AI in 15 — July 10, 2026

Seventy-five-point-two percent, versus zero-point-seven. That's OpenAI's new voice model against the old one on a web-search test — a gap so wide it's basically a different product. And on the same day, OpenAI reorganized its entire app and half its own developers revolted.

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

And I'm Marcus, your co-host. The launch week we've been tracking finally landed all at once — three products in a single day, and a user backlash that's arguably the real story.

It really is, Marcus. Our lead: OpenAI ships GPT-5.6, GPT-Live, and ChatGPT Work together — with government sign-off and a first-ever benchmark milestone. Then a run worth your time.

Grok's first independent numbers are in, and the surprise is the price tag.

Meta does something it has never done — it starts charging for its own AI.

OpenAI tells everyone to stop trusting a coding benchmark it used to champion.

The New York Times asks a judge to punish OpenAI for allegedly hiding evidence.

China wants to build its own AI chips — and quietly buy Nvidia's anyway.

And a former Fed chair takes a seat guarding a frontier lab.

Lead story, Marcus. Yesterday we dug into GPT-Live and previewed GPT-5.6. Now it's all live. What actually shipped?

Three things at once, Kate, and that's the point. GPT-5.6 went generally available in three tiers — Luna, the cheap fast one; Terra, the everyday workhorse at roughly half the cost of the last model; and Sol, the frontier reasoner tuned for biology, chemistry, and cybersecurity. GPT-Live, that full-duplex voice model we covered, is now on iOS, Android, and web. And the genuinely new one — ChatGPT Work.

Give me the headline capability first. There's a benchmark milestone here.

There is, and it's a real one, Kate. GPT-5.6 Sol became the first verified frontier model to actually beat an ARC-AGI-3 game — setting a new state of the art at seven-point-eight percent, per the ARC Prize folks. Now, seven-point-eight percent sounds tiny, but ARC is designed to be brutal — it tests genuinely novel reasoning, the stuff models usually can't fake. Being first past that bar is a legitimate marker. And notably, the Department of Commerce signed off on the broad launch after extra safety testing — that government review we've been tracking all week actually cleared.

So the review didn't stall it. Good. Now — ChatGPT Work. What is it?

It's the agent, Kate. You give it an outcome, not a prompt — "build me the quarterly sales deck" — and it goes and gathers context across your connected apps: Slack, Teams, Google Drive, email, your CRM. It breaks the goal into steps and comes back hours later with finished spreadsheets, slides, docs, even working web apps. OpenAI says five million people already use its Codex coding tool weekly, and this folds that ambition into everyone's account, free tiers included.

That sounds genuinely useful. So why the revolt you mentioned?

Because they shipped it alongside a total app reorganization, Kate. The separate Codex app got merged into one unified ChatGPT app — Chat, Work, and Codex all in one place. The old app got renamed "ChatGPT Classic," and the Atlas browser is being sunset. And developers on Hacker News were openly hostile — that thread hit three hundred and thirty-two points. The complaint: a workflow that worked fine got reshuffled, and casual chat now lives in what one person called an "unsearchable popup."

So the capability races ahead and the product gets more confusing at the same time.

That's the tension exactly, Kate. This is OpenAI trying to own the entire stack in one day — the smartest model, the most natural interface, and the agent that does the work. Technically it's an impressive flex. But the people who use these tools most got the least warning, and the loudest voices in the room today weren't cheering the ARC milestone — they were annoyed they had to relearn where their chat history went. Capability and usability are pulling in opposite directions.

Story two, Marcus. Yesterday's "one to watch" was Grok 4.5's independent numbers. They're in.

And they're interesting, Kate. Artificial Analysis puts Grok 4.5 fourth on general intelligence — behind Fable 5, Opus 4.8, and GPT-5.5. So "Opus-class" on general smarts? Not quite; it sits a notch below. But flip to coding and it jumps to tied-for-second on the Coding Agent Index, one point off Fable 5. So the specialist framing held up.

But you said the surprise is the price.

That's the actual story, Kate. Artificial Analysis ran the same coding task on three models: two dollars forty-nine on Grok, five dollars seven on GPT-5.5's Codex, and eleven-eighty on Fable 5 running Claude Code. That's roughly an eighty percent cost cut for near-frontier coding. When capability commoditizes, price becomes the battlefield — and Grok just undercut everyone.

So is it a giant-killer? Keep me honest.

I'll temper it, Kate. Those are "high" reasoning-effort scores, which cost more compute than the sticker suggests. Hallucination rates run higher than the leaders. And a coding-agent composite isn't the same as everyday reliability across every task. It's also not in the EU yet — expected mid-July. So: genuinely cheap, genuinely good at code, not yet the all-rounder. But at a fifth of the price, "good enough" reframes the whole race.

Story three, Marcus, and it's a real identity shift. Meta starts charging for AI.

For the first time ever, Kate. Meta Superintelligence Labs, under Alexandr Wang, released Muse Spark 1.1 — a multimodal reasoning model with a one-million-token context window for agentic tasks. But the real pivot is the paid Meta Model API, now in public preview. Consumers still get it free in the app, but developers now pay: a dollar-twenty-five per million input tokens, four-fifty output, with twenty dollars of free credit to start.

Why is that such a big deal? Everyone charges.

Because Meta's entire model was the opposite, Kate. Their whole strategy was giving AI away — open weights, free, undercut everyone by making the technology a commodity. Now they're metering it like OpenAI and Anthropic do. That's a strategic about-face for the company that made "free" its weapon.

So has Meta given up on the open playbook?

Not necessarily, and this is the sharp read from Hacker News, Kate. Meta doesn't need to match OpenAI or Anthropic on revenue. If it keeps shipping cheap, competitive models, it can deflate their pricing power — bleed the margins of the labs charging premium rates. A week ago the story was "OpenAI and Anthropic are irreversibly ahead." Between Meta charging and Grok undercutting, that narrative just got a lot messier. And Simon Willison already built a plugin for it during preview — the developer community moves fast.

Story four, Marcus. OpenAI audited a popular coding benchmark and told everyone to stop using it.

One it used to recommend, Kate — that's what makes it notable. OpenAI audited SWE-Bench Pro, a widely-cited agentic coding test, and found somewhere between twenty-seven and thirty-four percent of its tasks are broken. Their automated pipeline flagged two hundred tasks; five human engineers flagged two hundred and forty-nine. The problems: tests so strict they reject correct code, prompts too vague to solve, and low-coverage tests that pass incomplete fixes.

And they publicly walked back their own recommendation?

They did, Kate — retracted it outright, saying the benchmark's hit a noise ceiling around seventy percent as models jumped from twenty-three to eighty percent pass rates in eight months. Once a third of your test is faulty, a few points of difference between models is meaningless.

Which lands in a funny week, doesn't it. Everybody's quoting benchmarks.

That's the whole reason it matters, Kate. This same week, three labs are all touting benchmark wins — Grok's coding index, GPT-5.6's ARC score, Meta's numbers. And here's OpenAI saying, in effect, the referee might be broken. When the scoreboard is suspect, every "model X beats model Y" headline deserves a raised eyebrow. It's a healthy reliability check on the entire news cycle we're standing in the middle of.

Story five, Marcus, and this one's serious. The New York Times wants a court to sanction OpenAI.

It's a big escalation, Kate. A newspaper group led by the Times asked a Manhattan federal court to sanction OpenAI, alleging it lied about being unable to search its own systems for copyrighted articles. The plaintiffs say OpenAI had already run those internal searches, built a database of some seventy-eight million de-identified conversations to gauge its own infringement — and then deleted billions of ChatGPT outputs after the suit was filed, allegedly violating a preservation order.

Deleting evidence after a lawsuit starts — that's a different category of problem, isn't it.

Completely different, Kate. This isn't about copyright anymore — it's about discovery conduct. And courts punish hiding or destroying evidence far more harshly than the underlying claim, because it strikes at whether the process can function at all. OpenAI flatly denies it — their spokesperson says the Times is just trying to access private user conversations as its case weakens. So treat the allegations as allegations.

But if they're proven?

Then it's a generational mistake — that's the blunt read circulating, Kate, and I think it's fair. A lab can lose a copyright fight and survive. Losing the judge's trust on whether you played straight with the court is the kind of thing that shapes how every AI company gets treated in litigation for years. This is the one to actually watch on the legal side.

Story six, Marcus. Two China compute stories that mirror each other. DeepSeek's building a chip.

An inference chip specifically, Kate — not training. DeepSeek is reportedly reaching out to chip-design, foundry, and memory partners, quietly hiring silicon engineers for about a year now. The goal: cut reliance on both Nvidia and Huawei under US export controls.

Can they actually pull that off?

Big caveat, Kate — this is an early-stage report from anonymous sources, and "designing a chip" is a long way from manufacturing a competitive one. US curbs bar China from the most advanced foundries, and separate controls have choked access to the high-bandwidth memory that inference depends on. Designing is the easy half.

And here's the mirror image — Beijing may let them buy Nvidia after all.

Right, and this one's single-sourced, from The Information, so hold it loosely, Kate. Beijing is reportedly preparing to let Alibaba, ByteDance, and DeepSeek buy a limited batch of Nvidia's H200 chips — after months of resistance. Strings attached: training only, keep inference on domestic processors, and a total possibly capped under two hundred thousand units, less than half what they asked for.

So build your own silicon and still buy Nvidia. That's the tension.

In one sentence, yes, Kate. China wants independence and still needs Nvidia in the meantime. This reads as a pragmatic retreat forced by a real compute shortage — not a change of heart. The crunch is severe enough to bend policy that Beijing spent months defending.

Two lighter hits to close, Marcus. First — GLM 5.2 keeps having a moment.

It really does, Kate. Zhipu's open-weight GLM 5.2 sits near the top of the open-model rankings and rivals Opus on some coding benchmarks at a fifth of the cost. Two grassroots stories went viral this week: a developer's project called Colibri — five hundred-plus Hacker News points — that streams the model's weights so it runs on an ordinary slow computer. And a benchmark claiming GLM 5.2 does VAT bookkeeping nearly as accurately as a human.

A human bookkeeper. Really?

With a great piece of pushback, Kate — commenters noted the humans also did the harder job of actually finding the invoices in the first place. And the sharp question underneath: if the AI bookkeeper is wrong, who goes to prison? But the direction is unmistakable — frontier-adjacent capability is now cheap enough that people run it locally and point it at real accounting work.

And the human one, Marcus. Ben Bernanke joins Anthropic.

The former Fed chair, seventy-two, joined Anthropic's Long-Term Benefit Trust, Kate — the body that holds a special class of stock and can appoint board members, meant to steer the company toward safety over pure profit.

Does a central banker's name actually keep a frontier lab honest?

That's exactly the question people asked, Kate — and the reaction was mixed. Some read it as genuine credibility; others as loading the board with big names. It's a fair test of the whole governance idea: who's supposed to watch the watchers, and does a famous résumé accomplish it? I'd say prestige and oversight aren't the same thing — but at least someone's asking.

One to watch tomorrow, Marcus.

The GPT-6 rumor, Kate. Accounts claiming inside knowledge say OpenAI's next major model — freshly pretrained, possibly called GPT-6 — could land "within weeks," with xAI reportedly training a ten-trillion-parameter Grok. Treat every word of that as unverified.

Agree, or counter?

Agree, with my usual rule, Kate — watch what ships, not what leaks. AI timelines slip, and anonymous claims are free to make. Ask me when there's a model card.

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