AI news for builders, July 3, 2026
Claude Sonnet 5 just landed with near-Opus intelligence at Sonnet prices, two dollars per million tokens in, and it finishes agent jobs that used to stall halfway. That is the model most of you should be running by Monda…
The Board
- Anthropic in early talks with Samsung to manufacture a custom AI chip (2nm process); hired ex-OpenAI chip engineer Clive Chan. Existing AWS Trainium / Google TPU / Nvidia partnerships remain central. Reported by The Information (via Andrew Curran, TestingCatalog). https://techcrunch.com/2026/07/02/anthropic-is-discussing-a-new-custom-chip-with-samsung/
- OpenAI's own inference chip "Jalapeño" (with Broadcom) announced last week. Nvidia still ~74% of the AI chip market.
- Etched exited stealth (June 30): $800M raised, $1B+ in customer contracts, first-pass A0 silicon on TSMC N4P; 400+ engineers; "Sohu" transformer-only inference chip; racks ship this summer. Not yet independently benchmarked. https://finance.yahoo.com/technology/ai/articles/etched-emerges-stealth-working-chip-150000905.html
Level up
Build and commit a two-tier model-routing rule this week (Sonnet 5 for volume, Fable/Opus for hard reasoning).
Chapters
- 0:31The Rundown
- 4:20Sponsor: Outpace
- 5:32The Wire
- 10:04Repo Spotlight
- 12:26The Deep Dive
- 15:02The Close
Every story, number, and link in your inbox.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: Claude Sonnet 5 just landed with near-Opus intelligence at Sonnet prices, two dollars per million tokens in, and it finishes agent jobs that used to stall halfway. That is the model most of you should be running by Monday. 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 third, twenty twenty-six. I am Nova.
0:30Dex: And I am Dex.
0:31Nova: On today's Mainframe: Claude Sonnet 5 ships as a genuinely cheaper way to run agents, and Fable 5 comes back from the dead after Washington blinked. Anthropic starts designing its own chip with Samsung, while Etched walks out of stealth claiming it can hardcode transformers into silicon. OpenAI floats handing the government a forty-two-billion-dollar slice of itself. And Simon Willison quietly drops a coding agent you can read end to end in an afternoon. Let's get into it.
1:06Dex: Quick note before we go deep. The voices you are hearing, mine and Nova's, are AI. The reporting, the picks, the opinions: those are human-made. We just do the talking.
1:19Nova: So let's set the table. The whole industry right now is a fight over one thing: who controls the compute, and therefore who controls the price of intelligence. This week that fight got physical. Etched came out of stealth with more than four hundred engineers pulled from Nvidia, Broadcom, Google's TPU program, and SK Hynix, and achieved first-pass silicon success on TSMC's four-nanometer process. They claim eight hundred million dollars raised, over a billion in customer contracts, and state-of-the-art throughput on inference. The bet, Dex, is brutal and simple.
2:04Dex: Right. The transformer architecture processes text by contextualizing every token against all others through attention, and every major model, GPT, Llama, Claude, runs that at the core of every request. On a general-purpose GPU that means loading weights, dispatching a CUDA kernel, and managing thousands of threads, only some doing useful work, so GPUs typically hit thirty to forty percent of their theoretical capacity. Etched is saying: bake attention straight into the chip and reclaim the rest.
2:42Nova: And here is why that matters to you at the keyboard. Cheaper inference is exactly what makes Sonnet 5 possible. And it is not just startups. OpenAI teamed with Broadcom on its own inference processor, dubbed Jalapeño. And now, brand new, Anthropic is in talks with Samsung to be the manufacturing partner for a custom AI chip. Andrew Curran and TestingCatalog both carried this yesterday off reporting from The Information. The tell: Anthropic hired Clive Chan, the second engineer to join OpenAI's custom chip team, who spent two and a half years building the Broadcom-designed accelerator OpenAI unveiled as Jalapeño. Every lab is going full stack, because whoever owns the silicon owns the margin. For a builder, the takeaway is not the chip. It is the direction of the price line: down. Which is why you should be re-architecting your workflows around cheaper, more agentic models right now, not next year.
3:51Dex: One caveat worth naming. Nvidia still holds an estimated seventy-four percent of the AI chip market, higher than before the custom silicon race, because demand grew faster than the alternatives matured. This is a long game, not a coup.
4:08Nova: A long game with your bill on the line. And that is a good moment to talk about who actually ships when the model changes under you.
4:20Dex: This episode is brought to you by Outpace. Here is the thing about a week like this one: the model you built on Monday is deprecated by Friday. Sonnet 5 lands, Fable comes back, prices move. Most teams freeze. Outpace does not. It was founded by an operator who spent thirty years building real businesses: real estate, hospitality, food, even a theater. Now he 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 a faceless mega-agency. The agents do the work, one senior operator directs them, and you own the code and the keys at the end.
5:09Nova: So if you have got an idea stuck in the gap between "I could build this" and "I actually shipped it," here is what to do next. Book a thirty-minute call at outpace dot media. That is outpace dot media. Tell them what you are trying to launch, and walk out with a plan.
5:30Dex: outpace dot media.
5:32Nova: To The Wire. The lead is the one you build on. Claude Sonnet 5 shipped Tuesday, promising performance close to Opus 4.8 at much lower cost, and it is now the default for free and Pro plans and available on every subscription. Introductory pricing is two dollars per million input tokens and ten per million output through August thirty-first, then it rises to three and fifteen. Artificial Analysis put a number on it: on one agentic coding benchmark Sonnet 5 scores sixty-three point two, versus Opus 4.8 at sixty-nine point two and the old Sonnet 4.6 at fifty-eight point one.
6:18Dex: Two honest asterisks. First, price. Sonnet 5 uses an updated tokenizer, so the same input can map to roughly one to one-point-three-five times as many tokens depending on content. Simon Willison measured it: about one-point-four times more expensive for English, but roughly flat for Simplified Mandarin. Benchmark your own workload before you assume the bill holds. Second, one skeptical developer, Theo Browne, reported Sonnet 5 burned through six thousand dollars in benchmark testing, more than any model he had evaluated, while scoring thirty-seven percent on his coding test. Effort settings matter enormously here.
7:05Nova: They do, and that is the actual skill. Run Sonnet 5 at low effort for the boring stuff, dial it up only when the task earns it. Now, the story everyone felt: Fable 5 is back. We covered the shutdown when it happened. What is new as of Wednesday: the Commerce Department lifted the export controls on Fable 5 and Mythos 5, and Fable is available globally again across the Claude Platform, Claude dot A-I, Claude Code, and Cowork, included for up to fifty percent of weekly usage limits through July seventh, after which it runs on usage credits.
7:49Dex: And the delta that matters for builders: it is more guardrailed now. Anthropic retrained a safety classifier for the reported jailbreak, says it now blocks that bypass in more than ninety-nine percent of attempts, and any flagged Fable 5 request gets automatically rerouted to Opus 4.8 with a notification. Dylan Patel and kimmonismus are both raving about the model; kimmonismus also flagged the catch, that it is, in his words, heavily guardrailed. Expect some routine coding tasks to silently fall back to Opus.
8:27Nova: Which is the real lesson of the whole saga. Epoch AI dropped a genuinely alarming number this week: AI is finding software vulnerabilities at scale. In June, they report twenty-one organizations disclosed around fifteen hundred high and critical severity CVEs, more than three-and-a-half times the previous monthly record. If you ship software, that is your homework: assume the attackers have these tools too, and patch accordingly.
9:02Dex: One more from The Wire. OpenAI. OpenAI has proposed handing the US government a five percent stake, worth roughly forty-two-point-six billion dollars at its eight-hundred-fifty-two-billion valuation. Reported by the Financial Times, carried by TestingCatalog and Min Choi. The idea is that Anthropic, Google, and Meta would each hand over a similar stake into a public vehicle modeled on the Alaska Permanent Fund.
9:31Nova: And here I will take a stance. This does nothing for you this week, but watch it, because a government that owns five percent of your model provider has an incentive to keep that provider winning. That could mean friendlier releases, or it could mean a regulator that owns the thing it regulates. Andrew Curran adds a sharp note: any shareholding negotiated before an IPO would lock in Washington's position ahead of a public float. File it under context, not action.
10:04Dex: Repo Spotlight. Two picks, both hands-on.
10:08Nova: First, the trending one, and it ships builder value today: Claude Science. Anthropic introduced it as an AI workbench for scientists, an app that integrates the tools and packages researchers use, produces auditable artifacts, and provides flexible access to computing resources. It is in beta for macOS and Linux. The killer feature is provenance: when it generates a figure, it bundles the exact code, the environment it ran in, a plain-language description, and the full conversation history. Concrete move: if you are on Pro or Max and touch data at all, install it and hand it one messy analysis you have been avoiding. Grad and postdoc researchers can also apply for one of fifty projects with up to thirty thousand in credits, applications open through July fifteenth.
11:07Dex: And the hidden gem, straight from a practitioner, not the front page. Simon Willison released llm dash coding dash agent, version zero point one alpha, on July second. It is a minimal coding agent built on top of his widely used llm command-line tool. The point is not that it is the most powerful agent. The point is that it is small enough to read and understand completely.
11:35Nova: Which is exactly the skill Willison himself has been preaching this week: he added a storyboard feature so your coding agent can record a video demo of the features it builds, and he framed the design lesson beautifully. A well-designed help output functions as a skill file for a coding agent, giving it enough context to autonomously script its own work. Install path: pip install llm, then read the coding-agent source top to bottom. If you have only ever used agents as black boxes, spending one evening reading one you can actually hold in your head will change how you prompt every other agent you touch. Credit to Simon Willison, one of the most reliable hands-on voices in this space.
12:26Dex: The Deep Dive. Let's talk about the pattern under all of today's news: everyone is going full stack.
12:34Nova: Six weeks ago the story was just models. Today, the same companies that sell you tokens are racing to build the chips those tokens run on. OpenAI has Jalapeño with Broadcom. Anthropic is in talks with Samsung. Amazon and Google already offer custom TPUs. And Etched is betting the whole company that a chip built only for transformers beats a general-purpose GPU. Here is the John Oliver turn: this looks like fierce competition, and it partly is, but strip away the press releases and it is really one thing: a coordinated bet that the price of intelligence is about to fall off a cliff, and that owning the silicon is how you capture the difference.
13:24Dex: The skeptic's corner, though. Etched's chip has been demonstrated to investors and early customers but has not been publicly benchmarked by an independent third party. Nathan Lambert makes the sober point that both selling tokens and building the token machines can be great businesses at once. And Nvidia's seventy-four percent share says the incumbent is not going anywhere fast. So the direction is clear; the timeline is not.
13:55Nova: So how do you actually put this into action, instead of just nodding at chip news? Three steps. One: this week, take your single most expensive AI workflow and re-run it on Sonnet 5 at a low effort setting. Measure the cost and the quality side by side against whatever you use now. Two: pick the one task in that workflow that genuinely needs frontier reasoning, and route only that to Fable 5 or Opus, nothing else. That is smart routing, and kimmonismus is right that it is becoming the real moat. Three: write it down as a rule in a CLAUDE dot M D file in your repo, so your agent picks the cheap model by default and escalates on purpose. Do that, and when the price of intelligence drops again next month, your margins move with it instead of against you.
14:57Dex: That is the whole game. Cheap by default, expensive on purpose.
15:02Nova: Your Level Up for this week: build a two-tier routing rule, Sonnet 5 for the volume, Fable or Opus only for the hard reasoning, and commit it to your project. One config change, permanent leverage.
15:18Dex: And if you want every story and every link from today in your inbox, subscribe to the free Mainframe newsletter. It is the written brief, all the reporting, none of the digging.
15:30Nova: That is Mainframe for Friday, July third. Route smart, patch your software, and we will see you Monday. I am Nova.
15:38Dex: And I am Dex. Take yourselves to the next level.