Tokens & Signals for 4/16/2026. We scanned ~1,200 Twitter accounts (1234 tweets), 13 subreddits (54 posts), Hacker News (20 stories), 9 newsletter posts, 3 podcast episodes, 306 Discord messages, and leaderboard data for you. Estimated reading time saved: ~13 hours.
* Anthropic dropped Claude Opus 4.7. New tokenizer, "adaptive" thinking, 94.2% on GPQA Diamond — but heads up, token costs are going up. x.com/kimmonismus/status/2044787072947601796
* OpenAI launched "Codex for Almost Everything," pushing hard into agentic workflows on macOS — spreadsheets, cloud infra, 90+ plugin integrations. x.com/embirico/status/2044828916330983716
* Alibaba open-sourced Qwen3.6-35B-A3B, a 35B parameter MoE model efficient enough to run on two consumer GPUs while still crushing coding benchmarks. x.com/kimmonismus/status/2044780695361290347
* Physical Intelligence released π0.7, a robot foundation model that lets machines learn new physical tasks zero-shot from plain language instructions. x.com/svlevine/status/2044840590261796895
* @jukan05 on the chip market: "TSMC's Q1 report confirms the AI hardware supercycle is still very much alive, despite all the peak-valuation chatter." x.com/jukan05/status/2044650383512162502
* @rohanpaul_ai on new research: "Multi-lab study from MIT, Oxford, and CMU confirms that AI makes us faster but leaves us weaker at independent problem-solving later. Cognitive atrophy is real." x.com/rohanpaul_ai/status/2044860162016760104
* Figure.AI unveiled the 'Vulcan Balance Policy,' letting humanoid robots stay stable and keep working even with three failed leg actuators. x.com/adcock_brett/status/2044797356965757065
* Seedance 2.0 is live on Runway with 1080p resolution and better temporal consistency — quickly becoming a go-to for world generation. x.com/_akhaliq/status/2044812535044452435
Best to Build With Today
* Coding — gpt-5.2-codex (LiveBench leader for pure coding) or gpt-5.4-xhigh for agentic automation tasks.
* Reasoning — claude-opus-4-6-thinking-auto is the top performer for complex logic.
* Chat — gemini-3.1-pro currently holds the top ELO spot on the Chatbot Arena.
* Open-source — Qwen3.6-35B-A3B is the best new option for running locally, with MoE efficiency that actually works on consumer hardware.
* Voice — gemini-3.1-flash (now with Scene Direction tags for custom emotion and pacing).
Deeper Dives
🚀 Products & Launches
Anthropic Launches Claude Opus 4.7
Anthropic's latest model is built for extended context reliability and serious reasoning, especially on coding tasks. It cuts hallucinations by 15% over previous versions. The catch: the new tokenizer and "adaptive" thinking behavior can push token usage up by 1.0–1.35x.
Why it matters: If you're running automated agentic workflows, your cost assumptions just changed — time to recalibrate.
� Twitter� Reddit� Hacker News
OpenAI Launches 'Codex for Almost Everything'
OpenAI is done being a chatbot company. This is a full agentic platform — spreadsheet manipulation, cloud configuration, background automation on macOS. It runs on a distilled GPT-5 architecture and comes in at $0.05 per 1k tokens.
Why it matters: The shift from "chatting with AI" to "assigning tasks to AI" just got a lot more concrete.
� Twitter� Hacker News
Google Gemini 3.1 Flash Adds Scene Direction TTS
You can now drop [tags] directly into prompts to control emotional tone, pacing, and delivery in generated audio. Expressive voice without a price bump — that's a genuinely good deal.
Why it matters: Granular voice control is fast becoming a baseline requirement for next-gen interactive agents.
� Twitter
🧠 Models & Research
Alibaba Releases Qwen3.6-35B-A3B MoE Model
35B total parameters, but only 3B active per token — that's the MoE magic. It scores 73.4% on SWE-bench Verified and fits comfortably on dual high-end consumer GPUs.
Why it matters: High-performance agentic coding models are finally within reach for local setups. This one's worth trying.
� Twitter� Reddit� Hacker News
Physical Intelligence Launches π0.7 Robot Foundation Model
A 12B parameter transformer trained on 50,000+ hours of robot interaction data. It predicts joint torques directly from language and posts a 40% higher success rate on multi-stage assembly than its predecessor.
Why it matters: This feels like the "GPT-3 moment" for robotics — the move from narrow fine-tuning to genuine general-purpose generalization.
� Twitter
OpenAI Releases GPT-Rosalind for Scientific Research
Fine-tuned on the full PubMed archive and biochemical databases, it uses "self-consistent evidence synthesis" to check claims against peer-reviewed literature — and shows a 12% improvement in technical precision over standard models.
Why it matters: Generic models aren't cutting it for high-stakes research. Frontier labs are starting to take that seriously.
� Twitter� Reddit
💼 Industry & Business
Anthropic and OpenAI Expand in London
Both companies went all-in on London this week. Anthropic is scaling to 800 employees; OpenAI just opened a permanent hub. The race for European research talent is very much on.
� Twitter� Reddit
TSMC Posts Strong Q1 Earnings
NT$572.5B in net income, 66.2% gross margins. The AI hardware supercycle isn't slowing down — the numbers make that pretty hard to argue with.
� Twitter
Stripe Projects Integrates AI Providers
Stripe now supports direct integration with providers like Hugging Face, Cloudflare, and OpenRouter — meaning you can provision and pay for AI infrastructure straight from your Stripe dashboard.
� Twitter
Closing thought:
The gap between "chatting with AI" and "assigning tasks to AI" is closing fast. Between the robotics breakthroughs and the new agentic platforms dropping this week, the industry is clearly done just talking to these models — it's time to put them to work.