Tokens & Signals · Friday, May 29, 2026

Anthropic’s Trillion-Dollar Ascent: The New Market King

claude-opus-4.8grok-build-0.1gpt-5.2-codexclaude-opus-4-8-xhigh-effortgemini-3.1-prolfm2.5-8b-a1bstep-3.7-flashfaradayanthropicopenainvidiamicrosoftxaiamazonrobinhoodliquid-aiinherent-labsstepfuncomputer-useagentic-financemodel-efficiencycoding-agentscompute-costsmultimodalityedge-aiautonomous-researchwilldepueclementdelanguevictortaelinsama
Tokens & Signals for 5/29/2026. We scanned ~1,200 Twitter accounts (1168 tweets), 13 subreddits (54 posts), Hacker News (8 stories), 6 newsletter posts, 3 podcast episodes, 206 Discord messages, and leaderboard data for you. Estimated reading time saved: ~11 hours.

TLDR & AI Twitter Recap

* Anthropic is officially the biggest player in the room, hitting an eye-watering $47B ARR and a $965B valuation. x.com/scaling01/status/2060121210533359911

* OpenAI just brought native "computer use" to Windows 11 — models can now control your mouse and keyboard and run apps directly. x.com/OpenAI/status/2060428604727771421

* NVIDIA and Microsoft are dropping cryptic hints about a "new era of PC" at Computex on June 2 — probably an Arm-based N1X chip with integrated graphics. x.com/CodeByPoonam/status/2060421518870634950

* Claude Opus 4.8 is live with "Effort Control," a new dial that lets you trade reasoning quality for lower compute costs. x.com/theo/status/2060172445592789064

* xAI launched the Grok Build API (0.1) for just $1/$2 per million tokens, going straight after high-volume coding agents. x.com/xai/status/2060392249402552457

* @willdepue on enterprise mood: "Amazon killed their internal AI leaderboard because nobody wants to pay for 'AI for the sake of AI' anymore." x.com/willdepue/status/2060149225606779204

* @ClementDelangue on IP ethics: "The friction between rapid, open innovation and legally-sourced training sets remains our industry's biggest unresolved tension." x.com/ClementDelangue/status/2060175330665508917

* Robinhood is moving into agentic finance, letting you hook AI agents up to their API to trade stocks with set guardrails. news.ycombinator.com/item?id=48326659

* Engineering teams are hitting 3,000 tokens/second on standard GPUs — making real-time agents finally feel instant. x.com/VictorTaelin/status/2060398243125531116

* @sama on the compute paradox: "The cost of intelligence is dropping 10x every 18 months but the usage is rising 100x."

Go deeper on what matters to you

Tap to expand

Best to Build With Today

* Codinggpt-5.2-codex leads the LiveBench coding leaderboard.

* Reasoningclaude-opus-4-8-xhigh-effort is currently the top performer for deep logic.

* Chatgemini-3.1-pro is the overall champion on the Chatbot Arena leaderboard.

* Open-sourceLFM2.5-8B-A1B by Liquid AI for efficient, long-context edge applications.

* Value pickgrok-build-0.1 at $1/M input tokens is the best deal for high-volume agentic coding.

Deeper Dives

💼 Industry & Business

Anthropic ARR reaches $47B

From $3B to $47B in a single year. That's not a growth curve, that's a vertical line. A new $65B funding round puts Anthropic at a $965B valuation, making them the most valuable AI company in the game right now.

� Twitter�️ Podcast

Microsoft data: AI costs exceed human labor

The efficiency dream is running into a very expensive wall. Large enterprises are finding that running AI agents can cost more than just hiring people — Uber reportedly burned through its entire annual AI coding budget in four months, and Microsoft has had to throttle access to specific tools as agentic token costs spiraled out of control. Turns out infinite efficiency gains aren't so infinite.

� Reddit

Amazon ends internal AI usage leaderboard

Executives pushed back hard on "using AI just for the sake of using AI," and the internal leaderboard got the axe. It's a clear signal that companies are moving past the experimentation phase and actually asking whether any of this stuff pays off.

� Twitter� Reddit

🧠 Models & Research

Opus 4.8 performance and efficiency analysis

Claude Opus 4.8 is built for agentic workflows, scoring 69.2% on SWE-bench Pro. The headline feature is "Effort Control" — basically a slider that trades latency for reasoning depth, which matters a lot when long-horizon tasks can get expensive fast.

� Twitter� Reddit

Inherent Labs launches to automate AI research

Former DeepMind researchers raised $50 million to build "Faraday," a system designed to automate the scientific research process itself. The goal is AI agents that can recursively improve — which is either very exciting or the setup to a lot of sci-fi plots.

� Twitter

StepFun releases 196B MoE model

Step-3.7-Flash is a new multimodal Mixture-of-Experts model hitting 56.26% on SWE-Bench Pro, with only 11B active parameters. Sparse models are getting seriously competitive and this one makes a strong case.

� Twitter� Reddit

🚀 Products & Launches

OpenAI integrates Codex with Windows

Codex is now native on Windows 11, with full access to PowerShell, WSL2, and direct GUI control. It's built for persistent agentic threads and can visually inspect and interact with graphical interfaces — so yeah, it can actually see what's on your screen and do stuff about it.

� Twitter

xAI launches Grok Build API

Grok Build 0.1 is in public beta, tuned specifically for software engineering tasks like debugging and refactoring. At $1/M input and $2/M output, it's an aggressive price point aimed squarely at developers building autonomous coding agents.

� Twitter

Robinhood enables stock trading for AI agents

Robinhood is letting users connect AI agents to trade stocks and manage funds in a new beta program. There are built-in safety guardrails and manual approval steps, so your portfolio won't go completely rogue. Probably.

� Hacker News

Funding & Deals

* Inherent Labs raised $50 million from Index Ventures to build autonomous scientific research agents.

Launches

* Grok Build 0.1 — xAI's new agent-optimized coding model with 256K context.

* Step-3.7-Flash — Multimodal MoE model from StepFun with 196B total parameters.

* LFM2.5-8B-A1B — Liquid AI's edge-optimized model trained on 38T tokens.

Closing thought: Everyone's racing to build agents, but the industry is quietly shifting from "how much can we build?" to "how much is this actually costing us?" The winners won't just be the ones building the smartest agents — they'll be the ones building the most efficient ones.