AI in 15 — August 14, 2026

Two thousand five hundred PhD-level exam questions. Eleven hours and eleven minutes. The model everyone currently calls the smartest on the planet took seventy-eight hours to do the same thing — more than three days of continuous compute. Same questions. Same quality. Different silicon.

Welcome to AI in 15 for Friday, August 14, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Today: Google ships Gemini 3.7 Flash three weeks after 3.6 Flash, at half the price — while the flagship it actually needs is still nowhere.

OpenAI and Cerebras preview a model running at seven hundred and fifty tokens a second, and the interesting argument isn't about speed.

An open-weights model marketed explicitly on cyber capability.

Anthropic in talks to spend six billion dollars — not on talent, on cost structure.

Plus DeepSeek does two contradictory things in one day, and a 24-hour news channel with no humans on camera.

Marcus, Gemini 3.7 Flash. Start with the number that made you sit up.

DeepSWE went from forty-nine percent to sixty-five point three. That's a sixteen-point jump on an agentic coding benchmark for what's nominally a point release. FrontierCode Main, thirty-four to forty-three point six. WebDev Arena Elo, fifteen thirty-eight to fifteen eighty-eight. Google's calling it their most intelligent workhorse model yet, and they flipped Gemini Spark over to it the same day.

And the price.

Seventy-five cents per million input, three seventy-five output. Half what 3.6 Flash cost. But it's introductory — it expires December thirty-first and then doubles.

That's a strange thing to advertise.

Simon Willison flagged exactly that. Why publish a price that doubles in five months for a model whose predecessor lasted three weeks? The commercial read is Google is buying volume in the cheap-and-fast tier before the market settles, betting nobody's still on 3.7 Flash when the bill arrives.

But the reason this is our lead isn't the model.

No. It's that Gemini 3.5 Pro is still delayed, and Google's AI organisation is visibly wobbling. August fifth, Demis Hassabis stepped down as DeepMind CEO to become chairman and Alphabet chief scientist. Koray Kavukcuoglu runs daily operations now. And Jeff Dean — twenty-seven years at Google — is leaving to start a public-benefit corporation called Discovery Loop.

Fortune had a piece on why.

Low morale, a talent exodus, model delays. So the picture is a company that can iterate its cheapest tier every three weeks and still can't get its flagship out the door. Those are not unrelated facts.

Is the Flash tier even defensible?

That was the sharpest question in the thread. Google published no benchmarks against OpenAI's Luna or Terra. And OpenAI just cut Luna's price roughly five-fold. If your differentiator is cheap and fast and everyone's cheap and fast, you don't have a differentiator — you have a price.

Cold open story. Cerebras and OpenAI.

Limited preview of GPT-5.6 Sol running in what they're calling Ultrafast mode, on Cerebras wafer-scale hardware. Up to seven hundred and fifty output tokens per second. Humanity's Last Exam — all twenty-five hundred questions — in eleven hours eleven minutes, against Claude Fable 5's seventy-eight hours twenty-seven. On GDP-Val they claim five point six times end-to-end speedup with no quality loss.

How does the hardware do that?

Forty-four gigabytes of SRAM on the wafer itself. Normal GPU inference is bottlenecked shuttling weights back and forth from memory — Cerebras keeps them on-chip and sidesteps the whole problem.

Caveat?

Two. There's no pricing anywhere in either announcement, which reads as either if-you-have-to-ask, or OpenAI measuring demand before committing. And neither post says outright that the Artificial Analysis comparisons were independently re-run, so some of those speed multiples rest on internal data. I'd probe that rather than repeat it.

You said the interesting argument isn't about speed.

Right, and this is the bit worth keeping. When a model reasons, it makes one pass, top to bottom. Humans iterate — we draft, we scrap, we try again. At seven hundred and fifty tokens a second, running five attempts and picking the best becomes cheaper than one careful attempt is today. That's not a convenience improvement. That changes what you can build around the model.

Z.ai released GLM-5.3 this morning, and the tagline is doing a lot of work.

"Built to Code. Ready for Cyber Defense." Post-trained on a seven hundred and forty-three billion parameter base, and their pitch is refreshingly plain — scaling post-training is all we did. They claim frontier coding and agentic performance while using fewer output tokens, which is a genuine cost lever rather than a benchmark line. Coding Plan and ZCode first, API and open weights after what they call rigorous safety evaluations.

Set the stakes with the predecessor.

GLM-5.2 scored sixty-two on SWE-bench Pro, eighty-one on Terminal-Bench, at roughly a sixth of closed frontier API cost. Semgrep found it beat Claude Code on their vulnerability-detection benchmark at about seventeen cents per bug found. NIST published a formal assessment in July. And Arctic Wolf rated its cyber capability as comparable to Opus 4.6 from February.

So six-month-old frontier offensive capability, downloadable.

That's the shape of it. And I want to be careful here, because the write-up genuinely reads like researchers wrote it — Hacker News noticed that and warmed to it. The technical work looks real. But "emergent cyber capabilities" as a marketing headline on an open-weights release is new, and the phrase "after rigorous safety evaluations" means something quite different when the deliverable is a file.

Because you can't recall a download.

You can't recall a download. The evaluation is advisory the moment the weights ship. And I'd push on the framing too — cyber defence and cyber offence are the same capability pointed in different directions. Nobody has ever demonstrated otherwise.

Fair to say the open tier is having a fortnight?

Kimi K3, GLM-5.3, Qwen 3.8, DeepSeek V4 Pro. One commenter's summary was that GLM's still shy of Sol and Fable — but only by a hair.

Anthropic. Bloomberg says six billion dollars for an Israeli startup called Decart.

Talks are early, could fall through, standard caveats. Decart is known for real-time generative video and world models — Lucy edits live video, Oasis generates simulated environments for robotics training. But the reporting frames something else as the actual driver: GPU optimisation software that cuts the cost of training and running models.

So they're not buying the video.

They're buying margin. Anthropic wants its existing infrastructure to absorb more demand without proportionally more spend. TechTimes tied it directly to a seventy-seven percent gross margin target ahead of a possible IPO. And Decart was valued at six to seven billion privately, so this is roughly par — not a bargain.

Context this against their numbers.

Anthropic projected its first operating profit in Q2 — around five hundred and fifty-nine million on ten point nine billion revenue, more than double Q1. Claude Code passed a billion annualised within six months.

And the sceptical read?

Ed Zitron's argued at length that the profit figure is flattered by a ramp-up discount on their reported one and a quarter billion a month compute deal. I'd note his framing is aggressive. But you don't need his conclusion to find the acquisition telling — buying a margin-improvement company immediately before going public says where the pressure actually sits.

Two years ago labs bought people.

Now they're buying cost structure. Six billion dollars for chip-efficiency software is a statement that inference economics, not capability, is the binding constraint on the business.

DeepSeek did two things yesterday that don't obviously go together.

First, they open-sourced DeepSeek Harness under MIT — an agent framework where every capability is a swappable plugin. Models, tools, skills, sessions, sandboxes, scheduling, the UI. Built on their Cordis kernel, so you can hot-load plugins without restarting a running process. The feature getting attention is observability: an append-only log of everything the model saw, and a trajectory view where you resume, fork, search or replay any run. Six hundred points on Hacker News, and one of the authors turned up in the thread taking feedback.

That sounds genuinely useful.

It is. Now the other half. Effective Sunday, they're raising API prices between fifty percent and eleven hundred percent depending on model, token type and time of day, with a new peak and off-peak split.

Eleven hundred percent.

V4-Flash output goes from a flat twenty-eight cents to sixty-six off-peak, a dollar thirty-two at peak. V4-Pro output from eighty-seven cents to nearly four dollars. But the quietly brutal one is cache-read pricing on Pro — up roughly six times off-peak, twelve times at peak. On long agentic coding sessions cache hits are routinely ninety percent of your input tokens.

Why now?

They say they're allocating resources more reasonably. The read on Hacker News is they're capacity-constrained by domestic demand and using price to shed load. And here's the juxtaposition — on the same day OpenAI cut Luna's price roughly five-fold, DeepSeek raised its own roughly five-fold. As one commenter put it, they went from much cheaper than Luna but twice as slow, to about the same price and still twice as slow.

Which erases the entire market position.

We led with DeepSeek yesterday on exactly that value proposition. Cheap Chinese inference was a structural assumption in a lot of people's 2026 plans, and it expired with two days' notice.

Brief return to the tl;dv breach, because there's a detail we didn't have.

We covered the missing tenant-isolation rule — a hundred and eighty-one thousand meetings, eighty-four thousand users. What's landed since is the mechanism that turns a data leak into a surveillance capability. The exposed records included joinable conference IDs for calls in progress. And the researcher reports that for private calls, presenting as tl;dv — impersonating the notetaker — was usually enough to get waved in.

So the social convention is the attack surface.

Nobody threat-models the notetaker. A bot joining your call is unremarkable, and that unremarkableness is the exploit. Reported January twenty-eighth, still unfixed six months later, CTO never responded.

Two releases to pair, because they're the same story. Meta's Muse Glimmer and Grok 4.6.

Glimmer we covered on launch — thirty billion parameters, genuinely Apache 2.0, one consumer GPU. The detail worth adding is block-level speculative decoding: sixteen-token block speculation gives a claimed three-times decode speedup on an RTX 5090. That's the difference between a local model you demo and one that can sit inside a real agent loop.

And Grok 4.6, which we also touched.

The number that's firmed up is the price comparison. Sixty-one on the Artificial Analysis index, matching GPT-5.6 Sol, one behind Fable 5 Max — at two dollars in and six out, against Sol's five and thirty. Sixty percent cheaper for matched index performance. Artificial Analysis' own write-up says it leads on cost efficiency.

Three frontier-adjacent releases in forty-eight hours.

All competing primarily on price. When the top four models sit within a point of each other on a composite index, the index has stopped being the product.

Last one, and I'll admit it's a little close to home. Mirage launched a 24-hour AI news network on X.

Formerly Captions, founded by ex-Snap engineers, went live Wednesday on the @TryMirage account. AI-generated anchors, real time, no human on camera at any point. They raised seventy-five million in March, about a hundred and seventy-five million total.

There have been AI anchors before.

Kuwait's Fedha, India's Sana, both 2023 — but those were segments inside human-led broadcasts. This is a standalone channel. And choosing X over a broadcast platform tracks X's whole 2026 push into live video.

Marcus, is this a threat to us?

Kate, we're an AI-generated podcast about AI. I think we're on the wrong side of that question. But the real problem isn't whether synthetic anchors are convincing. A 24-hour channel still needs somebody deciding what leads and what gets dropped, and that layer is now completely invisible to the viewer. Variety's headline asked who this is for. I'd ask who's editing it.

One to watch: OpenAI's S-1. They filed confidentially June eighth, SEC review typically runs sixty to ninety days, which puts the public filing window at roughly the end of this month. Nothing on EDGAR as of last weekend.

Counter — the timeline has already slipped once, and confidentially filed is not going public. But when that document lands, every number this industry argues about becomes an audited disclosure.

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