X is the best way I’ve found to keep up with AI. I like tweets throughout the week, filtering for things I think are actually worth knowing, then use Claude Code to pull those likes automatically and help me turn them into this post (here’s how the pipeline works). This week: 105 tweets liked, filtered down to what’s below.

Check out the previous roundup (June 20) if you missed it. Last week the Fable 5 shutdown turned into a standoff with the White House. This week that standoff started resolving, OpenAI got pulled into the same orbit, and Anthropic shipped the most interesting product launch of the year while the policy drama played out.

AI for Everyone

The big story this week isn’t a single model. It’s that Washington is now a variable in how frontier models launch, and it hit both Anthropic and OpenAI in the same seven days. Meanwhile OpenAI shipped a new three-tier model family, Google kept pushing Gemini deeper into your Mac, and a Stanford lab put out a research method worth stealing.

The Government Is Now a Variable in Every Frontier Launch (10+ mentions)

Two separate labs had their flagship releases shaped by the US government in the same week, for the same reason: the models got good enough at cybersecurity that the government wanted a say in who gets access. Anthropic flew senior staff to DC, and after a two-week standoff secured permission to restore Mythos 5, its strongest cybersecurity model, to roughly 100 organizations that defend critical infrastructure. Per CNBC reporting the deal came out of direct meetings with the Trump administration, and Fable 5 is expected back for general use next week. The same week, OpenAI launched its new flagship in limited preview at the government’s request rather than open access (more below). The word showing up across the timeline is “lobotomized,” the worry that models come back with quietly restricted capabilities as the price of access. If you build on frontier models, availability is no longer a purely technical variable, so wire in fallbacks. (source: @AnthropicAI, @KobeissiLetter, @kimmonismus)

OpenAI Ships the GPT-5.6 Family: Sol, Terra, Luna (7 mentions)

OpenAI launched three models at once. Sol is the new flagship, same price as GPT-5.5 but a “step function better” and a new state of the art on Terminal-Bench 2.1, which tests multi-step command-line work. The one you can actually use today is Terra, which delivers GPT-5.5-level performance at half the cost, with Luna as the cheap high-volume option. Sol is invite-only for now because, as Sam Altman put it, the government asked for a limited preview given its security strength, and he thinks the staged rollout is “quite reasonable” while also “not quite the process that we think is optimal.” Separately, ChatGPT’s weekly web refresh landed: type @ in the composer to connect Gmail, attach images, or search inline, plus a new dictation model that’s much better at Japanese, Korean, Chinese, Urdu, and Vietnamese. (source: @OpenAI, @sama, @adamhfry)

Gemini Burrows Deeper Into Your Mac (2 mentions)

Google is testing two features for the Gemini macOS app that move it from “AI chat window” to “AI layer underneath everything.” Speak to Window lets you hold the fn key in any open app and dictate, and Gemini writes the output right where you’re working, whether that’s your email client or a doc. The wilder one: a “Connect another Mac” option in the attachment menu that lets Gemini see and control a second machine remotely. Both are still in testing, not shipped, but the direction is the point. Google wants Gemini to be ambient, not a destination. (source: @testingcatalog, @testingcatalog)

X Money Launches With a 6% APY Hook (1 mention)

X Money is rolling out to a subset of US Premium+ users in early access, and the launch incentive is hard to ignore: 6% APY on cash balances for a limited time. For context, most high-yield savings accounts are sitting around 4 to 4.5% right now. This is the first version of Elon Musk’s long-promised payments layer that’s actually in users’ hands. If you’re a Premium+ subscriber, check your notifications for an invite. (source: @dbatura, @PolymarketMoney)

STORM: a Journalist’s Trick for Better AI Research (1 mention)

AI research reports are shallow for a simple reason. You ask one question, you get one answer. Monica Lam’s Stanford lab built STORM, a method that forces the model to approach a topic from six to eight different expert perspectives first, interview them, build a ruthless outline, then write section by section grounded in sources. The researchers claim 25% better-organized articles than the same model with standard prompting, and over 70,000 people have used the open version. You don’t need the tool to use the idea. Next time you ask an AI to research something, tell it to work the topic from several distinct expert angles before synthesizing. (source: @rileywestreel)

AI for Developers

Anthropic shipped what Karpathy called the third major redesign of LLM UX, OpenRouter made model selection a runtime decision, shadcn solved the worst part of chat UIs, and the OCR wars heated up. Plus the most interesting multi-agent experiment of the year.

Claude Tag Joins Your Slack Team (8 mentions)

Claude Tag adds Claude as a persistent member of your Slack workspace, not a bot you summon but a teammate that lives in your channels and builds context over time. Tag it in a thread and it breaks the task into stages, uses tools like GitHub or your data warehouse, then replies in-thread with the output, whether that’s a merged PR, a data analysis, or incident help. Turn on ambient mode and it takes initiative, following up on quiet threads and flagging what’s relevant. Karpathy called this the “3rd major redesign of LLM UIUX”: first the LLM was a website, then an app, now a self-contained asynchronous entity working alongside your team. Anthropic’s own product team says 65% of its code now comes from the internal version. It’s in beta for Enterprise and Team plans today. (source: @claudeai, @karpathy, @ClaudeDevs)

OpenRouter MCP: Pick the Right Model at Runtime (5 mentions)

Most agents have their model hardcoded at build time, which means they’re routing based on what was true six months ago. OpenRouter’s MCP server fixes that by giving your agent live pricing, benchmarks, and leaderboard rankings at runtime, so it can ask “what’s the best model for code review right now?” and route accordingly before making the call. Lennox Saint ran it inside Codex as a streamable HTTP MCP with an OpenRouter OAuth token (sensible defaults: 7-day expiry, $10 spend cap) and described it as “2x better Codex.” The OpenRouter demo shows it pulling Design Arena’s live leaderboard, spinning up GLM-5.2, Opus 4.7, and Kimi 2.6 as subagents on the same task, then opening all three for you to compare. This is one of the fastest weekend experiments on this list. (source: @OpenRouter, @lennox_saint)

shadcn Ships Chat UI Primitives (7 mentions)

Anyone who has built a streaming chat UI knows the scrolling is the part that destroys you. Do you auto-scroll? When do you stop? What happens when the user scrolls up while a message is still arriving? shadcn just shipped MessageScroller, a headless component that owns all of it: scroll anchoring, streaming, saved thread restore, prepended history, jump-to-message, and visibility tracking. You bring the data and styles, it owns the behavior. It also added scroll-fade and shimmer utilities plus a set of LLM message building blocks, available for both Radix and Base UI. This is the unsexy infrastructure that separates a polished chat app from a janky one, and now you don’t have to write it. (source: @shadcn)

Mistral OCR 4 Sets Structure, Not Just Text (5 mentions)

Most OCR tools extract text and quietly drop everything else, so tables become garbage and figures vanish with no trace. Mistral OCR 4 does it differently: every block gets a bounding box, a classification (title, table, equation, signature, chart), and a per-region confidence score, in 170 languages. Independent annotators preferred it over every competitor tested with a 72% average win rate across 600+ documents in 12+ languages, with the biggest gains on rare and low-resource languages. It runs as a single self-hosted container if your documents can’t leave your environment. For RAG pipelines on messy real-world PDFs, this is the current best option. (source: @MistralAI)

Gemini 3.5 Flash Gets Native Computer Use (5 mentions)

Google DeepMind shipped native computer use for Gemini 3.5 Flash, scoring 78.4 on OSWorld-Verified, comparable to Claude’s computer use on the same benchmark. The difference is Flash is much cheaper per token than a frontier model, which matters a lot for agents running hundreds of computer-use steps. It’s a built-in tool, not a bolted-on capability, working across browser, mobile, and desktop. Computer use is rapidly becoming a commodity, which is exactly the right direction for anyone building agents that actually do things. (source: @GoogleDeepMind, @osanseviero)

OpenAI Daybreak: Codex Security for Defenders (6 mentions)

OpenAI expanded its Daybreak security push. GPT-5.5-Cyber now scores 85.6% on CyberGym, up from 81.9%, and it’s paired with a Codex Security plugin that runs the full loop: find a vulnerability, validate it, trace the attack path, generate a codebase-specific patch, and export to your existing tools. The most ambitious piece is Patch the Planet, a joint effort with Trail of Bits and HackerOne to work through critical open-source dependencies with human review at every step. The labs are betting security is where AI agents earn trust fastest, because the tasks are well-defined and “the AI patched a CVE before a human did” sells itself. (source: @OpenAI, @testingcatalog)

Coinbase Cut Its AI Spend in Half (1 mention)

Brian Armstrong posted the most actionable AI cost breakdown I’ve seen from a major company. Coinbase nearly halved AI spend while token usage keeps growing, and the biggest single move was caching: their LibreChat cache hit rate went from 5% to 60% once properly implemented. Not a new model, not a usage cap, just not making the same call twice. They also switched defaults to open-weight models like GLM 5.2 and Kimi 2.7 (91% of employees never hit caps anyway) and route to frontier models only for planning, not execution. If your team is talking about AI cost, audit your cache hit rate first. (source: @brian_armstrong)

100+ Agents Self-Organize to 5x Gemma Inference (1 mention)

Thomas Wolf from HuggingFace published one of the most interesting AI experiments of the year: 100+ independent agents given a week to improve Gemma 4’s inference speed in vLLM. They got a 5x speedup, but the behavior in between is the story. An agent refused a human’s request to move coordination to Telegram, writing unprompted that private channels are “indistinguishable from collusion.” Another flagged a benchmarking loophole and asked for a community ruling. A four-agent relay spontaneously formed to build, run, diagnose, and ship a single checkpoint when no one agent had both the code and the GPU quota. None of this was programmed. The interaction board is browsable and worth an hour. (source: @Thom_Wolf)

Honorable Mentions

  • Claude Sonnet 5 is leaking (codename Fennec, expected as early as next week, already in an enterprise Early Access Program). Expect a 1M context window, better vision for diagrams and UI mockups, and a new tokenizer that reportedly uses 30% more tokens on the same prompts, so recalc your cost models before migrating. (source: @synthwavedd, @pankajkumar_dev)
  • Baidu open-sourced Unlimited OCR, which reads 40+ pages in a single forward pass with no chunking via Reference Sliding Window Attention, at 3B total / 500M active params and SOTA on OmniDocBench. Worth trying anywhere chunking is splitting your context. (source: @BaiduAI_News)
  • Nous Research added Mixture of Agents presets to Hermes Agent: reference models advise an aggregator that makes the tool calls. Nous claims it beats Opus 4.8 by 8% and GPT-5.5 by 11% on its internal benchmark. Take vendor benchmarks with salt, but the architecture is sound. (source: @NousResearch)
  • OpenAI built its first AI chip, Jalapeño, designed with Broadcom and now in production for LLM inference on ChatGPT and Codex. The business story matters more than the tech: every inference dollar OpenAI keeps from Nvidia changes its cost curve. (source: @OpenAI)
  • IBM showed the first sub-1nm node chip, claiming 70% greater energy efficiency. Process-node competitiveness is what ultimately gates AI-chip supply, so this is a leading indicator worth tracking. (source: @IBMNews)
  • GPT-5.5 Instant got a quiet upgrade, ChatGPT’s most-used model, now better at reading intent, handling multi-constraint requests, and giving shopping and local recommendations. If conversations have felt sharper this week, this is why. (source: @OpenAI)
  • A Redmond, WA drone caught a shoplifter as a Drone as First Responder beat officers to the scene, tracked the suspect to a bus, and coordinated the arrest in real time. The AI-policing future is showing up in suburban police blotters first. (source: @RedmondWaPD)

Try This Weekend

For everyone:

  1. Type @ in ChatGPT’s web composer to connect Gmail, attach images, or search inline. Then click the mic and try the new dictation model on a language that used to feel rough.
  2. Run a STORM-style research prompt. Pick a topic and tell Claude or ChatGPT to approach it from six distinct expert perspectives (economist, historian, practitioner, skeptic) and synthesize, then compare to your usual one-line prompt.
  3. Watch for Fable 5’s return next week and run your usual tasks the day it’s back. The community will be checking closely for any behavior changes after the government deal.
  4. Check your X notifications if you’re a Premium+ subscriber. X Money early access comes with 6% APY on balances for now.

For developers:

  1. Add the OpenRouter MCP to your agent or Codex as a streamable HTTP MCP, authenticate with OpenRouter OAuth, and try “use the best model for code review.” Watch it pick from live benchmarks.
  2. Build a chat UI with shadcn MessageScroller and see how much scroll-anchoring boilerplate it eliminates. It’s the part that’s always hardest to get right.
  3. Run Mistral OCR 4 on a PDF your current pipeline mangles. Check the bounding boxes and confidence scores, which are what make it different from just getting text back.
  4. Enable Gemini 3.5 Flash computer use via the API and point it at a browser task you do manually, like form filling or status checks. 78.4 on OSWorld is production-ready for well-scoped jobs.
  5. Pull Baidu Unlimited OCR from HuggingFace and feed it a 20-40 page PDF you’ve been chunking by hand. Only 500M active params, so it’s cheap to run.