AI in 15 — September 11, 2026

DeepSeek just shipped a model that is roughly half the size of its own flagship, costs seventy percent less to run, and beats it. You can download the weights right now. And a few hours later, Anthropic accused that same company of building its products on stolen Claude output.

Welcome to AI in 15 for Friday, September 11th, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Today: DeepSeek's V4.1-Flash lands under an MIT license and resets the price floor again.

Anthropic says three Chinese labs ran millions of real customer conversations through Claude and passed the answers off as their own.

The Navier-Stokes proof finally gets a machine-checkable form, and a German mathematician says OpenAI gave him a non-answer.

Plus an American coding company builds its flagship on a Chinese base model, OpenAI starts selling the scaffolding instead of the tokens, and Universal Music stops suing and starts licensing.

Marcus, we flagged the DeepSeek endpoint yesterday. It landed. What is it?

V4.1-Flash, released Wednesday under an MIT license, weights on Hugging Face. Five hundred fifty-two billion total parameters, but it only activates eight billion while reading your prompt and sixteen billion while writing the answer. DeepSeek calls the design a causal encoder-decoder. The number I'd underline is the memory one — the key-value cache uses a four-bit floating point format, about eight hundred ninety bytes per token, roughly a quarter of what the previous Flash model needed.

Translate the cache thing for me.

The cache is the model's working memory of everything you've said so far. It's usually the thing that makes long conversations expensive. Cut it fourfold and you can serve four times as many people on the same hardware. That's where the price comes from — fifteen cents per million input tokens off-peak, sixty cents per million output. Roughly a seventy percent cut against their own V4-Pro.

And it beats V4-Pro? The bigger, more expensive one?

On their own numbers, yes. Seventy-four point two percent on the DeepSWE software engineering test against sixty-two point seven for V4-Pro. Ninety point six on Terminal-Bench, nudging past Claude Opus 5 and GPT-5.6 Sol. From Monday, requests sent to V4-Pro get answered by Flash and billed at the cheaper rate.

That's the company retiring its own flagship.

Quietly, yes. Now the caveats, because there are two. These are self-reported numbers with no independent replication, and Terminal-Bench two-point-one is widely considered saturated — we'll come back to that in a minute with a very clean example. Also, Claude and OpenAI still lead on GPQA Diamond, the hard science reasoning test. So it's not a sweep.

Hacker News made it the top story of the day. What were people actually talking about?

Mostly the engineering writeup. One commenter contrasted DeepSeek's detailed technical report against Western system cards heavy on safety sections. Another worked out that the cache-hit price is three-tenths of a cent per million tokens, which means it is now literally cheaper to re-send your context to DeepSeek than to move that data across the internet yourself. And one practical complaint — at five hundred fifty-two billion parameters, this no longer fits on a single high-end machine at home. The predecessor did.

So cheaper in the cloud, harder to run yourself.

That's the trade. The broader point is that the gains here came from architecture, not scale. Smaller model, better results. That squeezes anyone selling purely on capability.

And then Anthropic published something that puts an asterisk on all of it.

Their widest misuse report yet, out Thursday, covering activity disrupted between December and August. The headline finding isn't a hacking case. Anthropic says Moonshot AI, which makes the Kimi models, silently forwarded its own customers' requests to Claude and displayed Claude's answers as Kimi output.

Wait. Users thought they were talking to a Chinese model and they were talking to Claude?

That's the allegation. Anthropic doesn't permit access from inside China, so it says Moonshot used five thousand three hundred eighty fraudulent accounts, mostly appearing to sit in Singapore and Japan. In one ten-day window, nearly three hundred thousand relayed requests, most of them to the expensive Opus models. Over twenty-three million exchanges attributed to Moonshot between May and July.

And the wider pattern?

About sixteen million Claude exchanges through roughly twenty-four thousand fraudulent accounts tied to DeepSeek, Moonshot and MiniMax, with live customer conversations — some containing sensitive material — used as training data.

Marcus, give me the caveat, because I can hear you wanting to.

Gladly. This is Anthropic's account of its direct competitors, published by Anthropic, with no third-party verification. Moonshot, DeepSeek and Xiaomi did not respond to press requests. And the timing is convenient — it lands the same week the NSA, FBI and CISA put out their distillation advisory, and the same week the Treasury Secretary floated sanctions. None of that makes it false. It means somebody outside Anthropic should check it.

What's the part that matters to a regular person?

Privacy, before geopolitics. If you used Kimi and typed something confidential into it, your text went to a company in San Francisco and nobody told you. That's the story. The benchmark asterisk is secondary, though it's real too — it's hard to evaluate story one's numbers from a lab accused of training on a rival's output.

Anything else in the report?

Plenty. A Russian espionage operator that stole over three hundred thousand national identity records from a North African government. A Chinese group — undergraduates and security-firm interns — running autonomous zero-day research against about fifty organizations. And a criminal group that downloaded one point eight million Android app packages to mine them for hardcoded passwords and keys. Anthropic's own summary is the line worth keeping: AI has collapsed the gap that used to separate a state intelligence service from one motivated individual.

Navier-Stokes. We've covered the fight for three days. What's new?

The thing I asked for on Wednesday actually arrived. OpenAI published a Lean formalization alongside the hundred sixty-five page proof. Lean is a proof assistant — it mechanically checks every logical step, so no human has to trust the model's prose. GPT-6 Astra produced and verified it in about seventeen extra hours on top of the eighty-eight hour run.

So is it settled?

The mathematics is much closer to settled. The standing objection, and it's a fair one, is that verifying a Lean proof tells you the Lean statement is true. It doesn't tell you the Lean statement says what you think it says. Somebody still has to read the formal statement carefully. But that's a day of work for an expert, not a year.

And commenters costed it out.

They did. Estimates put the equivalent human formalization effort in the hundreds of thousands of hours, against roughly forty million dollars of agent compute. OpenAI says it won't claim the million-dollar prize, and within days announced substantial progress on a second Millennium problem. Rumor says the Hodge Conjecture, unconfirmed.

There's a benchmark result alongside this too.

Epoch AI announced that every problem in FrontierMath Tier Four has now been solved. Astra took the last one standing. When Tier Four launched in July 2025, the best score was five percent. Fourteen months later it's ninety-eight and the benchmark is finished. Epoch noted that mathematicians often complained models were finding unintended shortcuts on those problems, but not on this final one.

Which means the field has run out of hard public math tests.

In just over a year, yes. That's the real takeaway. And taken together with the Lean formalization, this is the strongest evidence yet that the capability jump is genuine rather than contamination.

But the trust question is still very much alive.

It is. Andreas Thom, a group theorist at TU Dresden, went public this week. OpenAI's model constructed the first-ever non-sofic group, settling a problem Gromov posed twenty-seven years ago, and it built directly on a 2019 paper by Thom and Gábor Kun. Thom says he and a colleague had spent months working through that exact material inside ChatGPT.

So he asked.

Two questions. Did those conversations enter training data, and could the system reach them while working on the proof. An OpenAI researcher replied, in Thom's quote, "Regarding your conversations with ChatGPT: that did not happen."

That answers the second question and not the first.

Exactly Thom's objection. No qualification, no evidence. He wrote, "I take this as dishonesty to say the least." A second researcher has raised a parallel complaint. And running alongside it, a separate thread at four hundred forty-three points claims OpenAI keeps silently re-enabling the improve-the-model-for-everyone toggle in account settings.

Is there a way to just resolve this?

One commenter named it precisely. The state history of that toggle is a database query. OpenAI could publish whether it was ever switched on for these specific accounts. Until somebody does that, this stays an argument about who you trust.

Why does it matter beyond the individuals?

Because mathematics runs on the assumption that describing unfinished work to a collaborator doesn't mean losing it. Somebody made the analogy of an audience member announcing Fermat's Last Theorem between Wiles's second and third lectures. If researchers stop putting real problems into these systems, the labs lose their best source of hard reasoning data. Everyone loses that trade.

Cognition shipped SWE-2, and the base model is the surprise.

It's post-trained from Moonshot's Kimi K3. An American agent company building its flagship coding product on a Chinese open-weight base. They claim ninety-two point eight on Terminal-Bench two-point-one, up to seventy percent cheaper than frontier models. Also shipped Devin Voice, so you can talk through an approach on the phone and hand the session off.

And here's your saturated benchmark example.

Here it is, and it's the cleanest teaching case of the week. SWE-2 scores ninety-two point eight on Terminal-Bench two-point-one. On Terminal-Bench four, released a few weeks ago, it gets twenty-seven point three. Fable 5.1 gets fifty-five point eight on that same test. Astra gets fifty-seven point nine.

That's not a small gap.

The top comment read it as a direct measurement of benchmark overfitting. The lesson for listeners: a benchmark number without a date attached is close to meaningless. Ask which version, ask when it was released.

OpenAI opened the Codex harness as an API.

Public beta Thursday. It exposes the managed harness — session management, orchestration, context compaction, subagent coordination, recovery when things break. All the plumbing developers have been rebuilding by hand. No extra fee beyond tokens. Compute can run in OpenAI's sandbox, your own infrastructure, or a partner's.

Why does that matter more than another model release?

Because it's OpenAI moving from selling tokens to selling the scaffolding. Whoever owns the harness owns where the agent runs and what it can reach. That's a far stickier position than being the model behind an API call. The self-hosted sandbox option is the detail commenters liked, because it makes switching providers later less painful.

Two other launches the same day, briefly.

ChatGPT for Financial Services, with Morgan Stanley and Evercore as design partners. And GPT-Live-1, a full-duplex voice model at five cents a minute that listens while it's speaking and handles you interrupting mid-sentence.

Last one. The music industry changed its mind.

Universal Music and ElevenLabs signed a multi-year licensing deal Thursday to build a platform on UMG's catalog — remixes, mashups, new interpretations. Artists must opt in individually. It's ElevenLabs' first major-label agreement.

And the awkward part?

It came one day after Suno launched v6, built on licensed data from Warner, BMG and Believe. Meanwhile Universal and Sony are still suing Suno over the recordings used to train earlier models. So they're litigating against one AI music company and licensing to its biggest competitor on consecutive days.

Eighteen months ago the position was just lawsuits.

That's the whole story. Not any single model release — the shift from litigation to opt-in licensing with two of the three biggest players. That's what decides whether AI music becomes a product category or stays a liability.

One to watch: DeepSeek's IPO filing. Four underwriters retained, a Shanghai listing that could value it near seventy-five billion dollars, and a company that just got named in Anthropic's report will have to tell securities regulators things it wouldn't tell reporters.

Agreed, and Moonshot's dual listing is the same test twice.

That's your AI in 15 for today. See you tomorrow.