GPT-5.6 pricing (Simon Willison): Luna $1/$6, Terra $2.50/$15, Sol $5/$30 per M input/output. On OpenAI's Agents' Last Exam, Sol scored 53.6, +13.1 over Fable 5; even Sol at medium beats Fable 5 at ~1/4 the estimated cost, and Terra/Luna beat Fable 5 at ~1/16 the cost (OpenAI's own benchmark; Willison notes Fable still wins some real coding jobs).
Takeaway: don't standardize on one vendor this week. Price per finished task, keep a 3-4 model routing dial, bank the spread. This frames today's Fable cliff (below).
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)
— Claude Code gets a built-in web browser on desktop (Anthropic, July 10). The Browser pane is a fully tabbed browser inside the Claude Code desktop app; Claude can open docs, issue trackers, internal web apps, or any URL, then read, click, and interact the way it already does with local dev servers. Launch: Cmd+Shift+B (macOS) / Ctrl+Shift+B (Windows). KEY DISTINCTION: the in-app Browser pane uses a clean, isolated profile with no logins/history (for building + testing); to have Claude act as YOU in logged-in sessions, use the Claude in Chrome extension instead. Sandboxed and configurable (you choose whether sessions persist). Update the desktop app first. Carried by TestingCatalog + ClaudeDevs.
The move: stop copy-pasting URLs/error pages into Claude Code. Point it at your local dev server, tell it to open the page, click the broken flow, read the console error, and fix it in one loop.
Cursor ships side chats, transcript search, and conversation hooks (changelog, July 10)
Side chats: open with /side, /btw, or the + button; each is a durable, full agent conversation that carries main-chat context, focuses on read/search/answer by default, and can be @-mentioned to pull its findings back into the main thread — so you can research a tangent without derailing the running agent. Search within a conversation with Cmd+F (match counter, jump between matches). New hooks observe/control the conversation itself: prompts, responses, thinking, subagents, compaction, turn completion (on top of existing tool-execution hooks). (TestingCatalog)
The move: while a long build runs, open a /side chat to ask "why this library?", get the answer, then @-mention it back — keep the long run alive, keep your head clear.
Fable 5 leaves subscriptions; meter starts July 13
Included on Pro/Max/Team/select Enterprise (up to 50% of weekly limits) through July 12 at 11:59:59 PM PT; after that, usage credits at $10/M input, $50/M output (>2x Opus 4.8). If credits are off, access simply ends. NOT permanent: a Claude Code lead engineer says Anthropic aims to restore Fable to subscriptions "as soon as capacity allows" (no date — plan around July 12, not a hoped-for restore).
The move: don't burn the weekend on throwaway output — bank the method (see Use Cases).
The browser isn't about browsing — it's about closing the loop.
Level up
Update Claude Code, launch the in-app browser (Cmd+Shift+B / Ctrl+Shift+B), and make Claude verify one front-end change in it before reporting finished — and bank one distilled Fable skill before the meter starts July 13.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: Claude Code just grew its own web browser. It lives inside the desktop app now, so Claude can open a URL, read the page, click through it, and debug your app without you ever tabbing over to Chrome. 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 Saturday, July eleventh, twenty twenty-six. I am Nova.
0:30Nova: On today's Mainframe: Claude Code gets a built-in, sandboxed browser, and we tell you the exact keystroke to launch it and the one habit to change this week. Cursor just shipped side chats and searchable agent transcripts, a genuinely new way to work without derailing your main agent. And the one with skin in it: Fable five, the smartest Claude, drops off subscription plans in two days. The meter starts July thirteenth, and we will tell you what it costs and how to bank its value before the window shuts. Let's get into it.
1:10Dex: Quick, honest note before we roll. Our voices are AI. The reporting, the picks, and the opinions are human-made. Nova, set the table.
1:23Nova: The board today is short because the interesting stuff is at your keyboard. One macro move that touches your wallet: Simon Willison ran the numbers on the new GPT five point six family, and the pricing is the story. The new models are priced per one million input and output tokens as Luna one dollar and six dollars, Terra two dollars fifty and fifteen dollars, Sol five dollars and thirty dollars. And on OpenAI's headline agent benchmark, GPT five point six Sol sets a new high, eclipsing Claude Fable five by thirteen point one points; even at medium reasoning it beats Fable five by roughly one-quarter the estimated cost.
2:14Dex: So the smartest tier is no longer the only place to get near-frontier work done.
2:19Nova: Right. And that matters this weekend specifically, because your smartest Claude is about to cost real money. The takeaway for your margins: do not standardize on one vendor, price per finished task, and keep a routing dial. Near-frontier is a commodity now. That is the whole board.
2:43Nova: Brought to you by Outpace. When the model shelf reshuffles three times a week, like it just did, you do not want to be the person re-wiring your stack every Monday. Outpace is a system, not a freelancer. A thirty-year operator, someone who actually built real businesses, real estate, hospitality, food, a theater, now aims a whole team of AI agents at one client's project at a time, strategy to ship. Not a faceless agency, not a one-person shop: the agents do the work, one senior operator steers them, and at the end you own the code and you own the keys. If the pace of all this is making you dizzy, that is the pitch. Book a thirty-minute call at outpace dot media. That is outpace dot media.
3:40Nova: Anthropic shipped a built-in web browser inside the Claude Code desktop app, announced July tenth. Claude Code on desktop now has an in-app browser: Claude can pull up docs, designs, or any site, and read, click through, and interact the same way it does with your local dev servers. Here is precisely what it does that is different. The Browser pane functions as a fully tabbed browser that sits alongside your workspace; Claude can open documentation, issue trackers, internal web apps, or virtually any website, then read page contents, click links, and interact with elements.
4:24Dex: And the launch command, because that is what people actually need.
4:28Nova: Ctrl Shift B on Windows, or Cmd Shift B on macOS. Update the desktop app first. Now the important nuance, because there are two different browser things from Anthropic and people conflate them. This built-in Browser pane uses a clean, isolated profile with no history or saved logins, ideal for development and testing; if you want Claude to work with your existing logged-in sessions, you use the Claude in Chrome extension instead.
5:01Dex: So the sandbox one is for building, the Chrome extension is for acting as you.
5:06Nova: Exactly, and that split is the whole security posture. It is sandboxed and configurable: you choose whether sessions persist. The concrete move this week: stop copy-pasting URLs and error pages into your Claude Code session. The update removes one of the biggest annoyances in AI-assisted coding, constantly copying links back and forth; developers can now just let Claude open the page itself, understand the context, and continue working. The next-level version: point it at your own local dev server, tell Claude to open the page, click the broken button, read the console error, and fix it, all in one loop. That closes the gap between writing code and confirming the code works, and most builders are still doing that by hand.
5:58Dex: Second story, Cursor. TestingCatalog flagged it, and it is fresh.
6:03Nova: Cursor shipped side chats, transcript search, and new agent conversation hooks, in the release notes as of yesterday. The headliner is side chats. You open a side chat to ask questions, explore ideas, and investigate tangents without interrupting your main agent conversation; use slash side, slash B T W, or the plus button to create one that carries context from the main chat. And here is the clever part: each side chat is a durable, full agent conversation that you can revisit later and at-mention to pull context back into the main thread.
6:43Dex: So it is not a throwaway. You can fork a question, get an answer, and bring the answer back.
6:50Nova: Right. By default side chats focus on reading, searching, and answering; use them to ask clarification questions, research alternatives without committing to a pivot, and sanity-check a decision while the main agent keeps running. The other two: you can search within a conversation with Cmd F, jump between matches, and keep searching through long transcripts, and there are new hooks. Cloud agents already had hooks around tool execution; the new ones let you observe and control the agent conversation itself: prompts, responses, thinking, subagents, compaction, and turn completion.
7:35Nova: Next time your main agent is grinding on a long build, do not stop it to ask "wait, why did you pick that library." Open a side chat with slash side, ask there, then at-mention the answer back in. You keep the long run alive and you keep your own head clear. That is a real workflow change, not a cosmetic one.
8:00Dex: Now the one people are anxious about. Fable.
8:04Nova: Yes, and this is the skin-in-the-game item, so let me lead with the stake and the silver lining together. On July twelfth, Fable five comes off subscription plans and moves to prepaid usage credits at ten dollars per million input tokens and fifty dollars per million output tokens; if credits are off, access simply ends. To be precise on timing, because it has moved once already: Anthropic extended the deadline to July twelfth at eleven fifty-nine PM Pacific, so the meter effectively starts July thirteenth. You have this weekend.
8:45Nova: That is the silver lining, up front. This exit looks temporary: a Claude Code lead engineer confirmed Anthropic aims to restore Fable five as a standard part of subscriptions as soon as capacity allows. But no date. No firm restoration timeline has been given, so plan around the July twelfth date, not a hoped-for reinstatement. For scale on the price: fifty dollars per million output is more than twice Opus four point eight, Anthropic's next priciest.
9:20Dex: So what do I actually do with the last weekend?
9:23Nova: Do not burn it on throwaway output. Bank the method. And that is our first use case, because the best move is documented. The builder at iwoszapar did exactly this: during the free window they had Fable five write down its own working discipline as six Claude Code skills, then blind-tested them on Opus four point eight, where they won twelve of fourteen gradings; because a skill is just a markdown file, the same six also run in Codex and Gemini. Whatever Fable five costs after July twelfth, that discipline is already banked on the model you run every day. That is the play: this weekend, run your hardest recurring job on Fable, then have Fable write down HOW it planned and structured the work as a skill file. You keep the judgment after the model gets expensive.
10:21Dex: Second use case, and this one uses today's news.
10:24Nova: Pair the new in-app browser with a verification habit. Most people ship front-end changes and eyeball them. Instead, this week: after any UI change, tell Claude Code to open the local server in its own sandboxed Browser pane, click through the actual flow, and confirm the thing works before it reports "done." You are turning a confident narrator into an agent that checks reality. It costs you one sentence in your prompt and it catches the "looks done, isn't done" class of bug.
11:04Nova: With GPT five point six now cheap at the low end, wire Luna or Terra as your high-volume route behind Claude for boilerplate and mechanical refactors, and reserve Fable-or-Opus judgment for the hard reasoning. Write it in your CLAUDE dot M D as a rule: cheap by default, frontier on purpose. When Fable gets metered Monday, that rule is what keeps your bill flat.
11:36Nova: One idea: the browser inside Claude Code is not really about browsing. It is about closing the loop. For a year, the pattern was, the agent writes something and then narrates that it worked, and you go check. Every capability landing this week, the sandboxed browser, Cursor's agent that tests its own PRs, side chats that let a second agent sanity-check the first, points the same direction: the agent is being wired to observe the result of its own actions before it hands you anything. The value is shifting from "the model that writes the best code" to "the harness that verifies its own work." That harness is the thing a client cannot get from an API key, and it is the thing that survives every model reshuffle.
12:34Nova: This week, add exactly one verification step to your most-used workflow. Make the agent prove it, open the page, run the test, click the button, before it says done. That one habit outlasts whichever model is cheapest on Monday.
12:54Nova: Level up this week: update Claude Code, launch the in-app browser with Cmd Shift B, and make Claude verify one front-end change in it before reporting finished. And bank one Fable skill before the meter starts July thirteenth. If you want every story, number, and link from today in your inbox, subscribe to the free Mainframe newsletter. It is the written version of this show, every morning.