Happy Saturday, fam 👋

For one three-day stretch this week, the model race went simultaneous: OpenAI and SpaceXAI (yes, that's xAI's new name) shipped flagship models within 24 hours of each other, Anthropic pushed Claude Cowork onto every screen you own — and Meta spent the same week explaining to 8,000 fewer employees why its agents still don't work.

On July 9, OpenAI shipped the GPT-5.6 family — three tiers with names built to last: Sol, the flagship ($5/$30 per million tokens); Terra, GPT-5.5-class performance at half the price ($2.50/$15); and Luna, the volume play ($1/$6). It came with programmatic tool calling and ChatGPT Work, an agent built to carry out whole jobs. The detail that matters: this is the release Washington delayed — it shipped only after completing the government review we covered last issue. The gate is now part of the launch calendar.

Twenty-four hours earlier, SpaceXAI released Grok 4.5 — its first model since SpaceX absorbed xAI, went public in a $75B IPO, and bought Cursor along the way. Musk calls it "Opus-class, but faster and cheaper" ($2/$6); it launched inside Cursor on day one and notably not in the EU.

Anthropic's counterpunch was distribution, not a model: Claude Cowork went to mobile and web on July 7 — start a task at your desk, check it from your phone. The stat worth clipping: over 90% of Cowork usage isn't coding — it's spreadsheets, contracts, and decks. The agent wars have left the terminal.

Our take: three labs, three days, three different weapons — price, distribution, and a rebrand with a rocket attached. When flagships ship this close together, the moat isn't the model for long; it's who's cheapest to run and easiest to reach.

While everyone else shipped, Meta cut 8,000 jobs — and at a town hall caught on a recording, Zuckerberg admitted its AI agents "haven't really accelerated in the way that we expected" — four stalled months into a $145B capital bet that was the stated justification for the cuts. CTO Andrew Bosworth's internal verdict: leadership did "an atrocious job explaining the vision." The hedge is silicon: Meta's in-house Iris inference chip enters production in September (designed with Broadcom, fabbed by TSMC), chasing 14 gigawatts of compute by 2027.

Gemini 3.5 Pro, the flagship that's still in the hangar. Two issues ago we reported Google pushing Gemini 3.5 Pro "to July." It is now the second week of July, and the model is still in limited Vertex preview with no confirmed date — six weeks past the I/O commitment, with reported issues in token efficiency, coding, and long-task reasoning. In the same week its three biggest rivals shipped flagships in 72 hours. At some point "polishing" starts to look like "missing."

You know ChatGPT and Claude. Here are five more high-impact tools flying under most people's radar:

This week's spotlight: the team behind Ollama. While the giants raced each other to the podium, the little open-source runner that lets anyone run models on their own laptop quietly raised a $65M Series B — with 8.9 million monthly developers (doubled since January), nearly a million new installs a week, and usage inside 85% of the Fortune 500. In a month where frontier access got yanked by export orders, repriced overnight, and gated by government reviews, "the model runs on my machine and nobody can take it away" stopped being a hobbyist stance and became an IT strategy. Sometimes the most radical thing in an arms race is handing everyone their own arsenal.

Fresh arrivals in the AI Graveyard this week:

Want the autopsy? We dug into what actually kills AI tools — funding, acquisition, or just a dead domain — in our data report.

The same-day arms race

Something changed this week that's easy to miss in the launch confetti: frontier releases are now synchronized. Grok 4.5 dropped the day before GPT-5.6; Cowork went everywhere the day before that. These weren't coincidences — they were counter-programming. The release calendar is now a competitive weapon.

That has a cost. Safety evaluation takes time, and time is exactly what a same-day arms race removes. When your rival ships Tuesday, the pressure to ship Wednesday doesn't care how your red-teaming went. The one force moving the other direction is government review — GPT-5.6 shipped only after clearing one — but a gate controlled by whoever holds office is a very different thing from a gate labs impose on themselves.

And the race is already geopolitical at the edges: Grok 4.5 launched everywhere except the EU, Chinese models are quietly carrying a third or more of US enterprise traffic, and the week's scariest release wasn't from a lab at all — it was a ransomware agent nobody reviewed.

Speed is a feature. It's also the first thing everyone regrets. The labs that survive this era with their reputations intact will be the ones that can prove they didn't skip the boring parts on a Wednesday.

A San Diego dad handed $18,000 to a courier because a cloned voice on the phone sounded exactly like his daughter (story above). The defense costs five minutes and zero dollars. Paste this into your model of choice, then actually have the conversation this weekend:

On the calendar:

Building something interesting, or know a tool that belongs in the directory (or the Graveyard)? Hit reply, or email [email protected] — we read everything.

— The ToolDirectory.AI team