Tokens & Signals for 6/24/2026. We scanned ~1,200 Twitter accounts (1236 tweets), 13 subreddits (66 posts), Hacker News (14 stories), 5 newsletter posts, 3 podcast episodes, 153 Discord messages, and leaderboard data for you. Estimated reading time saved: ~11 hours.
* OpenAI and Broadcom just unveiled "Jalapeño," a custom inference chip they built in nine months flat. They're going full-stack to slash compute costs. x.com/gdb/status/2069809298612621629
* Google DeepMind gave Gemini 3.5 Flash "Computer Use" capabilities — it can now move your cursor and click buttons like an actual human. x.com/testingcatalog/status/2069854445957996996
* OpenAI dropped GPT-5.5 Instant for Pro and Plus users, promising 40% faster time-to-first-token. x.com/OpenAI/status/2069843083701915755
* @karpathy on the new agent paradigm: "The cursor is the new CLI and the browser is the new terminal. We're all just learning a new API."
* Fable 5 is rumored to be making a comeback via Claude Code; Polymarket has it at 63%. x.com/kimmonismus/status/2069818134719004878
* The EU is funding "EUROPA," a 400B parameter sovereign AI model built to break its reliance on US tech. reddit.com/r/singularity/comments/1ue8yy5/the_e...
* Meta is the only major lab refusing to sign the government's voluntary AI safety review agreement. Things are getting tense in D.C. x.com/AndrewCurran_/status/2069555826029965773
* Qwen released AgentWorld-35B-A3B, a lean Mixture-of-Experts agent model that's putting up big numbers on long-horizon tasks. news.ycombinator.com/item?id=48654351
* @MTSlive on the DeepMind brain drain: "The constant flow of top-tier talent from Google to Anthropic is starting to look less like a trend and more like a permanent shift in power." x.com/MTSlive/status/2069853573098434949
* Stripe, OpenAI, Anthropic, and Bill Gates backed "Intercept" with $500M to go after respiratory viruses. x.com/kimmonismus/status/2069809077136339120
Best to Build With Today
* Coding — gpt-5.2-codex is the top performer for code generation.
* Reasoning — claude-opus-4-8-xhigh-effort is the current king for hard logic and math.
* Chat — gemini-3.1-pro is sitting at the top of the Arena ELO rankings right now.
* Agentic Coding — glm-5.2 is leading the pack for autonomous coding workflows.
* Value pick — glm-5.2 for coding; it runs with the top-tier models at just $0.25/task.
Deeper Dives
💼 Industry & Business
OpenAI and Broadcom Unveil Custom 'Jalapeño' Inference Chip
OpenAI and Broadcom teamed up to build "Jalapeño," a custom ASIC designed specifically for high-throughput, low-latency AI inference. They built it in nine months, and the whole point is to cut OpenAI's dependence on third-party GPU providers.
Why it matters: This is a massive strategic shift — OpenAI is moving toward full-stack control and starting to look like a hardware-independent infrastructure player.
� Twitter� Hacker News
Meta Refuses Government Voluntary AI Review Agreement
Meta is the only major frontier lab holding out against a U.S. government push for voluntary, pre-release safety reviews. Their argument: the framework just doesn't work with an open-weights strategy.
Why it matters: This standoff is putting a sharp point on the tension between open-source philosophy and federal oversight.
� Twitter� Reddit
Intercept Launches to Eradicate Respiratory Viruses
Intercept is a newly formed organization with $500M behind it from Stripe, OpenAI, Anthropic, and Bill Gates. The goal is to use advanced AI and biological modeling to wipe out common colds.
Why it matters: It's a high-profile moment where AI lab heavyweights are planting a flag directly in public health infrastructure.
� Twitter� Reddit
Senior Google DeepMind Researchers Depart for Anthropic
Jonas Adler and Alexander Pritzel have left Google DeepMind for Anthropic, adding to a notable pattern of talent moving in that direction.
Why it matters: Talent follows mission — and Anthropic keeps consolidating the safety-focused research community.
� Twitter
🧠 Models & Research
Google DeepMind Releases Gemini 3.5 Flash 'Computer Use'
Google rolled out "Computer Use" for Gemini 3.5 Flash — the model can now read screens, move cursors, and click things on its own. It scored 78.4 on the OSWorld-Verified benchmark.
Why it matters: This puts real pressure on Claude's similar agentic features. The race for OS-level AI control is heating up.
� Twitter� Reddit� Hacker News
EU Funds 'EUROPA' 400B Parameter Open-Source Model
The EU picked a consortium to build "EUROPA," a 400B parameter foundation model running on European supercomputing infrastructure — a sovereign alternative to US-built AI.
Why it matters: Europe is making a serious push for digital independence, and this is the most concrete step yet.
� Reddit
Qwen Releases AgentWorld-35B-A3B Multimodal Agent Model
Qwen's AgentWorld-35B-A3B uses a 35B parameter mixture-of-experts architecture to stay lean and fast, scoring 82.4% on the AgentBench suite.
Why it matters: The industry is clearly moving toward smaller, high-efficiency models for long-horizon planning — and this one makes a strong case.
� Hacker News� Reddit
Anthropic Fable 5 Rumored to Return via Claude Code
Strings leaked from recent Claude Code updates are pointing to a Fable 5 comeback, possibly with weekly usage limits baked in for subscribers.
Why it matters: A high-tier model return paired with a legal challenge could shake up how AI access is regulated globally.
� Twitter� Reddit
Funding & Deals
* Intercept raised $500M from Stripe, OpenAI, Anthropic, and Bill Gates to develop treatments for respiratory infections.
Launches
* Gemini 3.5 Flash 'Computer Use' — Available via API, letting models interact directly with OS interfaces.
* GPT-5.5 Instant — OpenAI's new low-latency model optimized for speed and complex intent.
Closing thought: Custom chips, screen-clicking models, and half-billion-dollar bio-startups — today was one of those days where the "AI future" felt like it jumped a few years forward in 24 hours.