GPT-5.6 shipped with a real price ladder: Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per M tokens. Luna nearly matches GPT-5.5's peak at under half the cost; Terra surpasses it at lower cost.
Independent read: Fable 5 still leads the Artificial Analysis Intelligence Index (59.9 vs Sol 58.9), but the frontier premium is collapsing across OpenAI, Anthropic, and Grok.
Operator takeaway: price per finished task, keep a routing dial, bank the spread. Do not standardize on one vendor this week.
A system, not a freelancer: a 30-year operator (real estate, hospitality, food, a theater) directing a full team of AI agents at one client project at a time, strategy to ship. You own the code and the keys. When the shelf reshuffles three times a week, you want someone steering the whole fleet. Book a 30-min call. https://outpace.media
The stories
LEAD (builder impact)
— GPT-5.6 goes fully public (July 9), and the real news is the API surface. Programmatic Tool Calling (Responses API) lets the model write in-memory JavaScript to coordinate tools, filter intermediate data, and pass only what matters forward — cutting round trips and tokens. Multi-agent beta runs concurrent subagents and synthesizes in one request. Receipts: PlayCo saw 63.5% fewer total tokens and 50.1% fewer model turns with PTC. Simon Willison: it "could help bridge the gap between MCPs and full terminal sessions." CAVEAT: PTC only wins when control flow is predictable (fetch/filter/transform/aggregate); if next step depends on the last result, keep direct tool calling.
The move: re-implement one predictable agent workflow with Programmatic Tool Calling on Terra/Luna; log tokens before/after.
The desktop merge — Codex folds into the new ChatGPT desktop app (Mac + Windows)
Codex is NOT gone: it's a dedicated coding mode alongside Chat and Work; old app renamed ChatGPT Classic. New Codex bits: inline diff editing, PR review in the side panel, faster GPT-5.6 Computer Use, multi-repo projects. Builder pick: use Codex mode (shows technical detail Work abstracts away). If you had Codex installed, update it — it becomes the new app.
Included on all paid plans through July 12 (50%-of-weekly cap). After: usage credits at $10/M in, $50/M out (>2x Opus). If credits off when allowance runs out, access stops. NOT permanent; restore to subscriptions promised "as capacity allows."
The move: distill the method, not the output (see Use Cases).
Re-implement one predictable agent workflow with Programmatic Tool Calling on Terra/Luna, and distill one Fable skill before July 12 (~20 min). Every run after is cheaper and outage-proof.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: OpenAI just shipped the whole superapp: GPT five point six, Sol, Terra, and Luna, plus ChatGPT Work, all landing at once. And the piece that matters at your keyboard is not the model, it is a new API trick that turns a dozen expensive model round trips into one program run.
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 tenth, twenty twenty-six. I am Nova.
0:38Nova: On today's Mainframe: GPT five point six goes fully public, and the real gift for builders is Programmatic Tool Calling and a multi-agent beta in the API. ChatGPT Work and Codex fold into one desktop app, and I will tell you the one mode you actually want. Fable five, the model you plan with, is still on borrowed time: your included window closes July twelfth, and after that it meters at ten dollars in, fifty dollars out per million. And a hidden gem: Tencent's Hy three, a two hundred ninety-five billion parameter open model that is free on OpenRouter for eleven more days. Let's get into it.
1:27Dex: Quick note before we roll: the voices you are hearing are AI. The reporting, the picks, and the opinions are put together by humans.
1:39Nova: The Board, and I am keeping it short. Here is the one macro move that hits your bill. GPT five point six landed with a real price ladder. Sol is five dollars in, thirty out; Terra two-fifty in, fifteen out; Luna one dollar in, six out per million tokens. Luna nearly matches GPT five point five's peak performance at less than half the estimated cost, while Terra surpasses it at a lower cost.
2:11Dex: And on the independent numbers, it is close at the top. Fable five still leads the Artificial Analysis Intelligence Index, fifty-nine point nine to fifty-eight point nine. But the cost story flipped hard against premium.
2:26Nova: So the read for your margins: near-frontier is now a three-way price war, OpenAI, Anthropic, and Grok. If your pricing assumes one vendor, that is a leak. Price per finished task, keep a routing dial, bank the spread.
2:45Nova: Which is exactly the kind of thing our sponsor lives in. Brought to you by Outpace. Picture an operator who spent thirty years building real businesses: real estate, hospitality, food, even a theater. Now that same operator points a whole team of AI agents at one client's project at a time, strategy to ship. Outpace is a system, not a freelancer, and not some faceless mega-agency: the agents do the work, one senior operator steers them, and at the end you own the code and the keys. When the model shelf reshuffles three times a week like it did this week, you want someone driving the whole fleet, not chasing each release. Book a thirty-minute call at outpace dot media. That is outpace dot media.
3:41Nova: The Wire. Lead story, and this is the one to act on. GPT five point six is fully public today across ChatGPT, Codex, and the API. But skip the benchmark theater; the real day-one news is the API surface. In the Responses API, Programmatic Tool Calling lets GPT five point six write and run programs in-memory that coordinate tools and process intermediate results, and a multi-agent beta lets it run concurrent subagents and synthesize their work in a single request.
4:19Dex: Translate that. Normally an agent calls a tool, ships the whole result back to the model, the model reads it, decides, calls again. Every round trip burns tokens and time. Programmatic Tool Calling lets the model write lightweight JavaScript that coordinates tools inside a hosted runtime, process intermediate results, filter unnecessary data, and pass only the useful material forward, which reduces round trips and can lower token use for bounded, tool-heavy workflows.
4:55Nova: And the receipts are real. PlayCo saw sixty-three point five percent fewer total tokens and fifty point one percent fewer model turns with Programmatic Tool Calling coordinating its tool pipelines. Simon Willison's read: it could help bridge the gap between MCP servers and full terminal sessions that compose command-line utilities in useful ways.
5:21Dex: One honest caveat on when to use it. The deciding variable is how predictable your control flow is: if the tool sequence is mostly known in advance, fetch, filter, transform, aggregate, it turns a dozen model round-trips into one program run; if the agent genuinely doesn't know its next step until it sees the last result, direct tool calling stays the honest architecture.
5:49Nova: So the move this week: take one existing agent workflow where the steps are predictable, a data pull, a filter, a summary, and re-implement it with Programmatic Tool Calling on Terra or Luna. Log tokens before and after. If your control flow is knowable, you just cut your bill in half and got ahead of every builder still doing one tool call per model turn.
6:15Dex: Story two, the desktop merge, because people are already confused. On July ninth, OpenAI merged the Codex app into the new ChatGPT desktop app for Mac and Windows; Codex is not going away, it remains a dedicated coding experience inside the app, next to Chat and Work. The older ChatGPT desktop app is being renamed ChatGPT Classic.
6:44Nova: The one mode you want as a builder is Codex mode, not Work. Codex mode shows more technical detail that ChatGPT Work abstracts away. And the new Codex bits are genuinely useful: inline editing in diffs, pull request review in the side panel, faster Computer Use powered by GPT five point six, and multi-repository projects. If you already had Codex installed, just update it; it becomes the new app.
7:15Dex: Story three, and this is the builder-stakes one. Fable five, the model a lot of you plan and architect with, is still living on an extension. Anthropic pushed included access on all paid plans through July twelfth. After that, it meters via usage credits at ten dollars per million in and fifty per million out, more than twice Opus. If credits are off when your allowance runs out, access just stops.
7:46Nova: The silver lining up front: it is not a permanent goodbye. A Claude Code lead engineer has said they aim to restore Fable to subscriptions as capacity allows. But you have got the weekend, so the move, and we will hit it again in Use Cases, is do not spend the last Fable days on throwaway output. Have Fable write down how it plans and structures your hardest recurring job as a skill file. Keep the method after the model gets rationed.
8:19Nova: Tool Lab. Two picks you can use today. First, trending: Tencent's Hy three. Released July sixth, a two hundred ninety-five billion parameter mixture-of-experts model under the Apache two point zero license, with free API access on OpenRouter through July twenty-first. Credit the Tencent Hunyuan team.
8:44Dex: The reason it matters for you specifically: it activates only twenty-one billion parameters per token and ships a two hundred fifty-six thousand token context window, so it is cheap to serve and can eat a whole repo. And the reliability work is the real story. Tencent reports SWE-bench variance within a few points whether the model runs inside Claude Code-style harnesses, Cline, or KiloCode. That harness-consistency is what usually breaks open models in real agents.
9:22Nova: The move: you have an OpenRouter account, so this week point one high-volume coding job, boilerplate, scaffolds, mechanical refactors, at the tencent slash hy three colon free route. Use reasoning effort high for the hard steps. Compare quality and cost against your Claude baseline. Just know the ceiling: on Tencent's own appendix, GLM five point two beats Hy three across the agentic coding suite, so keep hard reasoning on Claude. And the free window is a clock: it closes July twenty-first.
9:59Dex: Second pick, a hidden gem for the L4 crowd: the multi-agent beta in the GPT five point six Responses API. Not a repo, a capability most builders will not touch for weeks. One GPT five point six instance coordinates several subagents, assigns independent workstreams, and synthesizes their findings.
10:23Nova: Why bother when you can hand-wire subagents yourself? Because now it is one request with built-in synthesis, which means less orchestration glue for you to maintain. The move: take a task that naturally splits, say researching five competitors, and fire it as one multi-agent request on Terra instead of five sequential calls. Watch your wall-clock time drop. Just budget for it: parallelism reduces wall-clock time but increases token use and creates coordination failures a single-agent trace does not have. Test on a throwaway task before it touches client work.
11:07Nova: Use Cases. Two smart moves with what is already in your hands. One, the Fable distillation, and the clock is real. Before Saturday, run your hardest recurring job on Fable, then have Fable write down how it planned and structured that work as a Claude Code skill file. Skills are markdown, so that same file runs on Sonnet five, on Grok, on Codex. You are buying the expensive model's judgment once and keeping it cheap forever. Credit to the builder iwoszapar, who has been running this play all week.
11:46Dex: Two, a fresh one off today's release. If you build any tool-heavy agent, copy the Programmatic Tool Calling pattern even outside OpenAI: separate the deterministic plumbing, the fetch-filter-aggregate, from the expensive judgment call. Imagine collecting thousands of customer records: use a program to filter them before asking the model to judge, so the expensive reasoning model sees the relevant fifty records rather than every raw entry. You can wire that logic into Claude Code with a bash tool today. Cheap plumbing, expensive judgment only where it counts.
12:28Nova: The Deep Dive, one idea, higher altitude. This week four labs made near-frontier intelligence cheap and interchangeable: GPT five point six opened to everyone, Grok undercut on price, Hy three gave the weights away, and even Fable, the best, got rationed by capacity. If your edge was access to the single smartest model, that edge just repriced toward zero.
12:59Dex: But notice what did not get cheaper. The thing OpenAI actually shipped as its headline was not raw IQ; it was Programmatic Tool Calling and multi-agent, the plumbing that gets more useful work per token. The value moved from the model to the harness around it.
13:19Nova: So here is what I would actually do about it. Stop chasing the smartest model, and start building the layer that makes any model swappable underneath you: your routing rules, your distilled skills, your tool-calling programs that filter before they reason. That layer is what a client cannot get from an API key, and it is the only thing on your workbench this week that did not just get commoditized. Next reshuffle, you change one line of config and keep the spread.
13:53Nova: Level up this week: re-implement one predictable agent workflow with Programmatic Tool Calling on Terra or Luna, and distill one Fable skill before July twelfth. Twenty minutes, and every run after is cheaper and outage-proof.
14:10Dex: And if you want every story, number, and link from today in your inbox tomorrow morning, subscribe to the free Mainframe newsletter. It is the written version of exactly what we just walked through.
14:23Nova: That is Mainframe for Friday, July tenth. Build something today that the next model launch cannot take away from you. See you Monday.