The Codex Bleed: Anthropic Just Made Its Third Capacity Move in Five Weeks
Anthropic does not bump rate limits three times in five weeks because it wants to. It bumps them because something on the other side of the market is forcing the move. That something is Codex. The Claude Code subscription is the most profitable developer product Anthropic has ever shipped, and OpenAI just spent a month showing the world how to bleed it.
On May 13, the @ClaudeDevs account quietly posted that weekly Claude Code limits would be lifted by 50 percent for every Pro, Max, Team, and seat-based Enterprise user through July 13, 2026. The Anthropic comms machine did not put a CEO on stage for it. There was no blog post. No marketing reel. Seven days earlier, on May 6, Anthropic had already doubled hourly limits. Roughly a month before that, peak-hour throttling was removed for Pro and Max users entirely. Three live interventions on the same product surface in 35 calendar days.
The next day, May 14, Anthropic did something stranger. It quietly re-enabled paid Claude subscriptions inside third-party agent harnesses, the OpenClaw-style wrappers that route Claude tokens through their own scaffolding. But it did so behind a new, separate monthly credit meter. Subscribers can use Claude inside someone else's harness again, just not for free, and not as much as they want. That is not a feature announcement. That is a metering wall going up on the path Anthropic shut six weeks earlier.
The unlimited agent subscription is cracking. And the reason is the same reason every time you look at the data.
What the timeline actually says
Read in order, the moves tell you what Anthropic was reacting to before the press tells you. I am not guessing here. The releases are public, the dates are public, and the framing on each one is defensive.
| Date | Move | Read |
|---|---|---|
| Early April | Peak-hour throttling removed for Pro and Max | Friction reduction, not a capacity bump |
| May 6 | Hourly limits doubled across Claude Code tiers | First real capacity intervention |
| May 13 | Weekly limits +50 percent through July 13 | Time-boxed bump, not a permanent change |
| May 14 | Third-party harnesses re-allowed behind a separate credit meter | Concession on policy, but with a price tag bolted on |
The May 13 bump is the giveaway. A permanent capacity increase is a roadmap item. A time-boxed bump with an explicit July 13 sunset is a tourniquet. Anthropic is buying itself eight weeks. Whatever the strategy team is doing about Codex, they are doing it on a clock that ends mid-summer.
The token-efficiency gap is the whole story
The fundamental asymmetry sits at the workload level, not the price level. Multiple independent benchmark passes (and a useful Northflank writeup from this week) put Codex token consumption at roughly one quarter of Claude Code's for equivalent coding tasks. Call it 4.2 times fewer tokens on the workloads developers actually run.
Now layer that onto an unlimited-tier subscription. If a Max 20x user runs an agent loop that costs $400 in raw Claude inference for the month, Anthropic eats that delta against the $200 sticker. The same workload on Codex, by token count, costs OpenAI roughly $95. The two companies are not running the same business on the same tier even when they charge the same price. They are running fundamentally different margin curves, and the one with the wider gap between revenue and inference cost is the one that can keep raising limits.
| Tier | Sticker | Inference cost on hot workload | Gross margin signal |
|---|---|---|---|
| Claude Code Pro | $20 | $25 to $40 typical | Negative on active users |
| Claude Code Max 5x | $100 | $120 to $200 typical | Negative on heavy users |
| Claude Code Max 20x | $200 | $280 to $450 typical | Negative on power users |
| Codex Pro (comp) | $200 | $60 to $110 typical | Positive on most workloads |
The numbers in the right column are not Anthropic financials. They are workload estimates from running both harnesses against the same task surface. But the direction is what matters. Codex eats fewer tokens, which means the same flat subscription supports more usage, which means OpenAI can dangle more capacity before the math turns red.
Why Sam Altman gave away two months for free
Inside that asymmetry, Sam Altman's May offer (two months of free Codex usage for every new business customer) is not generosity. It is a calculated raid. Free months on Codex are cheap to deliver because Codex is cheap to deliver. Free months on Claude Code would have triggered an investor meeting.
OpenAI is using its cost advantage to do the most aggressive thing a competitor with better unit economics can do: it is buying churn. Every developer who tries Codex on the free window and decides to stay is a Claude Max subscription that never renews. Anthropic's only counter, given the cost structure, is to raise the perceived value of Claude Code on the existing price. Raise the limits. Concede ground on third-party harnesses. Buy back attention with capacity.
The $900B sitting on top of all this
The pricing context is important, and worth saying out loud. Anthropic is, at the same moment it is rate-limiting its own developer plan three times in five weeks, locking in terms for a $30 billion fundraise at a roughly $900 billion post-money valuation. Annualized revenue is on track to top $45 billion. The numbers Anthropic shows fundraisers are real, large, and growing faster than any private software company in history.
But the strain is sitting on a single product line. The agent subscription, sold at a flat monthly price into a workload that scales with developer ambition, is the piece of the Anthropic business model that does not benefit from the same superlinear curve as enterprise tokens. Enterprise customers pay metered. Pro and Max users pay flat. Codex is structurally cheap, which makes it the natural competitor to a flat-priced product whose underlying cost has no flat curve to match.
A $900 billion valuation does not save you from a $20 subscription bleeding $30 of inference. It just means you can afford to bleed for a while.
What developers should actually do
If you run agent loops for a living, three practical things matter this quarter.
One: do not plan around the unlimited promise. The July 13 sunset on the 50 percent weekly bump is the most honest statement anyone in this market has made about unlimited tiers in 2026. Treat it as a soft preview of where things land. Capacity floors are going up because providers are reconciling unit economics, and your workloads should be priced like they will be metered by end of year.
Two: try the harness you do not pay for. Codex is two months free for new business accounts. Claude Code on third-party harnesses now has a separate credit meter, but for a lot of teams the meter is generous enough to A/B test the workflows you actually run. Independent harness comparisons matter more this quarter than they have at any point in the agent era. Our harness leaderboard tracks the surface and is updated weekly.
Three: instrument your token bill. If you are not logging tokens per task on both harnesses, you are negotiating in the dark. The 4.2x asymmetry I quoted above does not hold on every workload. On some agent loops it collapses to 1.6x. On others it opens to 6x. The right number for your codebase, your test suite, and your coding style is something you measure, not something you read off a press release.
Our take
The cold war between Anthropic and OpenAI on developer agents is the most consequential pricing dynamic in AI right now, and most of it is happening in changelog entries and X posts. Three Claude Code interventions in 35 days is a signal. The third-party harness concession with a metered cap is a signal. The two-month free Codex offer is a signal. The $900B raise running in parallel is the context that makes the signals legible.
The unlimited developer subscription was a creature of the 2024 to 2025 AI discounting era, when every provider was racing inference cost to the floor and flat-priced consumer pricing could mostly absorb the loss. That era ended somewhere in the last six weeks, and the May 13 sunset date is the closest thing to a tombstone we have. The next subscription generation will be metered, capped, credit-windowed, or some combination of all three. Codex is what made it unavoidable.
We will be watching three numbers between now and July 13. First, whether Anthropic extends the weekly bump or lets it expire. Second, whether OpenAI broadens the free Codex offer to existing customers. Third, whether the third-party harness credit meter on Claude paid plans gets generous enough to use, or stays restrictive enough to defend the first-party harness. The order in which those three resolve will tell us what the agent subscription actually costs to deliver in 2026, and which lab is willing to admit it first.