Good morning. Anthropic wants to be a drug company now, Mistral dropped a math prover that’s finding real bugs in real code, and a developer figured out you can shave 60% off Claude bills by turning your prompts into PNGs. Also: DeepMind’s union talks are already sour, and there’s a curious spike in serious CVEs that lines up suspiciously well with Claude Mythos Preview.
Anthropic wants to make its own drugs. Alongside the Claude Science launch we mentioned Tuesday, Anthropic has now told The Verge it plans to develop its own therapeutics, targeting “neglected” diseases. It’s one of the more direct moves by a frontier lab into a customer’s industry — Anthropic sells Claude to pharma companies it would now be competing with — and the announcement is thin on specifics: no diseases named, no partners disclosed, no responses to press queries.
Mistral’s Leanstral 1.5 finds bugs while proving theorems. Mistral released Leanstral 1.5 under Apache 2.0, a 119B-parameter MoE (6B active) that saturates miniF2F and solves 587 of 672 PutnamBench problems. More interesting than the benchmarks: agentic proof engineering in Lean 4 uncovered five previously unknown bugs across 57 open-source repos, including an integer overflow in a varint decoder. The HN thread has some fair pushback — the comparison charts pit Leanstral against models from six months ago, and one commenter wondered why Lean 4 got priority over Isabelle/HOL or TLA+.
The PNG-your-prompts trick. A developer released pxpipe, a local proxy that renders bulky Claude Code context — system prompts, tool docs, history — into PNGs before sending, since image token costs are fixed by pixel count rather than content density. Claimed savings: 60–70% on dense code and JSON. HN commenters are skeptical it’s really a win: Anthropic’s backend likely OCRs the images anyway (burning more compute, not less), and someone who tried this with OpenAI last year found completion tokens ballooned enough to erase the savings. A pricing loophole waiting to close, probably.
A CVE spike that coincides with Mythos Preview. Epoch AI flagged that high- and critical-severity CVE disclosures from Microsoft, Google, Apple, and AWS jumped more than 3.5x above previous monthly records in June 2026, matching the timing of Claude Mythos Preview and Anthropic’s Project Glasswing (10,000+ vulnerabilities claimed, alongside OpenAI’s parallel Daybreak effort). Commenters split between “defensive discovery is finally scaling” and “someone with early Mythos access leaked it,” with one poking a hole in the timeline — the model was announced in April but released in June, so starting the clock in April is convenient.
DeepMind’s union talks stall out of the gate. Wired reports that the first meeting between the CWU and DeepMind featured only HR staff, no senior leadership. Union rep John Chadfield called it a “time-wasting exercise,” and employees say Google has been restricting internal chats about unionization and reprimanding organizers. DeepMind disputes the framing and says next steps were agreed on.
DeepMind partners with A24. In a less contentious DeepMind story, the lab announced a research partnership with A24 to develop AI filmmaking tools, with Google also taking a financial stake in the studio. Details are vague — no specific projects, no technical outputs — but the framing is “artists shape the tech, DeepMind gets creative feedback.” Expect more of these lab-studio pairings.
Local LLM discourse gets a reality check. James O’Brien’s guide to running SOTA models locally spans $2k (dual 3090s for Qwen3-27B) to $40k (quad RTX PRO 6000s for a pruned GLM-5.2). The HN response is unsentimental: $40k for “almost-Opus” is roughly 17 years of a $200/month Claude subscription, and the claim that GLM-5.2 runs comfortably on that rig is disputed — real inference reportedly wants 8x H200s. Commenters point to an M5 MacBook Pro at 48GB unified memory or Intel’s Arc B70 as saner starting points. Relatedly, righttointelligence.org is advocating for legal protection of local AI, though the HN thread notes the site cites no specific pending legislation and Cloudflare briefly flagged it for suspected phishing.
Two more from the developer tools corner. Mcpsnoop is a transparent proxy and TUI for inspecting real MCP traffic between clients like Claude Desktop and their servers — a gap the official MCP Inspector doesn’t cover. And a blog post on the “short leash” method for AI coding argues for tight, incremental review over autonomous runs; the HN thread is divided between “this is just how competent people already work” and “you’re wasting a capable model by hand-holding it.”
That’s the digest. If you find yourself tempted by the PNG-prompt trick, maybe wait a week — that loophole has a short half-life.