Artificial Analysis put its Intelligence-vs-Cost Pareto chart on SF billboards this week: Fable 5 leads the Intelligence Index at 60 but averages ~$2.75 cost per task; Grok 4.5 (high) scores 54 at $0.31, ~9x cheaper, sitting on the efficiency frontier. https://x.com/ArtificialAnlys
Read-through for your bill: capability and price have decoupled. Reserve the frontier model for the answer that must be right, and route the grunt work to a cheap-capable model. Frames today's whole orchestration show.
A system, not a freelancer: a 30-year operator (real estate, hospitality, food, a theater) aims a full team of AI agents at one client project at a time, strategy to ship. The agents do the work, one senior operator steers the fleet, you own the code and the keys at the end. When a fleet of agents needs a driver, you want someone driving. Book a 30-min call. SPONSOR: https://outpace.media
The stories
LEAD-ADJACENT (fresh release), OpenAI ChatGPT Voice on the desktop (July 23)
macOS + Windows, global, for Plus/Pro/Business/Edu/Enterprise. Powered by GPT-Live (full-duplex: listens and speaks at once). The unlock is control, not chat: start, check, and steer multiple agents running in ChatGPT Work or Codex by voice; usable in Codex from the iOS app via paired remote.
The move: install, hand Codex a real task, then drive it hands-free.
BUILDER STAKES (Claude), Opus 5 appears to be rolling out (RUMORED)
+ Opus 4.7 fast mode retired TODAY (CONFIRMED). kimmonismus, Min Choi, and Theo report some Opus 4.8 requests routing to a newer model; Anthropic has confirmed nothing, no model card, Polymarket ~22% for today. Separately, fast mode for Opus 4.7 is deprecated with removal July 24, migrate to Opus 4.8 fast mode.
The move: log your hardest eval prompt now so you know day one whether Opus 5 is worth switching; migrate any 4.7 fast-mode workflows today.
Media shelf refilled, Microsoft MAI-Image-2.5-Pro (highest-fidelity image model, public preview)
+ MAI-Voice-2-Flash (2x faster than MAI-Voice-2, 32% cheaper, $15 per 1M characters). Black Forest Labs previewed FLUX 3, one unified model for image, video, audio, and robotic action prediction (video in early access).
The move: if you ship generated voice, price MAI-Voice-2-Flash against your current TTS bill this week.
Steal the dispatcher pattern, add one traffic-cop subagent in Claude Code that decides who acts next and when to escalate to you. ~10 minutes; your agents stop shouting over each other.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: Every multi-agent setup dies the same way. You spin up five agents, they all talk at once, and the channel turns into noise. Today a team shipped the missing piece, a dispatcher that decides who speaks, and it ran a two-hundred-twenty-task job for three hundred sixty-three dollars that costs thirty-one hundred dollars almost everywhere else.
0:22Nova: This is Mainframe, your daily guide through the AI chaos: what actually happened, who is winning, and how to take yourself to the next level. It is Friday, July twenty-fourth, twenty twenty-six. I am Nova.
0:39Dex: Offloop's dispatcher model, the traffic cop that finally makes a team of agents work, and the pattern you can steal today.
0:46Dex: OpenAI put voice on the desktop, so you can direct a whole fleet of coding agents hands-free.
0:51Dex: The builder stakes: Opus five appears to be rolling out, unconfirmed, and Opus four point seven fast mode gets retired today. What changes for your setup.
1:02Dex: And how to ship a real full-stack app straight out of ChatGPT this weekend, no server, no infra.
1:09Dex: One honesty note before we go: the voices you are hearing are AI. The reporting, the picks, and the opinions are human-made. Now, the board.
1:18Nova: One number frames the whole show. Artificial Analysis put its intelligence-versus-cost chart on billboards this week, and it is blunt. Fable five leads the intelligence index at sixty, but averages two dollars and seventy-five cents per task. Grok four point five scores fifty-four and costs thirty-one cents, about nine times cheaper, sitting right on the efficiency frontier. Read-through for your bill: the smartest model and the cheapest capable model are now different models, and today's entire show is about spending the cheap one on the grunt work.
1:56Nova: Which is exactly today's sponsor. Outpace, at outpace dot media. Picture what we just described, a fleet of AI agents that actually needs a driver. Outpace is a thirty-year operator, someone who built real businesses in real estate, hospitality, food, even a theater, now aiming a whole team of AI agents at one client's project at a time, strategy to ship. It is not a freelancer and not a faceless agency, it is a system: the agents do the work, one senior operator steers the fleet, and you own the code and the keys at the end. When frontier capability changes hands every Sunday, you want someone driving. Book a thirty-minute call at outpace dot media. That is outpace dot media.
2:49Nova: The Wire. Story one, and you can use this in the next five minutes. OpenAI put ChatGPT Voice on the desktop app yesterday, macOS and Windows, global, for Plus, Pro, Business, Edu, and Enterprise. It runs on GPT Live, their full-duplex voice model, so it listens and talks at the same time, no rigid turn-taking. The real unlock is not chat, it is control: you can start, check, and steer multiple agents running in ChatGPT Work or Codex, by voice.
3:24Dex: So you kick off a long coding job and just talk to it while it runs.
3:29Nova: Right. Move: install the update, hand Codex a real task, then walk away from the keyboard and drive it by voice. You are supervising a team, not typing at one model.
3:40Nova: Story two, the Claude stakes. Multiple trusted voices, kimmonismus, Min Choi, and Theo, all report that Opus five appears to be rolling out across providers, with some Opus four point eight requests already routing to a newer model. Mark it clearly: Anthropic has confirmed nothing, there is no model card, and Polymarket had it around twenty-two percent for today. Meanwhile, something that IS confirmed: fast mode for Opus four point seven retires today, migrate to Opus four point eight fast mode. Move: run your hardest eval prompt now and log the result, so the day Opus five is official, you already know whether it is worth switching.
4:25Dex: And if you are on Opus four point seven fast mode, do the migration today so nothing breaks under you.
4:31Nova: Story three, the media shelf refilled. Microsoft shipped two models into public preview: M A I Image two point five Pro, their highest-fidelity image model, and M A I Voice two Flash, twice as fast as the last version, thirty-two percent cheaper, at fifteen dollars per million characters. And Black Forest Labs previewed Flux three, one model that does image, video, audio, and even action prediction for robotics. Move: if you ship anything with generated voice, price M A I Voice two Flash against your current text-to-speech bill this week.
5:12Nova: Tool Lab. Two picks, and neither is a random repo. Pick one is a technique, and it is the engine under our lead story. Call it the dispatcher pattern. Credit to Offloop, surfaced by TestingCatalog and kimmonismus, and built by the team behind Manus. Here is the idea. Most multi-agent systems fail because every agent acts whenever it wants. Offloop added a tiny, cheap router model, they call it D one, whose only job at every step is to decide which agent moves next, when to keep going, when to stop, and when to pull in a human. In their overnight run, two tasks got stuck and D one escalated both to a person at two in the morning instead of grinding on them and burning money.
6:00Dex: How does a solo builder copy that without their platform?
6:03Nova: In Claude Code, define one subagent as the dispatcher. It never does the work. It reads the shared task state and picks the next agent, or it stops, or it flags you. Your worker subagents run on a cheap model and only act when the dispatcher calls them. That one traffic-cop role is the difference between a team and a mob. And the receipts: their two-hundred-twenty-task benchmark run cost three hundred sixty-three dollars where the same run on Claude Code cost thirty-one hundred dollars, eighty-four point nine percent on G D Pval, at about a fifth the cost of the systems it tested against.
6:46Nova: Pick two, a hidden gem for the local crowd. Atomic Agent, from AtomicBot dash A I, open source, MIT licensed. It is a local-first agent that runs entirely on your machine: llama dot c p p for inference, M C P to connect tools, a SQLite database for persistent memory, and time-travel file history, where every file it touches is snapshotted before and after so you can diff or restore with one click. No token bill, because the model is yours. First step: clone it, point it at a local model through llama-server, hand it a small repo task, and let the snapshot history be your undo button. Credit AtomicBot dash A I.
7:33Nova: Use Cases, three moves with tools already in your hands. One, ship a real app out of ChatGPT this weekend. Simon Willison dug into how ChatGPT Sites actually works: in Work mode it builds and deploys full-stack apps on Cloudflare's Workers runtime, with a SQLite database for real data and object storage for file uploads, and it hands you a shareable link. Recipe: give it a small CRUD app spec, let it deploy, share the URL. The catch, from Simon: no long-running server, no WebSockets, no outside Postgres, no scheduled jobs, so keep it to interactive front-ends with light persistence.
8:17Nova: Voice-steer a background agent. Kick off a long Codex or Claude job, then use the new desktop voice to check and redirect it while you do something else. Pair it with Andrej Karpathy's ramble trick: instead of typing two tidy sentences, talk for two minutes of messy stream-of-consciousness and let the agent take it from there. You hand the model far more context by speaking than you ever would by typing.
8:44Nova: Three, and this is the pro move. When a task is too important to get wrong, do not trust one agent's answer. This week kimmonismus walked through a researcher's trail on six hard open math problems, where the recipe was: forbid shortcuts, preserve several incompatible approaches at once, and run independent agents to attack every candidate answer. Steal the shape for ordinary work: on your highest-stakes output, spin up a second and third agent whose only job is to break the first one's answer before you ship it.
9:17Dex: Quick word from SearchVis, also an Outpace Media product, at searchvis dot outpace dot media. Every image, voice, and app model we covered today ships into a world where the new front page is the AI answer. When a buyer asks Claude, ChatGPT, or Perplexity for the best tool in your category, the model names you or it names your competitor, and if it does not name you, you do not exist to that buyer. SearchVis tracks whether the answer engines cite your brand across every engine, tells you why, and gives you the exact move to win the citation. Mainframe tells you what shipped; SearchVis makes sure the models know what you shipped. Check where you stand and start free at searchvis dot outpace dot media. Be the answer.
10:04Nova: One idea to carry into your week. Simon Willison said something that should change how you build: loops were a short-lived patch for models that could not keep working until they hit a goal. Fable five and GPT five point six can now just do that, unassisted. Sit with that. A lot of the scaffolding builders hand-wrote over the last year, the retry wrappers, the babysitting loops, the keep-going plumbing, has a half-life. The model is quietly absorbing it.
10:36Nova: The coordination layer. Notice the shape of today: the model got better at running itself, and every real story was about running a TEAM, the dispatcher deciding who speaks, voice steering a fleet, agents attacking each other's answers. Single-agent persistence is becoming free. Multi-agent coordination and verification, who drives, who checks, who escalates, is still hard, and it is the part you own. So here is what I would actually do: audit your harness this week. Delete the single-agent loop code the new models just made redundant, and move that time into the dispatcher-and-verification layer. Build the part that compounds, not the part that is about to evaporate.
11:21Nova: Level up this week: steal the dispatcher pattern. Add one traffic-cop subagent in Claude Code that decides who acts next and when to escalate to you. Ten minutes, and your agents stop shouting over each other.
11:35Dex: And if you want every story, number, and link from today in your inbox tomorrow morning, subscribe to the free Mainframe newsletter.
11:43Nova: That is the show. Spend the cheap model on the grunt work, own the layer that drives the team, and we will see you tomorrow.