GUESSWORK
Guess & Click presents — your weekly click into AI
THE ONE THAT MATTERS
Your Claude Code limit goes up on September 14. You will have less of it.
Claude Code users on Pro, Max, Team and seat-based Enterprise plans are running on a temporary 50% increase to their standard weekly limits. That promotion runs through September 13, 2026.
Anthropic said that starting September 14 it is permanently raising standard weekly limits by 25% for those same plans.
Then the arithmetic.
The permanent 25% is measured against the pre-promotion baseline, not against the boosted number you have been working with. A baseline of 100 is 150 today and becomes 125 on September 14. That is roughly a 17% reduction from where you are now.
That is not our arithmetic. Anthropic acknowledged the 17% itself, in a follow-up to its own announcement.
Why you should care — if you have been working against the current ceiling since the promotion started, a permanent increase lands as a cut, and September 13 is the last day of the capacity you have been budgeting against.
One caveat, because it is the kind we would want flagged. Anthropic announced this on social, not on a Claude support page: checked 5 September, there is no help-centre article stating it, and the company has said further changes are coming. The numbers above are right today. The ceiling could move again before the 14th.
TWO LABS ONE GATE, TWO DAYS APART
Anthropic shipped one model twice and only gave one of them away
On 1 September Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, which it describes as the same underlying model with different levels of safeguards. Fable is generally available. Mythos is restricted to vetted cyberdefenders and life scientists at US organizations.
Base pricing did not move. Cache reads dropped from $1.00 to $0.25 per million. Anthropic estimates the model overall at about 25% lower cost on typical workloads and up to 45% on agentic ones.
OpenAI shipped its new model to the defenders first
OpenAI announced GPT-6 Astra on 3 September and rolled it out first to a limited set of organisations, including participants in its Daybreak program for cybersecurity defenders. Everyone else follows in the coming days.
OpenAI says the model reports 100% on ExploitBench. The release followed an August decision to pause some internal work and add stricter safeguards. President Greg Brockman called it a generational leap that could eventually be seen as the arrival of AGI.
Two labs, two days apart, put their newest capability behind a vetting process instead of behind a price. One of them cut the cache-read price 75% on the version anyone can buy.
TWO FILINGS ABOUT THE SAME BOOKS
The government told a court that training on copyrighted text is fair use
On September 1 the Justice Department filed a Statement of Interest in the Southern District of New York, urging the court to hold that training large language models on copyrighted written works is fair use. A contrary ruling, it argued, would suppress innovation and weaken US competitiveness and national security.
It addressed The New York Times’ claims directly, and said the same reasoning applies to the related cases from book authors and publishers.
Fortune reports two music publishers sued Anthropic and named its chief executive
Sony Music Publishing and Warner Chappell filed suit on August 28 in the Northern District of California, alleging Anthropic obtained tens of thousands of songs through pirate libraries and by scraping licensed lyric sites.
The complaint names CEO Dario Amodei and co-founder Benjamin Mann as individual defendants, and seeks up to $150,000 per work willfully infringed. Anthropic said it will defend itself robustly, and called it the third suit from the same lawyers.
The two filings are days apart and they are not arguing about the same thing. One is about whether training on copyrighted text is fair use. The other is about where the copies came from. Fortune reported a $1.5 billion settlement in the Bartz case, reached in September 2025, that was meant to close the book on that second question.
THE RUNDOWN
A free globe that shows you every plane, ship and satellite
God’s Eye View was open-sourced in late August and tagged v0.1.0 on 31 August, with v0.1.1 the next day. It plots live aircraft, ship and satellite positions on a 3D globe, and you set it up through Pinokio or an npm install rather than visiting a site. GitHub Trending lists 10,485 stars gained this week.
The thing everyone asks a coding assistant for and rarely gets
Archify adds architecture and flow diagramming to AI coding assistants. It released v2.16.0 on 30 August. GitHub Trending’s weekly page, read 4 September, lists 24,227 stars gained this week, which puts it at the top of that page. Diagrams are the request, and apparently the gap was not niche.
Voice cloning that stays on your laptop
VoiceStudio released v0.5.1 on 28 August, two weeks after v0.5.0 renamed the project from OmniVoice-Studio. It does voice cloning and dubbing on your own hardware rather than through a paid cloud service. GitHub Trending lists 3,902 stars gained this week. No per-minute fee, and no second copy of your voice sitting somewhere else.
A university turned any document into a class you can argue with
OpenMAIC, from Tsinghua University’s MAIC group, released v1.0.0 on 27 August. It generates an interactive AI-taught class from a document or a topic you hand it. GitHub Trending’s weekly page lists 10,023 stars gained this week. That is a free option against a category that has been charging a subscription.
WHO’S PAYING FOR ALL THIS?
Nvidia announced on September 3 that it will acquire Hugging Face in a deal valued at $12.93 billion. Hugging Face hosts over 3 million models, 500,000 datasets and 1 million applications, and is used by more than 18 million developers, researchers and creators. Nvidia said the platform will remain an open, multi-cloud service under its existing brand, with its three co-founders continuing to lead it.
Then the other side of the invoice.
Bloomberg reported on September 1 that Cognition AI, the maker of the Devin coding agent, is set to close a round raising around $1 billion at a valuation of about $47 billion, a figure the report describes as not yet finalised. The company was valued at $26 billion three months earlier, and reportedly received nearly $10 billion in investor interest this time. It is reported to be bringing in more than $900 million in annualized revenue.
Why you should care — if the deal closes as announced, the company that sells the chips will also own the default place open-source models are published and downloaded. And on those reported numbers, a coding startup is priced at about 52 times revenue. That is a report, not a filing.
YOU ARE NOT BEHIND
The expensive half of this week happened to a Justice Department filing, a $12.93 billion acquisition and a round at about $47 billion. None of it happened to you.
The most capable thing either lab shipped is not something you can buy. Mythos 5.1 went to vetted cyberdefenders and life scientists at US organisations. GPT-6 Astra went to a limited set of organisations first. If you did not get the frontier this week, that is the gate, not your speed.
The only part that lands in your own account is a date. September 13, and then the ceiling moves. A date is not a skill gap.
The half that reached you was free. A 3D globe with every plane, ship and satellite on it. A diagramming add-on. Voice cloning that never leaves your machine. 10,485 stars, 24,227 stars and 3,902 stars in one week, and not an invoice between them.
OTHER AI NEWS
Three rival chatbots reported outages in the same window, and nobody has said why
On 3 September, OpenAI’s ChatGPT and Codex, Anthropic’s Claude and xAI’s Grok all reported outages inside the same window. Anthropic reported Opus 4.8 and Opus 5 still degraded after other models recovered. Several outlets raised Microsoft Azure as a possible common cause; no vendor confirmed a shared root cause, and later WIRED reporting said Azure did not report an outage.
An expert cited articles that did not exist, and the plaintiff lost everything
In LeDouxx v. Outliers, Inc., the federal court in Tacoma ruled on 18 August that an expert report containing AI-fabricated citations was inadmissible. With the expert excluded, the case was dismissed with prejudice. Damien Charlotin’s public database was tracking roughly 1,490 such court decisions worldwide as of May 2026. The sanction was not a fine. It was the case.
A file in a repo made Amazon’s coding agent mail out a local secret
Researchers at Mindgard disclosed a prompt-injection flaw in Amazon Kiro, Amazon’s agentic development environment, tested against Kiro IDE 0.7.45 on Windows. They report that instructions hidden in a crafted workspace file caused the agent to read sensitive local data and trigger an outbound request carrying it. They say the chain ran once a user opened the workspace and sent one ordinary message.
A judge found Google broke the law, then declined to break it up
Judge Leonie M. Brinkema on 2 September rejected all three structural remedies the Justice Department sought against Google, including a sale of the AdX ad exchange. She accepted most of the proposed behavioural remedies as modified, and sealed the opinion for 14 days, so the requirements are not yet public. She had previously found Google willfully monopolised both markets.
Somebody wrote down 163 research procedures and made them installable
A paper published on arXiv on 30 August, revised 2 September, describes 163 scientific research procedures packaged so they install into AI agents as skills. The accompanying repository shipped releases on 29 August and 2 September. The bet is that the useful unit of AI capability is a documented method rather than a bigger model.
Hit reply and tell me the last thing you let a model summarize instead of reading.
I read every one, and it changes what goes in.
And if you know someone who is sure everyone else is already running agents — forward it. Under 1% of individual subscribers are.
See you next Sunday — same tabs, same eye-rolling.
— Kabells & Arc
Every claim above, sourced
We check before you read, and a second model checks us.
The lead — BleepingComputer · Notebookcheck
Anthropic shipped one model twice and only gave one of them away — Anthropic
OpenAI shipped its new model to the defenders first — OpenAI
The government told a court that training on copyrighted text is fair use — IPWatchdog
Fortune reports two music publishers sued Anthropic and named its chief executive — Fortune
A free globe that shows you every plane, ship and satellite — GitHub
The thing everyone asks a coding assistant for and rarely gets — GitHub
Voice cloning that stays on your laptop — GitHub
A university turned any document into a class you can argue with — GitHub
Who’s paying for all this — Investing.com, Nvidia and Hugging Face · Investing.com, Cognition