CLAUDE RESET ITS WEEKLY LIMITS TO COMPETE WITH OPENAI'S AGI MODEL

Anthropic reset Claude Code's weekly limits today, one day into the week that OpenAI spent telling everyone it may have built artificial general intelligence. Two labs, two very different announcements, one calendar. OpenAI put GPT 6 Astra into limited preview on September 3 and let its president say the AGI word into a microphone. Anthropic answered by moving the meter on how much of Claude you get.

The thesis: when two frontier labs charge the exact same price, the fight stops being about price and becomes a fight about access. Today Anthropic moved on access, and that is a more honest signal of where this race actually is than any benchmark either company published this week.

The Reset Landed Today

The mechanics are the part worth understanding, because the headline number and the number you feel are different. Anthropic set standard weekly limits in Claude Code at 25 percent above its original baseline, permanently. That baseline had been running 50 percent higher since May 13 on a temporary boost. Index the original at 100 and the picture is clean: users had been living at 150, the permanent number is 125, and the reset closes that gap.

Anthropic published both halves. It described the change as permanently raising weekly limits by 25 percent, and it also stated that compared to the boosted level it works out to a 17 percent reduction. Nothing was hidden. What developers reacted to was the ordering, the increase announced first and the decrease second, which is a framing critique rather than an accusation of dishonesty, and those two things deserve to stay separate.

The blast radius is narrower than the noise suggests. It hits Pro, Max, Team and seat based Enterprise plans. Free plans and consumption based Enterprise seats are excluded. It moves weekly limits only, leaving the five hour rolling window untouched, and it applies to Claude Code rather than every Claude surface.

Greg Brockman Said the AGI Word Out Loud

The other side of the week was pure ambition. GPT 6 Astra went to limited preview on September 3, pitched as a generational leap across cybersecurity, professional work, software engineering and science, with tighter task boundaries, better focus and stronger multi step workflows. It reaches ChatGPT Plus, Pro, Business and Enterprise, plus the API and Amazon Web Services. President Greg Brockman said it could eventually be seen as the arrival of AGI. The model went through a formal review with the Trump administration before release.

Now put the two price sheets side by side, because this is the number that explains the whole standoff. GPT 6 Astra launched at $10 per million input tokens and $50 per million output. Claude Fable 5.1 launched at exactly the same figures. Identical, to the dollar, on both sides of the ledger.

Two competing frontier products at the same sticker means the sticker has stopped being the product. What you are actually buying is how much you get to use and under what terms, which is precisely the dial Anthropic just turned.

Nobody Outside Anthropic Knows the Baseline

Here is the thing that should bother a serious buyer more than any percentage. Anthropic has never published an absolute number. Not tokens, not messages, not compute units. The limit exists only as a ratio against a baseline that has never been disclosed.

So both figures float free. Up 25 percent from what. Down 17 percent to what. You cannot audit a ceiling you cannot count, you cannot plan a quarter of team capacity against it, and you cannot compare it against a competitor in any unit that survives a procurement meeting. Anthropic cited sustainability and published no capacity analysis beside it. That may be entirely true. It is also unfalsifiable, and unfalsifiable is a bad property for the thing you are selling.

We have watched this meter move before and said so at the time, when the meter kept resetting on Fable 5. The model is superb. The unit of sale keeps sliding.

Check Your Burn Rate This Week

The counter deserves a fair hearing. Anthropic disclosed the reduction itself rather than letting someone else discover it. A permanent 125 is a more honest promise than a temporary 150 that could evaporate on any given Tuesday, and a lab that runs unsustainable limits until it snaps is a worse partner than one that resets to a level it can actually hold. Predictability is worth something a percentage does not capture. It is also fair to say capacity planning runs on quarters rather than news cycles, so the proximity to Astra may be coincidence.

Coincidence is just not how it reads from outside, when the ceiling moves in the same week a competitor claims general intelligence at an identical price.

So: measure your actual weekly burn this week, while you still remember what the boosted ceiling felt like, and find out whether 125 is a constraint or a number you never touched. Most people never hit it and will feel nothing. The heavy agentic users will feel all of it.

The prediction is that absolute numbers arrive within two quarters. Not because Anthropic wants to publish them, but because enterprise buyers holding two identically priced frontier models will eventually refuse to sign for capacity that nobody will quantify. The first lab to print a real number on the box wins the procurement conversation, and as of today neither one has.