Anthropic beats the Pentagon, Apple opens Siri to rivals, and Wikipedia bans AI writing

The Pentagon tried to blacklist an AI company for being too loud about disagreeing with them. A federal judge just said: not so fast.

That's where we are in 2026 - and it's only Saturday. Strap in.

Anthropic's Claude is currently at the center of a legal battle with the U.S. Department of Defense. But how many parameters does Anthropic's Claude 3 Opus reportedly have, according to industry estimates?

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

If you've been following the Anthropic vs. Pentagon saga - and it's been a wild one - here's the update you've been waiting for.

A federal judge in California sided with Anthropic on Thursday, granting a preliminary injunction that temporarily blocks the Department of Defense's "supply chain risk" designation against the company. Judge Rita Lin ordered the Trump administration to rescind the blacklisting while the lawsuit plays out in court.

The court documents make the reasoning behind the Pentagon's move almost surreal. The DoD's own records reportedly show that Anthropic was designated a supply chain risk because of its "hostile manner through the press" - in other words, it was punished for publicly criticizing the Defense Department's push to use Claude in autonomous weapons systems.

That designation would have cut Anthropic off from federal contracts starting next week. The injunction puts that on hold. The underlying lawsuit - over whether the government can force a private AI company to support applications it explicitly refuses to build for - is still very much alive.

This is the first serious legal test of where AI companies stand when the government comes knocking with demands they consider ethically off-limits. However it resolves, it sets a precedent the entire industry is watching.

Read the full story →

Here's a headline that would have sounded absurd two years ago: Apple is reportedly preparing to let Google's Gemini, Anthropic's Claude, and other third-party AI chatbots plug directly into Siri.

According to Bloomberg's Mark Gurman, iOS 27 will let users choose which AI chatbot Siri hands off queries to - similar to how ChatGPT is currently integrated, but opened up to any third-party chatbot downloaded from the App Store. Siri handles the interface; your preferred AI does the heavy lifting.

This is a genuinely strategic move. Apple doesn't have a frontier AI model competitive with GPT or Claude, and it's apparently decided not to fight that battle - at least not yet. Instead, Apple Intelligence becomes the OS-level glue, and the underlying model becomes a user preference. For consumers who've already built workflows around Claude or Gemini, this removes the last reason to avoid an iPhone on AI grounds.

For Anthropic and Google, getting baked into Siri distribution is enormous. For OpenAI, which currently has the exclusive ChatGPT deal, it means the moat just got significantly smaller.

Read the full story →

Switching AI assistants has always had one painful cost: you lose all the context your old tool built up about you. Google Gemini is now tackling that directly.

Google is rolling out two new tools on Gemini's desktop interface: "Import Memory" and "Import Chat History." The memory import works by having you copy a suggested prompt into your current AI chatbot, paste back whatever it returns, and letting Gemini process that context. The chat history import pulls in conversation logs from other services directly.

It's a smart play in what's quietly becoming an AI platform war. Anthropic added similar migration tools to Claude earlier this month, and now Google is matching the move. The subtext is clear: both companies think there's a meaningful population of users who are unhappy with their current AI but haven't switched because the switching cost felt too high.

If you've been thinking about trying Gemini but worried about losing months of personalization - this is the feature that lowers that barrier. Worth noting: the quality of the import depends heavily on how well your current AI can summarize its own memory of you, which varies a lot in practice.

Side note - if you're building anything in this space, 60sec.site is worth a look. It's an AI-powered website builder that gets you from idea to live site in about a minute. Great for developers or creators who want a fast landing page without the overhead.

Read the full story →

Wikipedia has officially updated its guidelines to ban editors from using AI to write or rewrite articles. The rule - added to the English Wikipedia's policies late last week - specifically calls out AI-generated content's tendency to violate "several of Wikipedia's core content policies."

The ban isn't total. Editors can still use AI for tasks like drafting talk page comments, translating content, or brainstorming - things that don't end up directly in article text. But using a model to generate or substantially rewrite encyclopedia entries is now explicitly off the table.

The core concern here isn't style - it's factual reliability. AI content generation tools produce fluent, confident-sounding prose that can embed hallucinations in ways that are hard for casual reviewers to catch. For a reference source that millions rely on daily, that's an unacceptable tradeoff.

Wikipedia's move is interesting because it's one of the few major platforms drawing a hard content line based on quality and trust rather than copyright or legal pressure. It's also a pointed signal about where the volunteer editor community stands - they built something that depends on verifiable sourcing, and AI text generation fundamentally works against that.

Read the full story →

Two notable speech model drops landed this week, and both are worth knowing about if you work in voice AI or are building anything that needs transcription or speech generation.

First, Mistral AI released a new open-source model for speech generation, targeting enterprise voice agents for sales and customer engagement. It puts Mistral in direct competition with ElevenLabs, Deepgram, and OpenAI in a market that's growing fast.

Then Cohere dropped Cohere Transcribe - a 2 billion parameter automatic speech recognition model designed specifically for enterprise transcription use cases. It's lightweight enough to run on consumer-grade GPUs if you want to self-host, and currently supports 14 languages. The small parameter count is the point: it's built for teams that want to run transcription privately, without routing audio through an external API.

The pattern here mirrors what happened with text models a year ago - the open-source options are closing the gap fast. For developers building voice AI applications, the combination of lower latency, lower cost, and data privacy from self-hosting is becoming a real alternative to the hosted API providers.

Read the Mistral story →

Read the Cohere story →

The answer is D - Anthropic has never publicly disclosed the parameter count for Claude 3 Opus! Unlike some competitors, Anthropic keeps its model architecture details private. Industry analysts have speculated figures ranging widely, but there's no official number. The secrecy is partly competitive strategy, partly a reflection of Anthropic's view that parameter counts are a misleading proxy for capability anyway.

The Apple Siri story got me curious: which AI assistant do you actually use day-to-day? Are you all-in on one (ChatGPT, Claude, Gemini), or do you mix and match depending on the task? Hit reply and let me know - I read every single response, and the answers always surprise me.

Also - if you want to dig into any of today's stories further, the full archive is at dailyinference.com. See you tomorrow!