Etched exits stealth (June 30): working transformer-inference chip, $800M raised (latest round at a $5B valuation), $1B+ in signed customer contracts; first-pass silicon on TSMC N4P; already running DeepSeek, Qwen, Llama. Claims ~20x H100 throughput via fixed-function attention + low-voltage design (treat as marketing until independently benchmarked). Karpathy among backers.
OpenAI floats giving the US government a 5% stake (~$42.6B at its $852B valuation), and proposes Washington hold 5% of each leading US lab (possibly Anthropic, Google, Meta). Reported by the FT; carried by Andrew Curran, kimmonismus, TestingCatalog. A peace offering after a month of model gating.
Operator takeaway: serving costs are on a downward escalator (inference is where the profit war moved). Price client work to keep the spread. The durable moat is the owned workflow, not the rented model.
A system, not a freelancer: a 30-year operator 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
Claude Science ships (Anthropic, Jun 30 / rolling this week)
A desktop AI workbench (beta, macOS + Linux, Pro/Max/Team/Enterprise), NOT a new model. Wraps existing Claude models (incl. Opus 4.8) in a full research env: runs analyses, orchestrates compute (SSH HPC or Modal), connects to 60+ scientific databases, and traces every figure to the exact code + environment + message history that made it. A reviewer agent flags incorrect citations, untraceable numbers, and figures that don't match their code, self-correcting as it runs. Honest limit: the reviewer does NOT rerun analyses or judge whether your method was right; it catches fabrication/drift, not bad judgment.
The move: copy the architecture even outside science — one worker agent produces, one reviewer agent verifies every claimed number traces to code that ran.
Covered yesterday: Fable 5 back globally (Jul 1). DELTA today: the June 2 EO set a 30-day clock for agencies to publish a model-benchmarking process; that lands July 2. Andrew Curran reports the White House is accelerating the announcement, incl. a benchmark co-designed with Anthropic, possibly this week. Tell (The Hacker News): Fable 5 never went through the voluntary path — Washington reached for export controls instead; it still has no binding process, only improvised ones.
A single classifier blocks the flagged jailbreak in >99% of cases and reroutes flagged requests to Opus 4.8; it also catches more benign coding/debugging requests as a side effect (CAISI reviewed the safeguards).
The move: if shipping on Fable 5, log which model actually served each request — silent Opus fallbacks change output + bill.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: Anthropic just shipped a whole new app that turns Claude into a lab bench: it runs your analysis, queries sixty-plus scientific databases, and traces every chart back to the exact code that made it. Claude Science is here, and it is a template for how any of us should build agents that prove their own work.
0:20Nova: 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 Thursday, July second, twenty twenty-six. I am Nova.
0:37Nova: On today's mainframe: Claude Science, a desktop workbench that gives every result an evidence trail, and the free skill libraries that let you steal the pattern today. Then the frontier gets un-caged: Fable five is back worldwide and the White House benchmark is due any minute. Plus OpenAI floats handing Washington a forty-two-billion-dollar slice of itself, and Etched comes out of stealth with a chip that could halve your inference bill.
1:12Dex: And two tools you can wire up before lunch.
1:16Dex: One honest note before we go deep: our voices are AI. The reporting, the picks, and the opinions are human-made. Newsletter link at the end.
1:27Nova: The Board. For weeks the story here was the frontier getting caged: Washington pulling Fable five, gating GPT five point six. This week the cage cracked open, and the money layer moved with it. Start at the bottom of the stack. Etched came out of stealth June thirtieth with a working chip, over one billion dollars in signed customer contracts, and eight hundred million dollars raised, the latest round valuing it at five billion dollars. In under three years it hit first-pass silicon on TSMC's N four P process, and its systems are already running DeepSeek, Qwen, and Llama.
2:11Dex: Karpathy is an investor. What is the actual claim?
2:14Nova: The transformer inference chip claims twenty times the throughput of Nvidia's H one hundred using fixed-function attention circuits and a low-voltage design. Treat the twenty-times number as marketing until Artificial Analysis benchmarks it. But the direction is real: the AI race is moving from training to inference, where running models cheaply and fast decides who profits, and capital is flowing to anyone who can chip away at Nvidia's lead. For a builder, that means the same thing it has meant all week: your serving costs are on a downward escalator. Price client work so you keep that spread instead of handing it back.
3:04Nova: It did. OpenAI has proposed handing the United States government a five percent stake in the company, worth roughly forty-two point six billion dollars at its eight hundred fifty-two billion dollar valuation. Altman and other executives proposed Washington hold five percent of each leading American AI developer, possibly including Anthropic, Google, and Meta, though it is unclear the others would agree. Read it plainly: this is a peace offering after a month of the government yanking models around. The lab that spent the spring fighting Washington is now offering it equity. That is the macro backdrop for today's lead: the frontier is negotiated now, so the durable leverage is the workflow you own, not the model you rent.
4:00Nova: Which brings us to today's sponsor, and it fits the theme.
4:04Dex: Today's mainframe is brought to you by Outpace. Here is the through-line: when the frontier can be re-priced or re-gated on a Tuesday, the winners are the operators who own the whole workflow, not just a prompt. Outpace is exactly that. It is not a freelancer and not a faceless mega-agency: it is a system. One senior operator who spent thirty years building real businesses (real estate, hospitality, food, a theater) now directs a full team of AI agents at one client's project at a time, strategy to ship. And at the end, you own the code and the keys. If the platform churn has burned you, book a thirty-minute call at outpace dot media. That is outpace dot media.
4:55Nova: The Wire. Lead story, and it is the builder move of the day. Anthropic launched Claude Science, a purpose-built AI workbench for scientific research. It is not a new model: it is a new app that wraps Claude's existing models in a full research environment that can run analyses, manage compute, query scientific databases, and trace every figure back to the exact code that generated it. It is in beta on macOS and Linux for Pro, Max, Team, and Enterprise plans.
5:34Dex: Why should a developer who is not a biologist care?
5:37Nova: Because it is the cleanest example yet of the pattern that matters everywhere: provenance by default. When it generates a figure, Claude Science includes the exact code and environment that produced it, a plain-language description of how it was created, and the full message history. As the pipeline runs, a reviewer agent inspects the outputs, flagging incorrect citations, untraceable numbers, and figures that do not match their underlying code, and self-corrects as it goes. That is the thing to steal. An agent that checks its own work against ground truth beats a smarter agent that hands you a confident wrong answer.
6:24Nova: A sharp one from the docs: the reviewer is not a guarantee of correctness. Anthropic explicitly says it does not rerun analyses and does not decide whether your method was right for the question. It catches fabrication and drift, not bad judgment. Your move: even if you never touch genomics, copy the architecture. Make one agent produce work and a second agent verify that every claimed number traces to code that actually ran.
6:58Dex: Story two, the frontier un-caged, and it is a real update, not a re-run. We covered Fable five returning globally yesterday. The new delta today: the June second executive order set a thirty-day window for agencies to produce a benchmarking process for new models, which lands today, July second. Andrew Curran reports the White House is accelerating that announcement, including a benchmark co-designed with Anthropic, possibly this week.
7:31Nova: So the standing framework everyone has been improvising toward may become real. And here is the tell from the hacker news reporting: Fable five never went through that voluntary path; the government reached for export controls instead. When Washington wants to move fast on a frontier model, it still has no binding process, only improvised ones. If a documented benchmark shows up today, that improvisation ends. For builders that is good news: predictable rules beat surprise kill-switches. But keep the open-weight fallback wired anyway.
8:13Dex: Story three, quick hit. Anthropic's Fable five fix was a single classifier that blocks the flagged jailbreak in more than ninety-nine percent of cases and reroutes flagged requests to the older Opus four point eight, and it catches more benign coding and debugging requests as a side effect.
8:34Nova: Which is the practical catch for anyone shipping on it: silent reroutes to Opus change both your output and your bill. Log which model actually served each request.
8:47Nova: Tool Lab. Two picks, both live off today's provenance theme. First, the trending one, tied straight to the Claude Science launch. It is the K-Dense scientific agent skills library. It turns any AI agent into a research assistant: the number-one agent skills library for science, used by over one hundred sixty thousand scientists, with a hundred forty ready-to-use skills plus a hundred-plus scientific databases across biology, chemistry, medicine, and drug discovery, compatible with Cursor, Claude Code, and Codex.
9:28Nova: You do not need the full Science app to get the guts of it. Install path: in Claude Code run slash plugin marketplace add K-Dense dash A I slash claude dash scientific dash skills. You can also reach every skill via their MCP server in any MCP-compatible client. And you do not need every Python package: only install the ones you need, since each skill specifies its own requirements in its SKILL dot M D. Credit to the K-Dense AI team. Concrete use: point it at a folder of CSVs and ask for a literature-backed analysis with citations that trace to source.
10:18Nova: This one solves the pain the moment you actually try to use those skills. It is claude dash skills dash M C P, also from the K-Dense team, and it is not on any trending list. It is an MCP server that lets any MCP-compatible assistant intelligently search and retrieve skills from the curated scientific repo and Anthropic's official skills. The clever bit is the design for real workflows: a lightweight fifteen-megabyte frontend starts instantly with no Cursor timeout, while the two-hundred-fifty-megabyte backend downloads in the background, using vector embeddings for semantic skill discovery. Install is one line: uvx claude dash skills dash M C P, or drop the M C P servers block into your config. It loads roughly ninety skills from Anthropic's official repo and K-Dense's scientific collection out of the box. Why you want it: once you have hundreds of skills, your agent cannot hold them all in context. This lets it search for the right skill on demand instead of you hand-wiring each one.
11:35Nova: The Deep Dive. Here is the week in one frame. Every big story keeps circling the same question: not can the model do it, but can you trust and prove what it did. Etched wants to make the doing cheap. Washington's coming benchmark wants to certify the doing is safe. And Claude Science, the smallest story by market cap, is quietly the most important for you, because it ships the answer to trust as a product feature: figures, tables, and notebooks can include the exact code, environment, and conversation that produced them, which is the difference between the A I made a chart and this chart has an evidence trail.
12:21Dex: The skeptic's read is that this is a niche science app.
12:25Nova: The skeptic misses the shape. What is new in Claude Science is not the model; it is everything around it: the tools, database connections, and compute integrations that let Claude run full analyses and keep track of what happened in a previous session. Strip out the word scientific and that is a description of every serious agent workflow you will build this year: real tools, real data, and a memory of what it did. The labs are converging on the same lesson: the moat is not raw intelligence, it is the auditable harness around it.
13:07Dex: So how does a builder put this into action this week? Give me the steps.
13:11Nova: Four steps, and none require a Ph.D. One: take one agent workflow you already run, a report generator, a data pipeline, an audit, and split it in two. A worker agent that produces, and a reviewer agent whose only job is to check that every number and citation traces to something real. Two: make the worker emit provenance by default. Every output ships with the code or query that made it and a plain-language note of how. Three: give the reviewer a hard rule copied from Anthropic's own docs: flag any result claimed as computed when no corresponding code ran. Four: install claude dash skills dash M C P so your agent pulls the right skill on demand instead of you gluing each one in. Do that and you have turned trust me into here is the receipt, which is exactly what wins the client who is nervous about AI.
14:17Nova: The Close. Your level up this week: split one workflow into a worker and a reviewer, and make every output carry the code that produced it. Provenance is the feature clients will pay for. If you want every story, number, and link from today in your inbox, subscribe to the free mainframe newsletter: the whole brief, every morning. That is today's mainframe. See you tomorrow.