AI in 15 — August 01, 2026

OpenAI cut its prices eighty percent on Thursday. It held that lead for about fourteen hours.

Welcome to AI in 15 for Saturday, August 1, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Today: DeepSeek answers OpenAI's price cut overnight — with a model built to slot straight into OpenAI's own coding agent.

Google fixed more Chrome security bugs in one month than in the previous two years combined, and Hacker News has questions.

A conjecture James Clerk Maxwell floated in 1873 falls — with the key idea suggested by a language model.

Thinking Machines puts a 276-billion-parameter model on the table under Apache 2.0.

Plus Microsoft's tiny cyber model and its asterisk, an AI hedge fund down sixty-seven percent in a month, and a laptop running a two-point-eight-trillion-parameter model. Slowly.

Marcus, we led yesterday with OpenAI's price cut. Walk me through what happened next.

At six fifty-six UTC on Friday morning, DeepSeek pushed V4-Flash-0731 into public beta. Fourteen cents per million input tokens, twenty-eight cents output. Luna — OpenAI's newly discounted fast tier — is a dollar twenty on output. So DeepSeek came in at roughly a quarter of the price of the thing that had just been cut eighty percent.

And on capability?

Sparse mixture-of-experts, 284 billion total parameters, only thirteen billion active per token. A million-token context window. Artificial Analysis — third party — scores it fifty on their Intelligence Index, rank three out of a hundred and one. And the jumps inside a single post-training cycle are what made people sit up: Terminal Bench went from sixty-one-point-eight to eighty-two-point-seven. DeepSWE went from seven-point-three to fifty-four-point-four. Sevenfold.

But the number that matters isn't a benchmark, you said.

It isn't. V4-Flash speaks OpenAI's Responses API format natively, and it's explicitly adapted for Codex — OpenAI's coding agent. That turns switching providers from a rewrite into a config change, on exactly the surface OpenAI is trying to defend.

That's quite deliberate.

There's comedy in the timing too. Tibo Sottiaux, who runs Codex and ChatGPT at OpenAI, asked on X at four thirty-five that morning what people wanted improved in Codex. Two hours and twenty minutes later, DeepSeek shipped a model tuned for it.

Okay, hit me with the caveats, because I know you have them.

Every headline number here is vendor-supplied on the vendor's chosen field. DeepSeek ran the public code-agent tasks on its own unreleased harness at maximum effort. Two of the nine benchmarks are its own internal sets. And GPT-5.6 doesn't appear in DeepSeek's comparison table at all — they're benchmarking against Claude Opus 4.8.

How does it do there?

Genuinely well on some things. Within four points on five of nine benchmarks, and within half a point on Agents' Last Exam — at roughly ninety times less per output token. But the gap holds where it always holds: repository-scale work. NL2Repo, fifty-four versus sixty-nine-point-seven. DSBench-Hard, fifty-nine-point-six versus seventy-one-point-seven. Opus wins those by double digits.

So what's the practical read for someone building on these APIs?

Frontier capability now has a shelf life of about two months, and the company that cut first held its advantage for less than a day. Model choice is turning into a procurement decision rather than a strategic bet. And the place where the price collapsed fastest this week was China — Chinese models reportedly account for something like forty-six percent of US enterprise token usage on OpenRouter.

Completely different. Google says it fixed one thousand and seventy-two Chrome security bugs in a single month.

Across Chrome 149 and 150, both shipped in June. For scale: the previous twenty-three Chrome versions, over two full years, fixed one thousand and thirty-six. So one month beat two years. Google credits an AI pipeline using Gemini models to find, triage, and write patches. One of the finds was a critical sandbox escape that had been sitting in the codebase for thirteen years.

Thirteen years. That's uncomfortable.

Chrome's director of engineering, Doug Turner, said LLMs have "fundamentally shifted the economics of cybersecurity, transforming vulnerability discovery into an automated, industrial-scale operation." Microsoft patched five hundred and seventy flaws in July and cited AI too. Apple, interestingly, shows no spike at all — four hundred and eighty-two bugs this year, roughly its historical pace.

Five hundred comments on Hacker News. What's the sharp objection?

The best one I've read all week. How many of those automated fixes got reverted? How many introduced a new bug? What's the false-positive rate on the finding agents? The blog post has counts for everything that went right and nothing for what could go wrong. A raw bug count is a weak proxy for security — and an internal push to hit a number before publishing a chart would produce exactly this chart.

Is there anything pointing the other way?

Yes, and it's a good datapoint. Firefox paid out nothing at Pwn2Own Berlin in May — first time since 2007. That suggests the easy bugs really are being harvested.

And this cuts both ways, presumably.

Of course it does. The capability that found a thirteen-year-old sandbox escape is available to anyone who wants to look. What it favours is whoever has the codebase and the compute to run it continuously — which is a real structural advantage for the big browser vendors, and not much comfort for everyone else.

Now this one I loved. A conjecture from 1873 has been disproved.

Maxwell — James Clerk Maxwell, the electromagnetism Maxwell — suggested that for n point charges, if all the equilibrium points in the electrostatic potential are well-behaved, there can be at most n-minus-one squared of them. For five charges, that ceiling is sixteen. Three mathematicians — Philip Arathoon, Gavin Ball and Matthew Kvalheim — posted a paper on Wednesday exhibiting a five-charge configuration with at least twenty-four.

And the AI part?

The authors credit GPT-5.6 Sol with suggesting the key construction idea. They then developed, verified and rigorously proved every step themselves. Worth being precise: the arXiv abstract doesn't name the model — that attribution comes from the authors' own posts and the body of the paper. Greg Brockman amplified it.

Is it a big deal mathematically?

The top Hacker News comment says no, and I think that's fair. The conjectures currently available for AI-assisted exploration tend to be the niche ones. Maxwell's isn't load-bearing for physics. Another thread notes AI-produced proofs are often correct but inelegant — machines don't care about beauty.

So what is interesting about it?

The division of labour. The model proposed a construction; humans verified it. That's a different mode from autocomplete and different from oracle. And the economics are appealing — a wrong suggestion is cheap, a right one is a paper. It also matters that this was hand-verified, unlike that recent embarrassment where an AI "proof" of the Collatz conjecture turned out to be exploiting a bug in the Lean theorem prover.

Best career advice from the thread?

If theory gets cheap, experimental ability becomes the bottleneck.

Mira Murati's Thinking Machines released Inkling-Small, and the weights are public.

Apache 2.0, no usage restrictions. Mixture-of-experts, 276 billion total parameters, twelve billion active, million-token context, natively multimodal across text, images and audio. A dollar twenty per million output tokens against four-oh-five for the larger Inkling.

And it beats its bigger sibling?

On reasoning, yes. Thirty-one-point-six percent on Humanity's Last Exam against Inkling's twenty-nine-point-seven. Eighty-point-two on SWE-Bench Verified, eighty-nine-point-five on GPQA Diamond. Where the big model still wins decisively is raw factual coverage — SimpleQA Verified is forty-three-point-nine for Inkling versus twenty-point-six for Small.

So what does that tell an engineer?

It's the most useful lesson in the story. Shrinking a model costs you memorised facts far more than it costs you thinking ability. Which is a strong argument for small models plus retrieval rather than large models plus hope. And an Apache-2.0 model with a million-token context from a US lab is a fairly direct answer to the open-weights momentum coming out of China.

Microsoft's cyber model — we touched this on Wednesday. What's new?

The detail underneath the chart. MAI-Cyber-1-Flash, five billion active parameters, scores ninety-five-point-nine-five percent on CyberGym inside Microsoft's multi-agent harness — about twelve points above Anthropic's Mythos 5, at half the cost. But read the chart label. The winning configuration is MAI-Cyber-1-Flash plus GPT-5.4. The small model handles roughly ninety percent of tasks; the hard ten percent routes to a frontier model.

So it's a routing result, not a small-model result.

Exactly, and that's precisely how you get to a fifty percent cost reduction. It's also Microsoft methodically reducing its dependence on OpenAI for routine inference while still needing OpenAI for the tail. Worth pairing with a post trending this week titled "Everyone is building LLM routers, we deprecated ours" — generic prompt-level routing is proving hard. Domain-specific routing with a purpose-trained specialist appears to work.

Money. Leopold Aschenbrenner's AI hedge fund fell sixty-seven percent in July.

Situational Awareness, run by the former OpenAI researcher. Letter to investors on Thursday. They sold the bulk of the public equities portfolio to Ken Griffin's Citadel this week. His words: "We came closer to permanent capital impairment than is acceptable to us." He blamed short sellers targeting the firm's positions and compared it to a bank run.

Context, though.

Important context in both directions. The fund reportedly peaked near forty billion in value and sold at around ten. But even after July it's still up about eighty percent on the year. This is a leverage blowup, not a thesis being disproven.

And the backdrop?

A broad semiconductor selloff. The PHLX Semiconductor index fell ten percent in a week, steepest since April 2025, roughly one-point-three trillion dollars erased. Analysts point to Moonshot's Kimi K3 landing on July seventeenth as a "DeepSeek 2.0" moment, and to Meta becoming a seller of GPUs rather than only a buyer — which undercuts the perpetual-scarcity thesis holding those valuations up.

Your read?

Capital markets and capability run on two different clocks. The models got cheaper and better this week while the stocks priced on them fell hard. The lesson here is about leverage, not about AI.

Two things going viral among developers. Start with the fun one.

WASTE — Weight-Aware Streaming Tensor Engine. It runs Kimi K3, two-point-seven-eight trillion parameters, in about twenty-nine gigabytes of RAM on a sixty-four-gig MacBook Pro, by keeping the weights on disk and streaming the experts it needs each token off NVMe. The container is nine hundred and eighty-two gigabytes on disk.

And the speed?

About half a token per second. One commenter costed the electricity at roughly five dollars per million tokens, excluding hardware — worse than just paying DeepSeek.

So why does anyone care?

Because it makes the sparsity of these architectures viscerally concrete, and it proves frontier-scale models will run on consumer hardware if you'll trade throughput for privacy. The other one is qm, from YC Software — a multiplayer agent harness where every person and every room gets its own scoped memory, files, permissions and sandbox. The signal there is that the interesting problem in agents has moved from the agent loop to scoping and permissions. Best comment in the thread: "I gave an agent its own Slack channel and it started scheduling meetings with other agents without me. I've never felt more like middle management."

One to watch: tomorrow, August second, California's AI Transparency Act becomes operative. Any generative AI system with over a million monthly users has to ship a free public detection tool and embed machine-readable watermarks. Watch whether the big providers actually ship by Sunday.

Counter: latent watermarks die to a screenshot, and the detection tools are graded by the same firms whose output they're meant to catch. Watch the compliance theatre, not the technology.

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