Everyone printed the same five claims this week. I went to the source on each. Three stood up, one is real but the number attached to it is not, and one I could not check. I will show you which is which, because that is the point.
In today's issue
Meta is not paying you. It is discounting you 95%
Anthropic's shopping agents are real. The 35% is not in them
Your Claude Code week drops 17% on the 14th, and the switch nobody mentioned
Fable 5.1 is top of one index and twice as talkative as the median
Google's weather model, and the scorecard
Meta is not paying you. It is discounting you 95%.
The headline everywhere was that Meta will pay you to use its new model. Read the source and the deal inverts. Meta added a contributor tier to Muse Spark, its model for coding agents. You get cheap tokens. Meta gets the right to train on what you send.
TechCrunch: “While 1 million input tokens under a standard agreement costs $1.25, under the contributor pricing model they cost just 10 cents. For output tokens, the standard price is $4.25 per million, but that same million costs just 20 cents.”
The trade, in their words: a discount “averaging out to about 95% for users who ‘contribute’ to the development of future models by sharing their prompts and model outputs.”
Meta's framing is that it “lowers the barrier to entry for prototyping, testing integrations, and scaling experiments where training on your data is acceptable.” Meta did not explain its rationale when TechCrunch asked.
What it means for you. Use it for prototypes, throwaway scripts and anything you would publish anyway. Keep client work, NDA work and the prompt chain that is your product on a paid tier. A 95% discount on a bill you can measure, against a cost you cannot, is a trade to make on purpose rather than by default.
Anthropic's shopping agents are real. The 35% is not in them.
One newsletter ran this two days running: Anthropic open-sources Claude shopping agents that lift cart sizes 35%. I followed the link.
The repository exists. anthropics/commerce-agents, Apache 2.0, about 2,000 stars, last pushed 1 September. Its own description: “Reference blueprint for building shopping and merchant agents with Claude. Examples in retail, commerce, telecom, and entertainment included.”
There is no 35% in it. No cart-size figure, no conversion figure, no basket metric anywhere in the README. The numbers it does contain are port numbers and Python versions.
I do not know where the 35% came from. It is not on Anthropic's newsroom or their engineering blog either. It may be a launch tweet or a partner claim. It is not in the thing you would download.
Worth more than the number: Shopify has already published working implementations of the blueprint on Shopify, pushed 2 September.
What it means for you. If you sell anything online, this is the repo to clone this week. Two agents, a shopping one that searches, compares, fills the cart and remembers the customer, and a merchant one that prices, promotes and drafts campaigns. Judge it on what it does in your store, not on a number nobody can source.
Your Claude Code week drops 17% on the 14th
Both numbers doing the rounds are true, which is why you saw both. Anthropic raised the permanent weekly limit 25%, and the temporary boost running since May ends the same day.
Baseline 100. The May boost made it 150. From 14 September the permanent limit is 125.
Anthropic's own words: “Compared to today, this works out to a 17% reduction in weekly limits on Claude Code.” BleepingComputer reports the original thread was deleted and reposted with that admission.
The only reason given: “Thanks for hanging with us while we figured out what we can sustainably serve going forward.” No capacity analysis was published, so I am not going to invent one.
Pro, Max, Team and seat-based Enterprise all move.
The part nobody covered: Anthropic's pricing page says you can “turn on usage credits to keep working at standard API rates” once you hit the cap.
What it means for you. Check that setting before the 14th. With usage credits on, the cap stops being a wall and starts being a bill. Decide which of those you would rather meet on a Thursday afternoon.
Fable 5.1 is top of one index and twice as talkative
Anthropic's Fable 5.1 shipped this week. Artificial Analysis has it at #1 of 202 on their Intelligence Index, score 57 against a median of 29.
Cost per index task: $6.12, at $10 per million input and $50 per million output.
It wrote 160 million output tokens across the evaluation, which they call “very verbose in comparison to the median of 79M.” Roughly double.
The line I could not place is 73.4% on Cursor's coding benchmark. Neither Cursor's blog nor their changelog mentions Fable 5.1. Their latest entry, 2 September, is self-hosted machines for cloud agents.
What it means for you. Smartest model on the board, and it gets there by writing twice as much. If you pay per token, that is the whole cost story. Short edits will feel cheap. Long agentic runs will not. Watch a week of your own bill before you plan around either.
Google's weather model is five times sharper
This one checked out clean. Google DeepMind, 3 September: WeatherNext 3 gives “a global weather picture roughly five times sharper than our previous model,” moving from a 25-kilometre grid at 6-hour steps to 5-kilometre hourly forecasts, and “up to 50% more accurate precipitation forecasts” a day or more ahead.
The scorecard
Stood up at the source: Meta's 95%. The 17% Claude Code cut. WeatherNext 3's fivefold claim.
Real thing, wrong number: Anthropic's commerce agents. The repo is there. The 35% is not.
Could not place: Fable 5.1's 73.4% on CursorBench. Not on Cursor's blog or changelog.
Could not check: Qwen's million-token context and its 27B model. Their site renders through JavaScript and serves a plain reader nothing. OpenAI's GPT-6 Astra claim: their newsroom returns a 403 to me. Both may be exactly right.
A roundup that never says which of its lines it verified is asking you to trust the newsletter instead of the source.
Also
Cursor, now part of SpaceX since 14 August, shipped self-hosted machines for its cloud agents on 2 September.
Your move
Which of these do you want me to run properly next week, on real work, with the bill printed? Hit reply and name one. Replying to this email is the whole mechanism. Most-asked wins.
Mike
Automate & Hustle. The AI week in five minutes, checked against the source. When I cannot verify something I say so. When I get it wrong I print that too.