Anthropic cuts Claude Code capacity 17% as Astra lands; memory squeezes laptops

Marcus here.

Today's three picks all end with someone reading a meter.

Anthropic's 50% Claude Code boost expires September 13. The permanent 25% raise that replaces it on September 14 leaves heavy users about 17% below the current ceiling, in the same week GPT-6 Astra arrives.

TrendForce puts the CPU, DRAM and SSD at 68% of a benchmark notebook's parts bill this quarter, up from 45% in early 2025. Nobody has paid that yet.

And roughly 18,000 posts by self-identified OpenAI agents turned up on a dormant German wiki that had been edited 20 times in the previous decade.

Stay curious,

Marcus Schuler

Anthropic's temporary 50% increase to Claude Code's weekly allowance ends September 13, and the permanent 25% raise replacing it on September 14 leaves eligible subscribers about 17% below this week's ceiling.

The 25% is measured against the pre-May baseline, not against the promotion now expiring. Five-hour session limits do not change.

GPT-6 Astra landed in the same week. An independent experimenter, CozyBlaze, ran it through every chamber of Valve's Portal with no human input, 3,336 tool calls over 23 hours and 43 minutes. On coding Astra sits level with the field rather than ahead, scoring 67 against 70 for Claude Fable 5.1 on a launch-week index.

Why This Matters:

• Teams that sized Claude Code seats against the current allowance will hit the weekly ceiling sooner from September 14 onward.
• Astra gives heavy users a live alternative in the week their ceiling drops, at roughly 75% more per task than Sol.

TrendForce put the CPU, DRAM and SSD at 68% of a benchmark notebook's parts bill in the third quarter, up from 45% in early 2025.

Manufacturers stocked components before the increases, sheltering buyers so far. TrendForce modeled that preserving the first-quarter 2025 gross margin on a comparable machine would take a price rise of roughly 80%, a calculation, not an announced increase. When the cheap inventory runs out, brands decide who absorbs it.

OpenAI publishes a rare internal ledger of how coding agents are changing work inside its own research organization.

As of mid-August 2026, the median OpenAI researcher used more than $600 of inference per day at API prices, and the research organization logged 3.1 agent-workdays for every eight-hour human workday.

• HyperVault, a TCS unit, will invest up to $7.4 billion in a one-gigawatt AI campus on 264 acres in Hyderabad, built in phases against customer demand.
• Authors are disputing publisher and agent claims on Anthropic's $1.5 billion settlement, which pays $3,000 for each of nearly 500,000 pirated titles.
• Uno researchers report up to 3× faster generation while preserving the base model's output distribution, with code and checkpoints public for outside testing.
• Anthropic shipped /skill-doctor in Claude Code 2.1.261 on September 4, reporting which loaded skills a session never invoked and what each costs in context.
• Nvidia released PAIR, free software that pools compatible Windows, macOS and Linux machines to run local inference jobs across a home network.
• Microsoft's David Fowler said typing code is "absolutely over", pointing to Aspire, where agents run services and return changes for a developer to review.

A forecast matters when it changes a decision. This prompt converts one prediction into a dated watch list.

Your raw input:

The prompt:

Why this works: The prompt separates a prediction from the evidence that would move a decision. Dated checks and the ban on invented thresholds turn a broad claim into a plan you can monitor.

What to use: A frontier reasoning model handles long reports best. A standard chat model is enough for a one-page forecast.

Jotform AI Data Assistant lets a team question, filter and chart form-response tables in plain English. It turns responses into a downloadable chart without first exporting them to a spreadsheet and writing formulas. The Starter plan costs $0 a month, and Bronze costs $39 a month or an effective $34 a month with annual billing, checked August 25, 2026.

How to use it:

1. Open Jotform Workspace, hover over the form, click More, then choose Submissions under Data.
2. Type "Show submissions with a pending status" in AI Data Assistant, then press Enter.
3. Ask "Create a bar chart showing submissions by status," then press Enter.
4. Click the Download icon on the generated chart to save it as a PNG file.
5. Ask "Export the filtered submissions as a CSV file," press Enter, then click the Download icon.

Where it falls short: It only works on data inside a Jotform Table. On Starter, form traffic is capped at 100 submissions a month and 500 stored submissions per account, checked August 25, 2026.

Researchers counted roughly 18,000 posts by self-identified OpenAI agents between May and early July 2026, about 17,000 of them edits on DseWiki, a site edited 20 times in the previous decade (collusion.wiki, September 4, 2026).

Our take: For five days in June the moderator deleted about 100 pages a day while the agents created roughly 400. He lost. On September 4 he closed the roughly 2,640-page wiki to open editing, so the encyclopedia he tended for a decade now survives as an exhibit. Nothing was stolen and nobody was harmed, unless you count the one person actually using the thing.

OpenAI's position is that no clear standard exists for reporting behavior like this, which is true and is also the entire problem. A framework is promised in upcoming weeks. Until it lands, the test for whether an agent is staying inside its bounds appears to be whether a volunteer moderator somewhere notices and says so out loud.

Marcus Schuler

San Francisco

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