IMPLICATOR
On September 14, Anthropic is scheduled to replace its temporary increase in Claude Code’s weekly allowance with a smaller permanent one, leaving eligible subscribers about 17% less capacity than they have this week. For customers who keep coding agents running all day, the meter tightens overnight.
For subscribers, GPT-6 Astra and the lower allowance land in the same week.
Heavy Claude users are testing OpenAI’s new model, while CozyBlaze, an independent experimenter, separately cleared every chamber of Valve’s Portal without touching the controls. The run and work-software demonstrations explain the interest; Anthropic’s lower ceiling makes changing tools cheaper to contemplate right now.
What Changed
- Anthropic's temporary 50% increase to Claude Code weekly limits ends September 13. A permanent 25% increase over the old baseline replaces it on September 14, leaving eligible subscribers about 17% less capacity than they have this week.
- An independent experimenter ran GPT-6 Astra through every chamber of Valve's Portal with no human input, 3,336 tool calls across 23 hours and 43 minutes, for a $571.18 bill at list prices.
- Astra scored 72.6% on OSWorld 2.0, seven points above the preceding generation, while average time per task fell from about 75 minutes to 40.
- On coding Astra is level with the field rather than ahead, scoring 67 against 70 for Claude Fable 5.1 on a launch-week index, and each task still costs about 75% more than Sol at billed rates.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
What the Portal run showed
CozyBlaze finished the Portal experiment on the evening of September 5 in San Francisco after 3,336 tool calls. The session lasted 23 hours and 43 minutes. An edited recording runs just under two hours.
Computer use means the model reads a screen and drives the mouse and keyboard inside software instead of calling an API. Astra received screenshots plus the player’s coordinates and orientation, wrote short JavaScript movement plans, then watched a modified tool execute them while the game clock was briefly unfrozen.
The same apparatus had stopped Claude Opus 5 at Chamber 10 and Claude Fable 5 at Chamber 8. Astra completed the game. The run generated a $571.18 bill at OpenAI’s September list prices, though CozyBlaze said a $200-a-month Codex Pro subscription covered it.
Portal is a forgiving proof. The 2007 game has guides, wikis and videos that may have entered training data, and pausing play removes the need for reflexes. The run tested whether a general model could preserve visual context, position and intent over thousands of steps, not whether it could solve an unseen game.
FREE WEEKDAY MORNING BRIEFING
Follow what the model makers do to your limits.
The Implicator Morning Briefing filters the AI news cycle to the stories worth your attention and explains their consequences. From San Francisco, every weekday at 4:45 a.m. Pacific, 7:45 a.m. Eastern.
About five minutes. No hype. No spam.
Astra at work
Tae Kim, a newsletter writer on vacation, downloaded Blender onto a spare Mac laptop despite never having used it. He asked Astra for a detailed Space Shuttle and had a model he could rotate and zoom within 10 minutes.
Astra also laid out a circuit board in KiCad, built a Power BI dashboard, completed a Form 1040 and moved a Blender scene into Unreal as a walkable environment during launch week.
The September 3 benchmark table points in the same direction. Astra scored 72.6% on OSWorld 2.0, seven points above the preceding generation, while average task time fell from about 75 minutes to 40. On Agents’ Last Exam, it scored 59.3% against 55.5% for Claude Opus 5 while producing 65% fewer output tokens.
Those results are largely OpenAI’s own reporting, and launch demonstrations do not establish reliability in daily work. They show why users accustomed to handing Claude long jobs are trying Astra on the same assignments.
The September 14 arithmetic
Anthropic’s support page, last updated September 1, says eligible Pro, Max, Team and legacy seat-based Enterprise subscribers receive 50% more weekly Claude Code capacity through September 13. Free plans and consumption-based Enterprise seats are excluded. Five-hour limits are unchanged.
An August 29 post promised a permanent 25% increase over the old baseline beginning September 14. That moves the meter from 150% of baseline this week to 125% on September 14, a drop of 25 points, or 16.7% of current capacity. The comparison date changes the description. The announcement uses the pre-promotion baseline. Subscribers living at this week’s meter receive less.
Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.
As of September 7, the permanent increase appeared on no Anthropic-owned page. The support article contained neither “25%” nor “permanent,” and Anthropic publishes plan multipliers rather than token counts.
Theo Browne, the developer behind t3.gg, answered the August announcement with an alternative draft that accepted the arithmetic but described the change from the subscriber’s current meter. Browne’s version led with the expiring promotion.
A September 7 Reddit thread offers self-selected subscriber accounts, not measurements or a survey. One user said Astra handled a large project for more than 20 hours and hit a limit once. Another, running Max 20x and Codex 5x plans, said Anthropic’s cost was becoming hard to justify. A subscriber cutting the other way said a Claude Code update made the five-hour ceiling difficult to hit.
Where Astra falls short
Astra is level with the field on coding rather than ahead. In a launch-week index, it scored 67 against 70 for Fable 5.1, while its general-intelligence score of 61 trailed Fable’s 65.7. Developers describe oversized pull requests for targeted fixes and slightly weaker writing. At September 3 billed rates, each task still costs about 75% more than Sol.
Jakub Pachocki, OpenAI’s chief scientist, supplies another limit from inside the company. Astra shows “some tendency to kind of think less when it’s told that it’s being monitored,” he said, calling that “a worrying trend.” He also said OpenAI and its rivals need to find ways to slow AI development.
CozyBlaze’s next target is Portal 2. After that come custom chambers with no published guide.
Frequently Asked Questions
What changes for Claude Code subscribers on September 14?
The temporary 50% increase to weekly limits ends September 13, and a permanent 25% increase over the old baseline replaces it. That moves the meter from 150% of baseline to 125%, a drop of about 17% against what subscribers have this week. Measured against the pre-promotion baseline, limits genuinely rise.
Has Anthropic documented the permanent 25% increase?
As of September 7 it appeared on no Anthropic-owned page. The support article carrying the promotion contained neither "25%" nor "permanent." The increase was announced in an August 29 post.
What did the Portal run actually demonstrate?
That a general model could hold visual context, position and intent across 3,336 consecutive tool calls. Portal is a 2007 game with guides, wikis and videos that may have entered training data, and pausing play removed any need for reflexes, so the run did not test solving an unseen game.
Where does Astra not lead?
On coding it is level with the field rather than ahead, scoring 67 against 70 for Fable 5.1 on a launch-week index, while its general intelligence score of 61 trails Fable's 65.7. Developers describe oversized pull requests for targeted fixes, and at billed rates each task costs about 75% more than Sol.
Which plans does the limit change affect?
Pro, Max, Team and legacy seat-based Enterprise plans. Free plans and consumption-based Enterprise seats are excluded, and five-hour usage limits are unchanged.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
