Vercel AI Gateway (200K+ teams, tens of trillions of tokens) production data, animated by Guillermo Rauch this week: Anthropic's spend share rose from 61% to 65% in May, holding 70-80% of spend across every high-stakes use case (app generation, back-office agents, coding agents).
The volume chart inverts it: DeepSeek jumped from under 1% of tokens to 17% in a single month (third place, ahead of OpenAI). In the AI coding-agent category, DeepSeek drove 49% of token VOLUME but only 4% of COST.
Read-through: premium reasoning lands on Claude Opus; cheap volume lands on Gemini Flash / DeepSeek. Spend follows high-stakes calls; volume follows low-stakes ones.
Operator takeaway: this is a routing rule to copy, not a chart to admire. Cheapest capable model by default; frontier only where the answer must be right. Bank the spread.
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. Book a 30-min call. https://outpace.media
The stories
Artifacts in Claude Code reach Pro/Max (UPDATE)
Covered the June 18 launch (Team/Enterprise only). DELTA (this weekend, via TestingCatalog): Pro and Max users can now access Artifacts in Claude Code. Turns a session into a live, shareable web page (PR walkthrough, dashboard, incident timeline) at a private link that updates as the session works, built from your codebase + connectors + conversation. CATCH: it's a capture of work, not an app — single self-contained HTML, 16 MB cap, strict CSP blocks all external network calls; styled pages cost more output tokens than plain terminal text.
The move: publish an artifact walking through your next diff and send the link to a client instead of a status email.
Condense opens public access — up to ~72% off coding-agent token bills (July 3)
A drop-in proxy between agent and model. Two in-house models: Helene 1 (extractive, keeps survivors verbatim, strips before cache) and Adeline 1 (packs settled agent loops to ~9% of original tokens). Skeleton (system prompt, user messages, final answers, paired tool calls+results) preserved byte-for-byte; code/patches/errors never touched. Cache reads were 67.7% of a typical bill across 12 real sessions. No system-prompt change; results untouched. 100M free tokens to test. Carried by TestingCatalog, Min Choi. Install (Mac/Linux): `curl -fsSL bash`.
Alibaba bans Claude Code company-wide (UPDATE to the June 24 distillation feud)
Effective July 10, employees banned from Claude Code and must uninstall all Anthropic products; pushed to in-house Qoder. Stated trigger: reports Claude Code inspected environment (time zone, proxy settings) to flag China-linked users — some called it a "backdoor." Anthropic disputes it: an employee said on X the feature was a March experiment to stop resellers and prevent distillation. Reported by The Information, Reuters, carried by kimmonismus, Andrew Curran. Builder lesson: audit what your coding agent phones home.
Discussions to manufacture Anthropic's custom AI chip on Samsung Foundry's 2nm process + advanced packaging (The Information, via Andrew Curran / TestingCatalog). Still preliminary: no detailed design/testing yet; chips not expected before late 2027. Follows OpenAI's Broadcom "Jalapeño." Cheaper per token is coming, but years out.
No broad GA since the June 26 preview. Sol/Terra/Luna remain limited to ~20 vetted API/Codex partners; not in ChatGPT; no waitlist. GA "in the coming weeks" (Axios: Altman hopes mid-July if review clears). Keep production on GPT-5.5.
Nobody controls the token count — so control the junk.
Level up
Install Condense on one long-running Claude Code session, run a real task, and read your own savings off the dashboard before trusting the headline. If it holds at half the tokens with byte-exact data untouched, that's your cheapest margin this month.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: Claude Code just stopped being a thing you watch scroll by in a terminal. As of this weekend, Pro and Max users can turn a live coding session into a shareable web page: a PR walkthrough, a living dashboard, an incident timeline, published to a private link that updates itself as the work happens. 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 Sunday, July fifth, twenty twenty-six. I am Nova.
0:37Nova: On today's Mainframe: Artifacts in Claude Code opens up beyond the enterprise wall, and it changes how you hand work to a client. A brand-new proxy called Condense claims to cut your coding-agent token bill by up to seventy-two percent, with published receipts. The Alibaba versus Anthropic feud goes fully cold-war, with a company-wide Claude Code ban landing this week. And on the money layer, Vercel's own production data shows exactly who is winning the token economy, and it is not who wins the volume chart. Let's get into it.
1:18Dex: Quick note before the Board: the voices you are hearing are AI. The reporting, the numbers, the calls we make on them, all human. Now, the big picture.
1:31Nova: Here is the state of play, and it is cleaner than the noise suggests. Vercel runs the AI Gateway, which sees production traffic from over two hundred thousand teams, tens of trillions of tokens. Guillermo Rauch animated the spend race this week, and the numbers are stark. Anthropic's spend share grew from sixty-one percent to sixty-five percent in May, and it takes seventy to eighty percent of spend across every high-stakes use case.
2:04Dex: But the volume chart tells the opposite story. In April, DeepSeek was less than one percent of gateway tokens. In May its volume share jumped to seventeen percent, putting it third, ahead of OpenAI.
2:19Nova: Right, and that is the whole game in one data point. In the AI coding-agent category, DeepSeek drove forty-nine percent of the token volume but only four percent of the cost. Cheap models carry the grunt work; the expensive models get paid for the answers that have to be right. Premium reasoning calls land on Claude Opus, cheap fast calls land on Gemini Flash, spend follows the high-stakes calls and volume follows the low-stakes ones. For you at the keyboard: this is not a chart to admire, it is a routing rule to copy. Send the boilerplate to the cheapest model that clears your quality bar, reserve the frontier for the calls that carry your reputation, and pocket the difference. Which is exactly what today's tool story lets you do harder.
3:20Nova: Brought to you by Outpace. Look at that spend chart again: the leverage was never the model, it was the operator who knew which model to point where. That is Outpace. A thirty-year operator who built real businesses (real estate, hospitality, food, a theater) now aims a full team of AI agents at one client's project at a time, strategy to ship. It is a system, not a freelancer, and not a faceless mega-agency. The agents do the work, one senior operator directs them, and at the end you own the code and the keys. If you have a project that has been sitting in a someday folder, book a thirty-minute call at outpace dot media. That is outpace dot media.
4:14Dex: The Wire. Lead story: Artifacts in Claude Code.
4:18Nova: We covered the June eighteenth launch, when this was locked to Team and Enterprise. The delta this weekend, flagged by TestingCatalog: Claude adds Artifacts to Claude Code, turning session work into live, shareable web pages, for PR walkthroughs, incident pages, dashboards, and checklists. The important shift is who can reach it. Pro and Max users now get access. Claude Code can capture work progress as an artifact, a live shareable page including PR walkthroughs, system explainers, dashboards, and release checklists that update themselves as your session works, whether it is investigating an incident, refactoring a service, or analyzing months of data.
5:08Dex: And it builds from what already exists. Claude Code builds an artifact using the full context of your session, including your codebase, connectors, and the conversation itself, so you don't need to wire up data sources or stand up infrastructure. You ask for a page, and it builds it.
5:26Nova: Know the fence, though. An artifact is a capture of work, not an application. Each one is a single self-contained HTML page capped at sixteen megabytes, wrapped in a strict security policy that blocks all external network requests. So no live back end, no form submissions. The move this week: next time you finish a real task, ask Claude Code to publish an artifact walking through the diff, and send that link to your client instead of a status email. You just turned invisible terminal work into a deliverable. One catch worth pricing in: generating a styled page uses more output tokens than plain terminal text, so a heavy artifact costs more.
6:17Dex: Story two, and it is a direct answer to that cost. Condense.
6:22Nova: This is the fresh one, opened to the public July third, flagged by TestingCatalog and Min Choi. Condense dot chat opened public access to a context-compression proxy for coding agents, a system that sits between the agent and the upstream model, shrinking each request before billing, and it cuts token bills by up to seventy-two percent on deep sessions. Here is the problem it targets, and it is real. In a working agent loop, the model is re-sent the system prompt plus the whole conversation on every turn, so by mid-session it has re-read the same early context hundreds of times. Across twelve real sessions, cache reads were sixty-seven percent of a typical bill.
7:11Dex: Two models do the work. How are they different from just summarizing?
7:16Nova: Good question, because this is the clever part. Helene one is extractive: it scores every token and keeps the survivors verbatim, a strict subset of what your agent actually saw. The second, Adeline one, takes whole settled agent loops the session has moved past and packs each into one short summary that keeps intents, file paths, identifiers, errors, and code, landing at roughly nine percent of the original tokens. And the safety design matters here. The skeleton, meaning the system prompt, user messages, final answers, and paired tool calls with their results, is preserved byte for byte and never rewritten. Code, patches, and error output are never touched. That is the difference from a lossy image trick like the one we covered yesterday: this keeps your exact bytes where exactness matters.
8:22Nova: We have tracked this feud since Anthropic accused Alibaba of a distillation campaign on June twenty-fourth. The new development this week, carried by The Information, Reuters, and kimmonismus: Alibaba banned employees from using Claude Code and asked them to remove all Claude models from their work computers. Effective July tenth, employees are banned from Claude Code in the workplace and must uninstall all Anthropic products, and Alibaba is recommending its self-developed platform, Qoder, instead. The stated trigger: reports that Claude Code contained mechanisms to inspect users' computing environments, including time zone and proxy settings, to identify whether users were operating from China, which some researchers described as a potential backdoor.
9:18Nova: Anthropic disputed the characterization; an employee wrote on X that the feature was introduced as an experiment in March to combat unauthorized account resellers and prevent model distillation. The builder read: this is two superpowers fighting over model theft, but the lesson for you is smaller and sharper. Coding agents touch your source, your systems, your keys. Companies are no longer letting any AI tool into their workflows: if a tool can tap source code or company data, leadership wants to know exactly what it does and who controls it. Audit what your agent phones home. That is not paranoia now, that is diligence.
10:05Dex: Repo Spotlight. Two picks, both about the same enemy: wasted tokens.
10:12Nova: First, our trending pick is Condense itself, because it is genuinely a same-day install. On Mac or Linux you run: curl dash f s S L, https colon slash slash c l i dot condense dot chat slash unix, piped to bash. Drop it in front of your coding agent, no key swap. It requires no change to the system prompt, so the tool's results remain untouched. The move: run one long Claude Code session through it, watch the dashboard, and read your own savings before you trust the headline. They are even giving one hundred million free tokens to prove your agent is wasting context. Credit to the Condense team.
11:05Nova: The hidden gem is a mindset tool, not a proxy: OpenWiki from LangChain, by Brace Sproul's team. We name-checked it Saturday; today the how-to, because it solves the bloated CLAUDE dot M D problem directly. It is a CLI that writes and maintains documentation for your codebase built for agents: it creates a wiki in an openwiki folder, and automatically appends prompting to your AGENTS dot M D or CLAUDE dot M D files to tell your coding agent to reference it when searching for context. Most agents already read CLAUDE dot M D for instructions, but that file is not the right place to store hundreds of pages of docs. It should point the agent toward the right context, then let it retrieve what it needs. Then you make it durable: add the included GitHub Action to open a PR once a day with documentation updates. Install it, point it at your messiest repo, and stop stuffing everything into one instruction file.
12:13Dex: The Deep Dive. The through-line today is a single question: who controls how many tokens your work costs? Nobody fully does, and that is the whole story.
12:25Nova: Look at what stacked up this week. Artifacts get cheaper to distribute but cost more tokens to generate. Condense ships to cut those tokens by up to seventy percent. Vercel's data shows the frontier getting more expensive per token even as cheap models flood in. And zoom out to the hardware: this week, per The Information via Andrew Curran and TestingCatalog, Samsung is in discussions with Anthropic to manufacture its custom AI chips, considering Samsung Foundry's two-nanometer process and advanced packaging. Every layer of this industry is now bending toward one goal: bringing the cost per token down. But here is the honest part. Those chip discussions remain preliminary, custom chip design typically takes eighteen months to two years, and Anthropic won't see chips shipping until late twenty twenty-seven at the earliest.
13:30Nova: Years out. Which means the only cost lever you actually control today is the one at your keyboard: how much junk you send the model. That is why the boring proxy beats the exciting chip this quarter. So here is how you put it into action, concretely, this week. Step one: instrument one real Claude Code workflow, log tokens per completed task, get a baseline. Step two: install Condense in front of it, run the identical task, and read the savings off your own numbers, not the marketing. Step three: for that same repo, run OpenWiki so the agent pulls context on demand instead of dragging a bloated instruction file into every turn. Step four: write the routing rule down, cheap model by default, frontier on purpose, exact bytes for anything with a hash, a key, or an ID. Do those four and you have turned a chaotic pricing landscape into a margin you own, on models you can already reach.
14:43Dex: And that margin is yours to keep, not to hand back to the client.
14:51Nova: Your one move: install Condense on a single long-running Claude Code session, run one real task through it, and read your own savings off the dashboard before you trust the number. If it holds at half the tokens with your byte-exact stuff untouched, that is your cheapest margin this month. And if you want every story, number, and link from today in your inbox, subscribe to the free Mainframe newsletter. It is the written brief, no fluff, every morning.
15:24Dex: That is the show. We will see you tomorrow.
15:26Nova: Send the model less junk. Keep the spread. This has been Mainframe.