly limits. That is not necessarily a contradiction: the product executive explicitly called the removal temporary. It does mean users should treat the in-product usage dashboard and the latest company notices as more current than a static comparison table during the promotion.

Why limits are now product strategy

For a conventional chatbot, a limit can feel like a distant ceiling. For a coding agent, it can shape the work itself.

A developer may hand an agent a migration, a difficult debugging session or a test suite that needs several rounds of repair. Interrupting that loop after a fixed window is not the same as delaying one more chat message. It can break concentration, postpone verification and make the user reorganise the task around the vendor's meter.

That gives AI companies several reasons to compete on allowance. More included capacity can attract heavy users, encourage developers to move larger tasks into the product and generate the real-world activity needed to improve it. A temporary promotion can also absorb frustration after a launch surge or after unexpected consumption patterns.

There is a cost on the other side. Long-running agents use expensive inference, and the heaviest users are not evenly distributed. Providers therefore need limits that protect capacity without making the product feel unreliable. The result is a form of competition that looks less like a benchmark launch and more like a mobile-data plan: the headline capability matters, but so do the allowance, reset rules, overage price and small print.

More capacity is not the same as lower cost

The immediate offers are favourable to users, but they should not be mistaken for a settled reduction in the cost of AI coding.

Anthropic's 50% weekly increase has a stated end date. OpenAI has not said how long the five-hour restriction will remain suspended. Neither company has promised that current promotional capacity will become a permanent baseline.

There is also no universal conversion between a percentage of allowance and a quantity of finished work. A short, well-scoped task on a small codebase may be cheap. An open-ended task with a large context, high reasoning effort, browser use and several agents can consume much more. A nominally larger allowance can disappear quickly if the product or model uses more tokens to complete the same job.

For teams, the useful question is therefore not simply, 'Which plan gives me more messages?' It is, 'How much reviewed, working output do we get for the total cost, including interruptions and overages?'

How to compare the two offers without being misled

This is a good week to test Claude Code and Codex, but a poor week to treat a promotional meter as a permanent promise.

A fair comparison starts with the same real task: one that is large enough to require planning, implementation and tests, but clear enough that a human can judge the result. Record how much supervision each agent needs, whether it finishes its own verification, how often it drifts beyond the request and what happens when the included allowance is nearly exhausted.

Then repeat the exercise after the promotions end. The best tool is not necessarily the one that runs longest during a free-capacity window. It is the one whose normal limits, output quality and overage model fit the way the team actually works.

The broader lesson is already visible. AI coding companies are no longer selling only access to a model. They are selling a working relationship with that model. In that relationship, intelligence matters. So do continuity, transparency and the simple ability to finish the job.