AI Morning Brief · Oct 24

Alex: Good morning, it is Saturday, October twenty-fourth, and this is the AI Morning Brief. I am Alex.

Jordan: And I am Jordan. On today's show: Kimi's new flagship gets within a hair of Claude Opus on coding, DeepSeek closes another massive funding round, and Anthropic uses Claude to mathematically prove that open source code is safe.

Alex: Plus a giant new open model from the InternLM team, a video generator that builds ten minute films from a single prompt, and this week's pricing corner.

Jordan: Let us start with Kimi. Moonshot AI dropped Kimi K3.1 Pro yesterday, and the headline number is striking.

Alex: It is. A mixture of experts design with two point eight trillion total parameters, about one hundred four billion active per token, and a one million token context window. On Moonshot's own charts it scores sixty-five point seven percent on Terminal-Bench four point oh, against sixty-six point four for Claude Opus five point five.

Jordan: So basically a photo finish with a model that costs several times more.

Alex: Exactly. Weights are promised by November fourth, and the API is live now at two dollars per million input tokens and ten for output. That is roughly two thirds of what the previous generation cost, and it is already on OpenRouter.

Jordan: An open weight model nearly tying Claude Opus on coding, at a fraction of the price.

Alex: That pattern keeps repeating this month. Now the money story: DeepSeek just raised about fifteen billion dollars.

Jordan: Another round? They only raised earlier this year.

Alex: Second one, per Bloomberg and the Wall Street Journal. Total is now over twenty billion in six months, at a pre money valuation near seventy-four billion. Reported backers include CATL and Tencent.

Jordan: Where does that kind of money go for a company selling its API for pocket change?

Alex: Infrastructure and open weights. Founder Liang Wenfeng told investors annualized revenue hit one billion dollars, with gross margins in the seventy to eighty percent range on API access.

Jordan: And the timing is pointed, right?

Alex: The round landed just days after Washington ordered federal agencies to pull DeepSeek technology over security concerns. The market is betting the open weight pricing playbook wins anyway.

Jordan: Now a very different story from Anthropic.

Alex: Anthropic teamed up with a verification company called Theorem and used Claude to formally verify one hundred twenty-three open source libraries at the assembly code level. Not tests, actual machine checked mathematical proofs.

Jordan: So proof that the code has no memory safety bugs, instead of just testing for them.

Alex: Right. Roughly one hundred eighty million lines of proof, bugs patched, a subset released publicly. Work that took specialists years per project got done in weeks. The proofs only cover what the specification states, but for security critical code this is a real shift.

Jordan: And Claude Security got an update too.

Alex: Continuous scanning with automatic fix tracking, now in public beta for Enterprise customers. It rescans every commit and marks findings fixed when the repair lands. Thousands of organizations have surfaced over a million vulnerabilities with it since April.

Jordan: Two quick hits before pricing. InternLM released Atria-Dawn yesterday.

Alex: A seven hundred forty-four billion parameter agentic model with a five hundred twelve thousand token context window, weights up on Hugging Face and ModelScope.

Jordan: And VideoGen Model One?

Alex: A new video model that turns a single prompt into a finished multi shot video up to ten minutes long. A one million dollar creator credit program opened this week, with partner rollout starting Monday.

Jordan: Time for the pricing corner. What do the numbers say?

Alex: Two things. First, Kimi K3.1 Pro at two and ten dollars per million tokens undercuts its own predecessor by a third while nearly matching Claude Opus on coding. The frontier quality to open weight price ratio keeps getting better.

Jordan: And the second?

Alex: DeepSeek's effective price is even lower than the sticker. Per recent OpenRouter data, the official endpoint reports a ninety-one point five percent cache hit rate, so the input price actually paid averages around one point seven cents per million tokens.

Jordan: Plus a calendar reminder: DeepSeek's free search and fetch calls end on October thirty-first.

Alex: Right. Two weeks of free grounded search for agents, then billing starts.

Jordan: That is the week in AI. Open weights keep pushing the price floor down while quality climbs up.

Alex: Cheap compute, mathematically proven code, and video that writes itself. See you tomorrow morning.