Tokens & Signals for 4/27/2026. We scanned ~1,200 Twitter accounts (1658 tweets), 13 subreddits (82 posts), Hacker News (21 stories), 7 newsletter posts, 9 podcast episodes, 379 Discord messages, and leaderboard data for you. Estimated reading time saved: ~18 hours.
* OpenAI and Microsoft are divorcing their revenue-sharing deal. Microsoft sticks around as a cloud provider, but the AGI clause is dead and OpenAI is now fully multi-cloud. Big move toward an IPO. x.com/sama/status/2048755148361707946
* Ineffable Intelligence, led by David Silver (the AlphaGo guy), just raised a wild $1.1B seed round at an $8B valuation. They're building something called a "Neural-Logic Fabric" for autonomous reasoning. x.com/IneffableLabs/status/2048775952876851547
* DeepSeek-V4-Pro dropped: A 1.6T parameter MoE model that beats previous open-source benchmarks by 12%. Oh, and they slashed input cache pricing by 90%. x.com/deepseek_ai/status/2048440764368347611
* @karpathy on the "vibe coding" wave: "Software is eating the world, and now English is eating software."
* China just blocked Meta's $2B acquisition of Manus AI, citing national security concerns. Autonomous research agents are apparently "strategic assets" now. x.com/testingcatalog/status/2048724839704252655
* Anthropic is in hot water over "hallucinated" billing. Developers are getting charged for massive API spikes that don't match their logs, and refund requests are sitting in limbo for weeks. x.com/valigo/status/2048462664897016056
* @theo on the billing chaos: "When your billing system hallucinates harder than your models, you have a problem." x.com/theo/status/2048456227538231751
* GitHub Copilot is killing flat-rate subscriptions in favor of usage-based billing. Heavy agentic usage blew up the old economics. news.ycombinator.com/item?id=47923357
* Cognition AI launched "Devin for Terminal." Their coding agent now runs natively in your local shell, with cloud handoff for the heavy stuff. x.com/cognition/status/2048821234281181302
* 4TB of data from 40,000 AI contractors was stolen from Mercor. Second major security breach this month tied to the OpenAI contractor pipeline. news.ycombinator.com/item?id=47919630
* An amateur mathematician used ChatGPT 5.4 Pro to crack a 60-year-old math problem in 80 minutes. The model connected graph theory and topological combinatorics in ways human experts had never tried. reddit.com/r/ChatGPT/comments/1swn1bs/chatgpt_5...
Best to Build With Today
* Coding: gpt-5.4-xhigh (Top-tier agentic coding performance).
* Reasoning: claude-opus-4-6-thinking-auto (Leaderboard topper for hard logic).
* Math: gpt-5.5-xhigh (96.3 on LiveBench Math).
* Chat: gemini-3.1-pro-preview-high (Overall ELO leader).
* Open-source: DeepSeek-V4-Pro (1.6T MoE, high performance and low cost).
Deeper Dives
💼 Industry & Business
OpenAI-Microsoft Partnership Restructured
OpenAI and Microsoft have ended their profit-sharing deal, effective 2030. Microsoft stays on as a key cloud partner, but OpenAI can now sell its models across AWS, Google Cloud, and anywhere else. The AGI-triggered "kill switch" clause is gone too.
Why it matters: OpenAI gets the operational freedom to run as a truly independent, publicly traded company — no more Microsoft veto power hanging over everything they do.
� Twitter� Hacker News
China Blocks Meta's Manus AI Acquisition
Chinese regulators killed Meta's $2 billion bid to acquire autonomous agent startup Manus AI. Officials flagged "national security" and data sovereignty concerns, essentially calling Manus's autonomous research tech too valuable to hand over.
Why it matters: Agentic AI is now officially geopolitical. Global consolidation in this space just got a whole lot harder.
� Twitter� Hacker News
Mercor Security Breach
Mercor got hit, and attackers walked away with 4TB of data — voice samples and records from 40,000 AI contractors.
Why it matters: The AI industry's dependence on third-party data labelers is a serious, under-secured supply chain problem. This won't be the last breach.
� Hacker News
🧠 Models & Research
GPT-5.5 and "Stream-of-Consciousness"
OpenAI's latest, GPT-5.5, uses native "stream-of-consciousness" processing to work through complex multi-step reasoning. It also cuts latency by 40% on long-context tasks.
Why it matters: It's becoming the blueprint for how agents should plan and think before they act.
� Twitter
DeepSeek-V4-Pro
A 1.6T parameter MoE model that clears previous open-source benchmarks by 12%. Their new "expert-routing distillation" training approach is more efficient and 25% less energy-hungry.
Why it matters: DeepSeek keeps finding ways to push high-end inference costs down, and the rest of the industry has to keep up.
� Twitter
The 60-Year-Old Math Problem
An amateur researcher used ChatGPT 5.4 Pro to solve a discrete geometry problem that had stumped mathematicians for six decades — just by asking the model to connect different math domains.
Why it matters: AI is becoming a genuine discovery engine, finding non-obvious connections that human experts miss because they're too deep in their own lane.
� Reddit� Hacker News
🔥 Takes & Drama
Anthropic Billing Backlash
Developers are getting charged for "hallucinated" API spikes with no corresponding usage in their logs, and refund tickets are going nowhere for months.
Why it matters: Enterprise trust lives and dies on billing transparency. If the invoice doesn't match the work, developers start shopping around.
� Twitter� Discord
Funding & Deals
* Ineffable Intelligence: Raised $1.1B (Seed) at an $8B post-money valuation. Led by Sequoia and a16z, they are building neuro-symbolic superintelligence.
Launches
* Devin for Terminal: Cognition AI's agent now lives in your shell for local/cloud hybrid coding workflows.
* DeepSeek-V4-Pro: 1.6T MoE model with massive cost cuts for input caching.
Closing thought: The shift from "chatbots" to "autonomous agents" is changing everything — how we bill for compute, who controls the strategic silicon, all of it. We're leaving the app-store era and moving into an agent-first world, and honestly, it's happening faster than anyone expected.