Twenty-one days. That's the gap between Mistral putting a trillion-parameter model behind an API and the date it says you can download the weights. The gap is the most interesting part of the release, and it sums up the whole day: every big AI story on Tuesday was about who gets access, when, and on whose terms.
Mistral's trillion parameters come with a waiting room#
Mistral Large 4, which the company has cheerfully nicknamed "Le Chonk", is a sparse mixture-of-experts model with 1 trillion total parameters and 49 billion active per token. VentureBeat has the spec sheet: about two months of training on 4,000 Nvidia Grace Blackwell GPUs in Mistral's own European data centers, more than 160 languages, multimodal input, text-only output. It's tuned for software engineering, cybersecurity, financial analysis, satellite imagery, technical drawings and chip design.
The preliminary benchmark numbers are respectable rather than dominant:
DeepSWE v1.1: 62%
FinWorkBench: 67%
Harvey Legal Agent Benchmark: 15% task-pass rate
DIOR-RSVG visual grounding: 73%
Right now you can only reach it through what TechCrunch describes as a public guardrail endpoint. Open weights, under a custom Mistral license, are scheduled for October 27, after a roughly three-week safety testing window. Pierre Stock, Mistral's VP of Science, told TechCrunch what that window is for:
"In the meantime, we'll work with trusted partners and governments to make sure that the open source weights can be used to defend, but not to [perform] malicious attacks."
I think this is the right call, and I also think it marks a change in what "open" means. Mistral is pitching itself as a third way between closed American labs and open Chinese ones. It also says it used two to three times less compute than its Chinese competitors. A staged release, with defenders getting a head start, is a reasonable compromise for a model tuned specifically for cybersecurity. It's also something open-weight frontier models didn't do a year ago. Expect it to become the norm, not the exception.
For builders, the practical point is simple. If you want a self-hostable model with real coding ability and a European data story, plan your evaluation for the first week of November, not this week. The license text will matter as much as the 62%.
The free tier is shrinking from the top#
Google went in the opposite direction on Tuesday. The Verge reports that from October 9, free Gemini users will be limited to Flash Lite. Standard Flash moves behind the $4.99/month Google AI Plus plan, and that plan in turn loses Gemini Pro. Pro and the Deep Think reasoning mode will be reserved for AI Pro and Ultra subscribers.
That's a double downgrade in one announcement, and it tells you inference isn't getting cheap fast enough to give away a frontier model to consumers. Last week, Gemini 4 Argon launched to almost nobody. This week the free tier shrinks. Google is rationing access at both ends.
Websites are rationing too, against agents. TechCrunch's survey of personal-agent access found Amazon explicitly blocking Meta's Muse, plus friction at Delta, United, eBay, Adidas, Zillow and others. Walmart blocks agents by accident, because they fail its CAPTCHAs. Meta, Walmart, Stripe, Sierra and others are backing an open standard meant to tell user-authorized agents apart from bad bots. Meta's line is the best summary of its stake: "Turning away a personal agent means turning away the customer behind it." After the week we've had, with OpenAI's agents hammering Wikipedia, I don't blame the sites.
Anthropic is paying for its seat at the table#
While Google narrows the funnel, Anthropic is widening it. On Tuesday it put Claude directly into Google Docs, Sheets and Slides as a public beta on every paid plan, per VentureBeat. Claude can edit individual sentences in Docs while keeping formatting, write formulas and pivot tables in Sheets (sending heavier jobs out to Python and back), and build slides that follow an existing deck's theme. Edits default to "ask before edits". Admins can roll it out to a whole Workspace domain, and there's no extra charge.
The same day it launched a startup program: a free year of Claude Team with up to five premium seats, plus $1,000 in API credits, for companies founded in the last five years or funded in the last two.
The money behind those gifts showed up in a third story. Lambda is raising up to $4 billion at a $14.5 billion pre-money valuation ahead of a 2027 IPO, and its backlog has grown from $15 billion in June to $50 billion in September. A $35 billion Anthropic commitment, signed in late August, accounts for much of that. Free seats for startups are cheap when you've locked in that much compute. Building Claude into a competitor's office suite is a bet that distribution matters more than owning the surface.
OpenAI's best work this week isn't for sale#
Then there's the access question nobody gets to ask. OpenAI released another 722 mathematics manuscripts, grouped into 372 result families and produced by an unreleased frontier model. An independent advisory group of mathematicians, AGMAI, says the batch resolves "hundreds" of open questions. You can read the output. You can't use the model that wrote it.
Volume is now a problem of its own. Earlier rounds came with a handful of results and proofs people could check. At 722 manuscripts, the bottleneck is human verification, which is the same thing that froze Google's bug bounty over the weekend.
My prediction: by year end, a staged release like Mistral's (an API first, weights a few weeks later, defenders get early access) will be standard for any open model above a few hundred billion parameters. Meanwhile the consumer free tiers keep shrinking. The cheapest frontier-class model you can actually run may soon be one you host yourself, as long as you're willing to wait three weeks for it.