Anthropic Cut Claude Code Weekly Limits 17 Percent Today. The Harness Just Got Capped Four Days After OpenAI Uncapped Theirs.

At 04:00 UTC this morning, Anthropic rolled off the 50 percent temporary boost it added to Claude Code weekly limits on May 13 and replaced it with a permanent 25 percent lift over the pre-May baseline. Anthropic scheduled the change on August 28. On paper, it is a permanent increase. Against the capacity a paying customer had at 03:59 UTC this morning, it is a 17 percent reduction on Pro, Max, Team, and seat-based Enterprise.

The August 28 announcement led with the plus 25. The community read the math on the same thread, appended a Community Note, and forced Anthropic to delete the post and repost a clarification that concedes the sentence "compared to today, this works out to a 17 percent reduction in weekly limits on Claude Code" in plain text. That is a rare artifact from a frontier lab. It is also the price of the last four days, because the cut lands inside a window TensorFeed has been tracking as the harness-is-the-product thesis, and the tell is which side of the ledger Anthropic just moved.

The Math

One baseline, three states, one plain read.

Anthropic can honestly say the September 14 cap is 25 percent above what a Pro seat could spend on May 12. A Pro seat that started paying on May 14, or scaled its Claude Code workload on the 150 baseline over the summer, sees a 17 percent cut. Which of those two audiences drives the harness thesis is not a difficult question. The developer running Claude Code every day since the May boost is the exact person the Fermat receipt, the Navier-Stokes scale delta, and the Prove2Me piece were about.

The Same Week, Two Directions

Line the last nine days of harness pricing up on one page.

Four rows below the receipt row point one direction, the fifth points the other. Two of the four are competitors and one is Anthropic's own token-line move on Fable 5.1. The pattern is real, and Anthropic is now on both sides of it inside two weeks.

The reconcile is not hard, and it is worth naming plainly. Anthropic sells Claude Code two ways. On the seat, a fixed monthly fee that has to bound worst-case usage or the margin collapses. On the API, a metered token line that scales with load. The seat is where the harness thesis meets the capacity wall first, because the seat is the product a developer opens in the morning to write code and the API is the product a platform calls in the background to route tokens. The September 1 cache-read cut moved the API line down. The September 14 cap moved the seat line up. Both are consistent with a lab that has decided the seat cannot absorb another quarter of the same load curve on the same monthly price.

What This Says About Capacity

The read the launch post did not want on the record: Claude Code usage has grown faster than the seat price can carry. The 50 percent boost added on May 13 was a bet Anthropic could hold the top of the load distribution on the existing subscription tiers without eating margin. Four months later the boost is gone, the ceiling on power users lands 17 percent below yesterday, and the messaging thread ate a Community Note before the change landed. That sequence is the shape of a company recalibrating on real numbers, not on a plan.

Read against the $200 billion Google TPU commitment, the September 14 cut is a footnote on a calendar. That capacity arrives in 2027, and the seat has to survive four more quarters against a load curve the token line is designed to catch. Anthropic can loosen the API and tighten the seat at the same time without contradiction, and this week that is exactly what happened.

What It Does to the Claude Code Audience

The Claude Code developer audience is the audience Anthropic has been pitching hardest in public for the last month. It is the audience the Fermat receipt lives inside, the audience the Claude Science posts are written for, and the audience OpenAI just aimed the Agents API at with a partner sandbox list of nine and no session premium. The 17 percent cut asks that audience to run smaller sessions, ship fewer subagents per day, or move to the API and pay per token. All three options are available. Two of them are the ones OpenAI wants the developer to pick.

The counter-read Anthropic is entitled to: the Max tier and seat-based Enterprise absorb the top of the distribution, the 25 percent above May baseline is genuine headroom for the median developer who signed up after May 13, and the whole quibble is a window effect from a promotional boost that was never permanent. All correct on the numbers, all consistent with the observation that the median developer is not the developer running the loop the harness thesis is about. The workflow that generates a six-billion-token Fermat run is not the workflow that fits inside a Pro seat at any multiplier.

The Adjacent Reads

Two pieces already on the page make the cut readable without a full recap. The Agents API writeup argued OpenAI decided the runtime is a distribution channel for tokens, priced at zero on the harness line, offloaded to nine partner sandboxes on the compute line. The Fugu Max writeup argued Sakana decided the harness is a paid product priced at $2 input and $6 output. Both moves shrink the wedge between a session and a token. Anthropic's September 14 cut runs the wedge the other way on the seat. A Claude Code Pro seat is now cheaper per month than a full-tilt Codex session on a partner sandbox, and it carries a lower ceiling than it did yesterday. The developer picking a coding harness in Q4 is doing so against a three-way pricing sheet, not a two-way one, and the seat side just moved.

Our Take

The interesting fact is not the 17 percent. It is that Anthropic scheduled the cut on August 28, four days before the Fable 5.1 cache-read cut, seven days before the Prove2Me post, thirteen days before Agents API, and it landed today anyway. Nothing that happened in the last two weeks changed the calendar. Read forward, that tells you the seat economics are not a knob Anthropic can turn against a competitor's pricing move. Read backward, it tells you the harness-is-the-product signal was a message about where the value lives on the API, not a promise about where the ceiling sits on the seat.

Both reads matter for anyone building a coding product on someone else's API tier. If your product is a wrapper on Claude Code, the top of your addressable load just moved down. If your product is an API-native agent runtime paying per token, nothing about today changes your unit economics, and the September 1 cache-read cut is still the piece of Anthropic's pricing sheet that decides your Q4. Which side of that line your product sits on is the question this week put on the pricing page.

The messaging misfire on August 28 is worth one line of its own. A frontier lab whose public thesis for four weeks has been that the coding harness is the product line published an announcement that led with a plus 25 and buried a minus 17, then deleted the thread when developers ran the math in public. The developers ran the math correctly on the first pass, and the correction shipped in under 24 hours. That is a functioning feedback loop. It is also a data point on how much slack the Claude Code audience has left for optical framing on the tier that carries the load they generate.

Three Signposts for the Next 60 Days

Whether OpenAI or Sakana publishes a coding-workload benchmark specifically against Claude Code sessions with the September 14 caps priced in (the direct test of whether the Agents API and Fugu positioning treats today as a marketing opening or a technical one). Whether Anthropic ships a metered Claude Code tier that lets a power user pay past the weekly cap on the API price sheet (the direct test of whether the seat is a permanent bound or a promotional entry point to the metered line). Whether the next Claude Code weekly-limit change lands with an announcement that leads with the net change against current capacity (the direct test of whether the Community Note on August 28 changed the messaging playbook or only got one thread deleted). Two of the three fire and the seat side of the harness pricing sheet gets rewritten inside a quarter, one way or the other.

Adrian Vale, September 14, 2026. TensorFeed tracks AI model releases, provider status, and the pricing layer underneath them. Sources for this piece: Anthropic's deleted and reposted August 28 announcement thread through ClaudeDevs on X, BleepingComputer and Implicator coverage on September 12 and 13, and the Fable 5.1, Agents API, and Fugu Max receipts referenced above.