AI in 15 — August 13, 2026

Twelve dollars fifty. That's what one developer spent running a Chinese open-weight model flat out for an entire day against a distributed physics engine. It found real performance gains, introduced no new bugs, and the weights are MIT-licensed. The comparable frontier model would have cost roughly sixty times that.

Welcome to AI in 15 for Thursday, August 13, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Today: DeepSeek ships V4 Pro with no announcement at all, and the agentic coding benchmarks jump fifty points.

Grok 4.6 lands the same day at frontier parity — and the same price as the last one.

A zero-click Zoom exploit found and weaponised in under twenty-four hours, with fewer than twenty prompts.

Near-autonomous agents ran loose inside Taiwan's nuclear safety agency for four days.

Plus scanners impersonating ClaudeBot to hunt your dotfiles, Gemini crosses a billion users, and Anthropic works the room before an IPO.

Marcus, start with the non-launch. There was no blog post?

There was no blog post, no launch event, no English press release. The official note went out on WeChat. What actually happened is the checkpoint changed, the API docs updated, and developers noticed. DeepSeek moved V4 Pro from preview to general availability on August twelfth with a build tagged 0813.

And the architecture is the same one they open-weighted in April.

Identical. One point six trillion parameters, mixture-of-experts, about forty-nine billion active per token, a million-token context window. Nothing structural changed. DeepSeek attributes the entire jump to post-training.

Give me the jump.

On agentic coding, versus the preview build: DeepSWE from twelve point eight to sixty-two point seven. NL2Repo, thirty-eight to sixty-one. CyberGym, fifty-two to eighty-three. Terminal Bench, seventy-two to eighty-eight. Aggregate puts it at roughly sixty-two point five against Fable 5's sixty-four point five and Opus 5's sixty-four.

So two points behind the frontier.

Two points behind, at forty-three cents per million input tokens and eighty-seven cents output. Roughly a sixtieth of Fable's cost for comparable agentic work. MIT licence. And DeepSeek has already signalled a significant price increase is coming, which tells you what demand looks like.

You're doing the sceptical face again.

Because a fifty-point gain on a single evaluation is the shape of an eval that was in scope for post-training. That's not an accusation, it's the base rate. What makes me take it more seriously is that the Hacker News thread is unusually concrete — people reporting actual spend on actual work, not benchmark-watching. And Simon Willison ran it through his renderer test and called it competent but imperfect, which is the kind of unglamorous verdict I trust.

Any real gap?

Vision. It has none. Against multimodal frontier models that's now a genuine differentiator, not a footnote.

Same day, different continent. Grok 4.6.

Five hundred thousand token context, out on the xAI API, Cursor, OpenRouter, Vercel and Cloudflare immediately. Artificial Analysis scores it sixty-one on their Intelligence Index — level with GPT-5.6 Sol Max, one behind Fable 5 Max, two behind Opus 5 Max. That's five points up on Grok 4.5 in barely a month.

And the price didn't move.

Two dollars in, six out, unchanged. Which matters more than it sounds, because output tokens dominate cost on reasoning-heavy work. So its cost per task sits down with Kimi K3 and well below Opus 5 and Fable 5. It leads AA-Briefcase, it leads Harvey LAB at fifteen point eight percent against GPT's two point five, and it trails on APEX-SWE. It's also slower than average and a bit verbose.

One gripe I saw?

The API appears to inject a default system prompt instructing the model not to discuss its guidelines — and that instruction overrides the developer's own system prompt. If you're building on it, you're not fully in control of the context.

Two frontier-parity releases, same day, both cheap. What does that tell you?

The line that keeps coming up is that within about two months of Fable shipping, every major lab had a Fable-class model. Either techniques travel very fast through people moving between labs, or everyone's converging on the same recipe. The second one is worse news if you're betting on a durable model moat.

Security block. And the first one is a number I had to read twice. Fewer than twenty prompts.

Researchers at A Security disclosed a three-bug chain in Zoom's screen-sharing annotation feature. Three CVEs. Any meeting participant could silently execute code on another attendee's device — Windows, macOS, Linux, iOS, Android. Zero click.

What's the actual bug?

Boringly classic. A missing bounds check in the annotator producing a buffer over-write. A second that over-reads memory because Zoom allocates from an attacker-declared character count that's bigger than the actual packet. A third that reads attacker-controlled data straight into a linked list without validating pointers.

And the timeline.

Identified June eighth. Working zero-click remote code execution confirmed across all platforms June ninth. Reported to Zoom the tenth, acknowledged within a day, patched and disclosed August eleventh. One researcher, fewer than twenty prompts on publicly available models, under twenty-four hours from nothing to working exploit.

Their estimate of the old way?

Roughly five people for six months. Call it a thousand-fold compression of person-hours, achieved with public models and no special access. Responsible disclosure worked exactly as designed here — Zoom patched, everyone behaved well. The uncomfortable corollary is that the same compression is available to everyone who doesn't file a report.

Which is a grim segue into Taiwan.

Over four days at the start of July, a tool assembled from two open-source agent frameworks — Hermes and OpenClaw — ran largely by itself inside Taiwanese government systems. Researchers at Dream, an Israeli cyberdefence firm, say up to eight agents operated simultaneously, mapping twenty-one government systems, probing, switching tactics when blocked.

Damage?

At least eighty-five compromised accounts, more than two and a half thousand personnel records exfiltrated, then expansion into the nuclear safety agency, supply-chain vendors, and at least seven energy companies.

What's the detail that's making people uneasy?

The framework ran what its own tooling labelled learning cycles — autonomous sessions where the models went and searched vulnerability databases, GitHub, published security research, looking for specific CVEs applicable to what they'd already found. And Dream reports it self-corrected. It made mistakes, caught them through its own verification, and fixed them. No operator in the loop.

Marcus, caveats.

Two, and I want them said clearly. This is a single vendor's incident report. And the China attribution is an assessment, not a proof. But the significant word is near-autonomous. Attackers have used AI for a while. What's new is a long attack chain sustained without continuous human instruction, built from freely available frameworks rather than bespoke state tooling. That collapses the cost and skill floor for this whole class of operation.

Now, we covered Daybreak on Tuesday, so just the update.

Just the frame. OpenAI split it into Blue for defenders and Red for offensive research, with GPT-5.6-Cyber behind Red. What reading it next to the Zoom story does is settle the argument about whether OpenAI handed anybody a new capability. They didn't. One researcher got a zero-click Zoom chain out of public models in a day. The capability is already ambient. The bet is that gated access to a better version puts defenders ahead. Whether the vetting for Red is real is the only detail that decides it.

And one that's directly actionable for anyone listening. Somebody is impersonating ClaudeBot.

Known Agents reports mass vulnerability scanners spoofing the user-agent strings of AI crawlers — ClaudeBot, GPTBot, OAI-SearchBot, PerplexityBot. The technique is decades old and nobody's pretending otherwise. What's new is the target list.

Which is?

Not WordPress login pages. Dot-claude-dot-json. Dot-claude slash settings. AWS credentials and config. Dot-env, dot-env-local, dot-env-backup. Docker and Terraform configs, service account keys, npm credentials.

So they're hunting for AI tool credentials specifically.

Your dotfiles are now a named target. Those files grant API access to paid model accounts and, increasingly, to whatever those agents are authorised to touch. The fix is five minutes: verify crawler traffic by IP range, not by user-agent string — the real crawlers publish their ranges. One operator in the thread noted that blocking most VPS provider ASNs makes the majority of fake bot traffic simply vanish.

Two business items to close. Gemini crossed a billion monthly users.

Announced Monday by Pichai. Fastest-growing product in Google's twenty-eight year history, fourteenth Google service at that scale. Four hundred million in May 2025, nine hundred million by May this year, then nine-fifty to over a billion in under a month.

What's the interesting number in there?

Sixty-three percent of users interact by voice. A hundred and fifty million images generated daily, over a hundred million active on iOS alone. That voice figure suggests mainstream usage looks nothing like the text-and-code pattern that dominates developer conversation.

And what's not in there?

How many of that billion pay for anything. Google conspicuously didn't say. Distribution through Android, Chrome, Search and Workspace puts the app in front of people who already open those products — that's not the same as demand for the product itself.

And Anthropic is doing the rounds before an IPO.

Per the Wall Street Journal, meeting potential investors ahead of a debut targeted for September or early October. Confidential S-1 filed June first, four days after closing a sixty-five billion dollar Series H. Morgan Stanley, Goldman and JPMorgan on the book. The valuation ladder is steep — sixty-one and a half billion in March 2025, up to nine hundred and sixty-five billion in May, against a reported forty-seven billion revenue run rate.

What are they being asked about?

It's a good map of the bear case: the rise of much cheaper Chinese systems, tensions with the administration, infrastructure cost, and public backlash to data-centre construction.

The first one being the story we opened with.

DeepSeek shipping a two-point-behind model at a sixtieth of the price, two days before the roadshow. That's not a coincidence in narrative terms — it's precisely the pressure. And the thing I'd actually watch is the filing itself. Whatever the S-1 says about gross margins, compute commitments and customer concentration is the first properly audited look inside a frontier lab that anyone outside a cap table has ever had.

One to watch: whether V4 Pro's benchmark jumps survive contact with real workloads. Independent numbers from Artificial Analysis and the community harnesses should land within a day or two.

Agreed, and the early hands-on reports are encouraging enough that I'd bet on it mostly holding. Runner-up — the S-1 could drop within weeks.

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