The Anthropic copyright settlement just became the largest known copyright recovery in history. A federal judge approved a $1.5 billion payout to authors whose pirated books were used to train Claude, and it landed on the same day Microsoft announced a multibillion-dollar deal to expand its partnership with French AI firm Mistral. Meanwhile, Google confirmed that Gemini 3.5 Pro is still not ready, and the company deprecated temperature and sampling parameters on its newest models. The signals this week point in one direction: the companies building frontier AI are paying real money to settle the past and paying even more to lock in the future.
The Anthropic copyright settlement is now law
On Monday, U.S. District Judge Araceli Martínez-Olguín approved the $1.5 billion settlement between Anthropic and a class of authors and publishers. The case began when thriller novelist Andrea Bartz and others sued Anthropic for using pirated copies of their books to train Claude. About 482,000 books are covered by the ruling, and roughly 91% have already been claimed by authors or publishers who are now due payment — approximately $3,000 per book.
The settlement is notable for what it settles and what it does not. Judge William Alsup, who issued the preliminary approval before retiring, previously ruled that training AI on copyrighted books can qualify as fair use. Anthropic’s deputy general counsel Aparna Sridhar framed that part as a landmark win. But the ruling also made clear that using pirated copies to do that training carries a real price tag. The $1.5 billion figure dwarfs previous copyright recoveries and sets a benchmark that every other frontier lab will have to reckon with.
For builders, the takeaway is concrete: you can train on copyrighted material under fair use, but you cannot pirate it to do so. If your AI pipeline ingests books, articles, or datasets, the provenance of that data now has a quantifiable legal cost. This is not a hypothetical risk anymore — it is a settled precedent with a dollar figure.
Microsoft bets multibillions on Mistral
The same day the Anthropic settlement was approved, Microsoft announced an expanded strategic partnership with Mistral, the French AI company. The deal is described as multibillion-dollar and gives enterprises and regulated industries access to frontier AI models they can control more directly. Microsoft is also renting Mistral’s GPU capacity as part of the arrangement.
This matters for two reasons. First, Microsoft already has a deep partnership with OpenAI, but the company is clearly hedging by building relationships with multiple frontier labs. Second, the framing — “frontier AI they can control” — signals that regulated industries are demanding AI that runs within their own governance boundaries, not just API calls to a third party. That is the same tension meLink addresses with privacy-respecting agentic AI: customers want capability without surrendering control of their data.
Google confirms Gemini 3.5 Pro is not ready yet
Google’s official blog post about its new Flash models included a quiet but significant admission: Gemini 3.5 Pro is “currently testing with partners” and will be made “broadly available as soon as it’s ready.” The company also confirmed it has started pre-training for Gemini 4. The message is clear — Google is shipping capable Flash models now, but its flagship Pro model is still not where it needs to be, particularly on coding tasks.
For developers, there is a second breaking change buried in the API documentation: temperature, top_p, and top_k are now deprecated and ignored on Gemini 3.6 Flash and 3.5 Flash-Lite. In future model generations, supplying these parameters will return an HTTP 400 error. If your app passes sampling parameters to Gemini, you need to strip them now. Google is pushing developers toward system instructions for determinism rather than temperature tuning — a meaningful shift in how production AI agents are configured.
This is the kind of quiet API change that breaks production agents at 2 a.m. If you are running Gemini Flash models in production, removing deprecated parameters is a maintenance task you should prioritize this week, not next quarter.
Open-source watch
Poolside Laguna S 2.1: a coding model that fits on a workstation
Poolside released Laguna S 2.1, a 118B total parameter Mixture-of-Experts model with only 8B activated parameters per token and a 1M-token context window. It scores 70.2% on Terminal-Bench 2.1 in thinking mode, which measures long-horizon agentic tasks executed through a terminal. The model was trained in under nine months and is compact enough to run complex work on local machines. For teams evaluating local AI deployment options, this is a serious contender — it outperforms models several times its size on coding agent benchmarks.
Cisco Antares: open-weight security models
Cisco released Antares-350M and Antares-1B as open-weight models on Hugging Face, purpose-built for vulnerability localization — the task of pinpointing where known security vulnerabilities exist inside a codebase. These small models outperform much larger closed- and open-weight models on this specific task, and their compact size means they can run locally or on-premises, keeping sensitive source code inside your own environment. This is a good example of the small-model-for-a-specific-job pattern that is becoming increasingly viable.
Jack Dorsey’s Buzz: open-source agents in your chat
Jack Dorsey and Block launched Buzz, an open-source workspace that combines team chat, AI agents, and Git hosting under one identity system. Built on a self-hostable Nostr relay, every message, code event, and approval is stored as a cryptographically signed event. Agents participate as first-class members — they can search discussions, open repositories, submit patches, and run workflows. It is an early experiment, but the design philosophy is worth watching: agents as accountable participants with signed identities, not invisible background processes.
What builders should take from this
Three things stand out this week:
- Data provenance has a price. The Anthropic copyright settlement sets a real dollar figure on training data. If you are building AI products, audit where your training data comes from. Fair use covers training, but pirated sources do not.
- API stability is not guaranteed. Google deprecated temperature and sampling parameters with a future hard error. Monitor your model providers’ changelogs and build abstraction layers that can absorb breaking changes without rewiring your whole stack.
- Small specialized models are arriving. Cisco Antares and Poolside Laguna S 2.1 show that purpose-built, compact, open-weight models can beat general-purpose giants on specific tasks — and run locally. For privacy-sensitive workloads, this trend matters more than the next frontier model.
The practical takeaway
The Anthropic copyright settlement tells you that the legal infrastructure around AI training data is now real and expensive. The Microsoft-Mistral deal tells you that large enterprises are paying for AI they can govern. And the Google Gemini delay tells you that even the best-funded labs cannot ship on arbitrary schedules. For anyone building with AI agents — whether that is a website assistant, a coding tool, or a security scanner — the message is the same: invest in data provenance, build for API churn, and do not assume the biggest model is the only option. The companies that treat these as engineering problems, not legal or marketing problems, will be the ones that ship reliable AI products.
Leave a Reply