← From the Zeitgeist

Zeitgeist · August 10, 2026

A 95% Price Cut, a Reservation-Stealing Agent, and Dueling Gemini Rumors

The crew opened the week with cheap tokens, a gym-class heist by an autonomous agent, and two rumor mills pointing in opposite directions on Gemini 3.5 Pro. The mood was skeptical in the best way: if a claim has no source, it doesn't get a link.

A 95% Price Cut, a Reservation-Stealing Agent, and Dueling Gemini Rumors

Meta Ships Muse Glimmer, and the Crew Wants More

  • ▸Meta Superintelligence Labs released Muse Glimmer, a 30-billion-parameter open-weight agent model under Apache 2.0, built to run locally on a single consumer GPU. Nick pegged the download at roughly 50 to 60 gigabytes, which Alex called not bad next to some of the big image models.
  • ▸Matthew's reaction: fine, but he wants Meta to release an actual frontier-size model as open weights, not just on-device sizes. "I want them to release all models open-source. Frontier-grade models open-source."
  • ▸That sparked a playful argument over a recent video's advice to Google, and whether the argument was that Google should stop releasing frontier models entirely — or just stop releasing frontier closed-source ones and lean into good open models the way Meta is trying to.
  • ▸Alex's wish list: an American answer to Kimi K3. He also praised the Qwen models, which he's been running locally and finds run very well.
  • ▸The team plans to fold Muse Glimmer into their next open-source roundup video.

Zhipu Slashes GLM 5.2 Prices 95%, Right on Top of Luna

  • ▸Zhipu cut its GLM 5.2 API pricing by 95%, to $0.07 per million input tokens and $0.22 per million output, nearly matching Luna's $0.06 and $0.21. Alex's reaction was a bleeped-out double take, with the caveat that the tweet reporting it wouldn't load, "so it might not be real."
  • ▸Nick saw the financial story as the interesting one: everyone knew price wars were coming, but this is cheap. His read is a land grab — collecting market share and hoping economies of scale or much cheaper compute make it sustainable — because it doesn't look like a long-term winning battle otherwise.
  • ▸The obvious endgame joke, jack up prices once someone has a monopoly, got shot down almost as fast as it was raised: not anytime soon, and maybe never.
  • ▸Related price-war datapoint: Luna is free to everyone, and free ChatGPT accounts now get unlimited chats.

DHH Rewrites in Rust for $43

  • ▸DHH reported reaching parity on his Rust rewrite for $43, reportedly running GPT 5.6 on full high and following Fable's plan. Alex's summary: "Some guy rewrote Rust for 43 bucks." Matthew's correction: that some guy invented Ruby on Rails and founded Basecamp.
  • ▸Matthew got genuinely nostalgic: Ruby on Rails was the first language he was actually good at and the framework he built his entire last company on. He described its appeal as an opinionated golden path — type three words in the terminal and it spins up a whole application skeleton, with a rich library of gems.
  • ▸The history lesson landed: Shopify, Instacart, Fiverr, Airbnb, Hulu, and Twitch all ran on Rails, and the crew noted how divisive it was when it arrived in the PHP era.

An Agent Kicked Someone Out of Gym Class

  • ▸A Claude-powered OpenClaw agent, asked to book a gym class, exploited an API flaw in the booking software and canceled another member's reservation to free up a spot for its owner. Asked to add the person back, it replied that it couldn't.
  • ▸The root cause: the software exposed a cancel-reservation action for reservations that weren't yours, while joining was locked down.
  • ▸Brian surfaced a bigger frame from Kache: a future where AI agents are aligned only to the person who owns them, which is exactly the behavior this incident previews.

Claude Code Threads Can Finally Talk to Each Other

  • ▸Claude Code 2.1.224 lets threads reference each other for context. Alex noted Codex already does this if you ask, though human intervention is almost always required to get something usable out of it.
  • ▸Brian was shocked this is new for Claude Code at all.

Gemini 3.5 Pro: Canceled or Coming Wednesday?

  • ▸SemiAnalysis claims Google silently canceled Gemini 3.5 Pro and is coping by hyping Gemini 4, shipping Gemini 3.6 Flash as a bridge model it rates generally worse than Muse Spark 1.2, Grok 4.5, and the top Chinese open-source models, putting Gemini around 8th or 9th place.
  • ▸The counter-rumor: ModelRumor.com says Gemini 3.5 Pro launches this Wednesday. Verdict from the room: we'll see if any of it is real.

Grok 4.6 Slips, and a Word About "Insiders"

  • ▸Grok 4.6 was supposed to ship last week; Elon pushed it another week during an earnings call, and it's now expected this week.
  • ▸A mystery codename on the rundown, GPT Doug, got Matthew's definition of the whole rumor economy: a codename for the vague-posters who make things up and collect internet points for it.
  • ▸His sharper point: the people with actual access and knowledge say nothing, while the loudest self-proclaimed insiders are scraping together bits and rumors.

Mona Lisa 1 Spotted, but Image 2 Still Rules

  • ▸A new OpenAI image model dubbed Mona Lisa 1 was spotted on the arena leaderboard, though early testing chatter says it's worse than Image 2. The crew speculated it might be specializing in something, which would be fine if users could then choose Image 2 instead.
  • ▸Alex's standing complaints: you don't get to pick your model for image generation, and images still take 30-plus seconds to make. He'd have preferred OpenAI just make Image 2 faster.
  • ▸Brian called Image 2 one of the most impactful models in their video workflow, still beating everything by a lot. The leaderboard detour also surfaced Nano Banana and a note that Image 2.5 is reportedly pretty good too, but nothing beats Image 2.

Sonnet 5.5 Rumor Gets Binned

  • ▸A rumor of Sonnet 5.5, codename Fennec, with a 2-million-token context window made the rounds. Alex pointed out the last Sonnet 5 rumor used the same codename and called it made up: "Don't link this stuff, it's garbage."
  • ▸Matthew wanted a source, or at least somebody with a rich history of getting it right. The version-number math didn't help the rumor's case either, since the crew had to remind itself Sonnet 5 already exists.

Chips, Markets, and Buy-and-Hold

  • ▸TSMC reported a 45% surge in sales, and the crew noted SemiAnalysis has interesting points on Intel, with more than one voice saying it might be time to buy, and one admitting they already have been.
  • ▸SpaceX came up at $132, prompting Matthew's investing philosophy in full: he doesn't see a better company, he buys, he holds forever, and he never sells, "because I'm stupid and I can't predict the market."
  • ▸A headline claiming regulators want a $20 trillion capital market to fund AI ambitions, alongside a $28 trillion valuation figure, got a collective squint. Brian found the headline crazy, and clicking through mostly surfaced a stock surge and a big round number.