Mistral researchers use Claude Code daily; Le Chaton Fat pitched

Trump administration — concluded Monday's working-group meeting with Anthropic without lifting Fable 5 export controls; per WIRED via Hugo Lowell, Commerce/CAISI still believe Mythos can be jailbroken through Fable, while a WH official told Polymarket the admin "has suspicions China has accessed Mythos," and Andy Jassy is reported to have called Treasury Secretary Bessent directly about the vulnerability. The exec trip didn't work — a WH source warning of "a de facto licensing regime where every future model asks permission" is now the new policy fear, not the original ban. x x x x

Salesforce — signed an agreement to acquire Fin (formerly Intercom) for ~$3.6B in cash, with Eoghan McCabe staying CEO and Des Traynor leading R&D after close in Salesforce FY27 Q4. The cleanest late-stage SaaS-to-AI pivot exit yet — Intercom rebranded to Fin "weeks ago" off the back of its Apex model and Operator agent, then sold the entire customer-agent category into the incumbent CRM platform inside the same quarter. x x

Anthropic — was hit with a class-action lawsuit over Claude Max usage limits, with the filing alleging Max 5x delivers only ~3.5x Pro (not 5x), Max 20x ~6x Pro (not 20x), and Max 20x just ~1.7x Max 5x, on top of "unclear and unspecific" limit calculations. The first consumer-protection suit against a frontier lab to attack the subscription math itself — pricing-tier promises are now an SEC-style disclosure problem, not a developer-relations one. x x

Vercel — pushed Functions to a 30-minute max runtime on a new microVM-based Fluid compute platform, with Guillermo Rauch framing it as the conclusion of a multi-year compute investment that also underpins Builds, Sandbox, multi-concurrency, Active CPU pricing, and Secure Compute. The platform's runtime cap was the single biggest blocker to long-running agent workloads on Vercel — they just removed it and turned their compute layer into a direct AWS Lambda alternative for AI work. x x

Factory — launched 2.0, repositioning from a coding-agent product to a "software factory" platform that drives the full engineering workflow from tickets to incidents, with Codydearkland framing it as exposing multiple AI capabilities rather than a single agent. Third meta-harness pivot in two weeks after Vercel HarnessAgent and Databricks Omnigent — the layer above coding agents is now its own contested product category, and pure agent vendors are racing to escape it. x x

Mistral employees using Claude Code over their own models — turned into a same-day chorus questioning what European AI sovereignty actually means in practice: Pawel Huryn flagged that Mistral researchers internally use Claude Code, willchen500 amplified it, and plbiojout argued that Mistral employees use their own models "one day a week for feedback" and the rest of the time live in Claude or Codex, framing the company's roadmap as the Capgemini/Accenture consulting playbook rather than the frontier-model playbook. Lands at the exact moment Le Chaton Fat is being held up as the European answer to the Fable ban — the loudest internal signal against that thesis came from inside Mistral. x x

Mitchell Hashimoto — argues any agent-generated diff over 1500 lines means the problem needs to be decomposed; lays out his current daily pattern as "draw the owl" → atomize into incremental tasks → parallelize new agents on each, and warns the "if it works who cares what the code looks like" mindset assumes the agent has a perfect model of "works" — which it doesn't, especially on API design, perf, and compatibility. x x

antirez — pushes back hard on the "Chinese labs are just distilling US models via API" framing: API-based distillation is mathematically impossible at frontier scale, and anyone repeating the line doesn't understand how training actually works. x

Aaron Levie — the future of AI is distributed intelligence — networks of cheap specialized models with smart routing — not a single smarter frontier model; once model swap-out cost goes to zero, the routing layer becomes the load-bearing piece and frontier capability is just one input. x

Paul Graham — the real winner of the Anthropic export-control fight is open weights — every uncertainty event on a closed lab moves enterprise cost-of-capital math closer to owning the model rather than renting it, and the past week has been the loudest such event yet. x

Paweł Huryn — argues Fable 5 is not a "disabled model" but the Mythos model wearing a safety filter — what export controls actually pulled is the wrapper, not the underlying capability, which is why Anthropic's framing of the incident keeps drifting between "we restricted access" and "we lost our flagship." x

HowToPrompt__ — relays a research analysis of Claude Code's leaked source: only 1.6% of the codebase touches the model, the other 98.4% is a 7-mode permission system, a 5-layer context-compaction pipeline, subagent worktree isolation, and four extensibility hooks. The whole product is a fortress of deterministic software around an existing model — Anthropic's bet is that managing the LLM beats making a smarter one. x

Thorsten Ball — AI is the first major tech revolution that contributes to a shortage of software engineering work rather than a cull — productivity gains are getting reinvested into more projects, more rewrites, and more agentic surface area, not fewer engineers. x

Cartesia Sonic-3.5 + Ink-2 — streaming TTS and STT models released, Cartesia claims #1 position on both axes simultaneously x

Kimi open-source coding model — released at 180 tok/s with a stated 1M context for max reasoning x

GLM 5.2 — combines 1M context with max reasoning, recirculating as the open-source coding leader x

Gemma 4 12B Coder — runs offline local code generation on consumer hardware x

Cloudflare AI Gateway — adds durable-objects buffering for inference calls x

OpenRouter Fusion — multi-model panel beats solo Fable 5 and Opus 4.8 on benchmark x

Conductor Pro — first paid plan opens with cloud workspaces for remote parallel agent execution x

Mastra — integrates Claude Code, Cursor, and Codex agents under one harness x

Argent — Claude Code now controls iOS and Android simulators simultaneously x

Greptile t-rex — runs PR branches in sandbox to find bugs before merge x

OpenAI Codex — gains Chrome DevTools Protocol support for browser automation x

Field Theory v0.3.2 — open-source context-management tools for agents x

Generative UI for Claude Code/Codex/Pi — enables real-time agent-driven UI rendering x

Printing Press — converts websites into CLI tools agents can call x

Linux kernel index for agents — new tool indexes the entire kernel for AI consumption in three minutes x

OpenAI Partner Network — launched with $150M investment and target of 300K certified consultants by end of 2026 x

OpenAI Codex — offering $1,200 in credits and 6 months ChatGPT Pro free to developers with a public GitHub repo x

Anthropic privacy policy — now requires age and identity verification for some users x

Anthropic Agent SDK — paused planned quota changes after user backlash, restoring programmatic Claude Code subscription use x

Google Open Knowledge Format — formalized as standard for AI-readable websites x

MDN MCP server — Mozilla ships MCP endpoints for AI agents and IDEs x

NVIDIA DGX Station — desktop AI workstation with 768GB unified memory ships x

Hetzner — triples US instance pricing while grandfathering existing customers, triggering migrations x

Supabase Realtime — adds binary message support to cut JSON overhead x

LangChain + Fireworks — fine-tuned Qwen judge model claims 100x cheaper agent evals x

NVIDIA Romania — reportedly weighing a $4B AI data center investment x

22-year-old MIT grad — raises $21M Series A with AI deployed in Fortune 100 customers x

n8n — German AI startup doubles down on London expansion x

Midday founders — join Ramp to bring design-first finance software to 70K+ businesses x

React Native 0.86 — ships comprehensive edge-to-edge support with no breaking changes x

React Navigation — automatic deep linking for shared mobile URLs x

Fable internals leak (cont.) — dbreunig flags the leaked Fable system prompt repeats copyright guidance six times at varying severity x

Anthropic-uses-Claude-Code internally (cont.) — 100% of PRs and 80–90% of code review at Anthropic now go through Claude Code, recirculating with new amplifiers x

AMD 235B-on-a-lunchbox (cont.) — Lisa Su's $1,499 Ryzen AI Max+ 395 demo keeps recirculating with no new product news x

Local-AI insurance advisories (cont.) — same talk track as 06-14's sovereign/local Notable disc, now with "own intelligence not rent it" wrappers x x

Mistral Le Chaton Fat (cont.) — extends 06-15's Top dev with parody amplifiers (Super-Chaton Fat 500T, "EU bans Le Chaton Fat") and a scaling01 critique of the perf claims x