Matt Pocock called it a poisoned chalice: Anthropic's June 15 change cuts programmatic Claude Code usage 5–20x, dressed up as a "dedicated monthly credit." Wes Bos burned $91 in under four hours running a single unsupervised coding session, which is the cleanest public datapoint of why Anthropic had to close the unmetered loop. Within hours, Sam Altman offered enterprises two months of Codex free to switch. The loss-leader era for agentic coding is over, and the next 30 days are a one-shot share grab.
The split underneath this is clean. Anthropic is choosing its buyer — Fortune 10 contracts, 500% NDR, run-rate revenue from $250M to $30B in two years — and repricing the developer subsidy accordingly. Cline open-sourced its full coding-agent harness the same day, topping Terminal-Bench 2, and that is the natural escape hatch for the power users Anthropic just squeezed. One camp is compounding on enterprise lock-in. The other is betting the open harness layer becomes the durable moat. Those two bets don't end in the same place.
Top developments
Anthropic — quietly ended Claude Code's programmatic-usage subsidy: from June 15, claude -p, the Agent SDK, GitHub Actions, and third-party apps built on the SDK move off subscription rate limits to a fixed monthly credit, which power users measured as a 5–20x cut. Within hours Sam Altman offered eligible enterprises two months of Codex free for switching. Anthropic is conceding that the loss-leader unmetered-loop era is over and OpenAI is treating the next 30 days as a one-shot share grab. x x
Anthropic — overtook OpenAI in business adoption for the first time on Ramp's index (34.4% vs 32.3%, with Anthropic share quadrupling year-on-year while OpenAI rose 0.3%), and CFO Krishna Rao disclosed run-rate revenue went from $250M to $30B in two years with 9 of the Fortune 10 now paying customers and NDR over 500% annualized. The enterprise distribution race has flipped at the same moment Anthropic is tightening the consumer-developer subsidy — the company is choosing the buyer it wants. x x
Notion — shipped a full developer platform in one go: a CLI (
ntn), Workers running on Notion's infra, database sync from any source, external-agent API, an Agents SDK, webhook triggers, and an Agent Orchestration layer that routes work between humans and tools like Decagon and Claude. Notion is repositioning from a doc to the universal context layer agents read and write to — competing directly with Linear's agent platform and the Salesforce/Atlassian data-layer plays a16z flagged this week. x xModal — is raising at a $4.5B valuation on roughly $300M annualized revenue, most of it from agent-driven sandbox usage per The Information. Inference-and-sandbox infra is now the second concrete agent-economy unit-economics datapoint after Cognition's $445M ARR — capital is repricing the picks-and-shovels layer for agentic workloads, not the model labs themselves. x
Cline — open-sourced its rebuilt coding-agent engine and shipped the Cline SDK, claiming the top spot on Terminal-Bench 2 with the same harness that powers its CLI and VSCode extension (plugins, sub-agents, cron jobs, multi-provider routing included). Lands on the day Anthropic restricts programmatic Claude Code usage; the open-source harness layer is the natural escape hatch and Cline is the first to ship a credible one. x
Notable discussions
World models vs LLM-only agents — Yann LeCun argued you cannot build a reliable agentic system without a world model and LLMs structurally lack one, drawing 1.9k likes and a personal retweet from LeCun; Deedy Das called world models the critical AI research direction with billions already committed; Dan Jeffries pushed further that five capabilities beyond world models are still missing, and Jonas Geiping flagged that modern coding agents lean on message-passing instead of proper training. The frame this week is shifting from "scale the LLM harness" to "the LLM alone was never the architecture" — and the people funding the post-LLM bet are getting a credible second wave of voices. x x x x
Sharp takes
Matt Pocock — calls Anthropic's June 15 pricing change a poisoned chalice — a 10x cut to claude -p dressed up as a monthly bonus — and says it's time to try Codex; the framing landed because the official announcement led with "dedicated monthly credit" rather than acknowledging the underlying tightening. x
Jeremy Howard — flags that Anthropic's policy quietly redefines "interactive" to mean "using an Anthropic front-end," so any claude -p or Agent SDK call now burns credits even when a human is in the loop — the "interactive use unchanged" line in the announcement is, in his read, materially misleading. x
Wes Bos — burned $91.34 in 3.5 hours running Claude Code with /goal to reverse-engineer an unfamiliar codebase end-to-end; the concrete bill is the cleanest public datapoint of what an unsupervised long-horizon coding session actually costs, and helps explain why Anthropic is metering the SDK lane. x
Demis Hassabis — restates his "cure every disease in a decade" claim on the back of the Isomorphic $2.1B close, projecting Isomorphic's first AI-designed cancer drug enters human trials this year and the endgame is per-patient drugs designed overnight; a useful counterweight to the day's coding-agent navel-gazing. x
Marc Andreessen — coined "AI vampires" for programmers running Codex/Claude Code: a16z portfolio leading-edge devs are 20x more productive year-over-year and sleeping less; argues marginal productivity theory predicts SWE demand rises (and the hiring data is rising), but warns software is now being summoned faster than it can be understood. x
Andrew Ng — pushes back on AI-jobpocalypse framing with the actual numbers: SWE demand rising, devs a rising share of new hires, AI-exposed industries seeing above-trend wage growth, open PM roles at a 2022 high — useful as a cold-water counterweight to the day's layoff and unemployability threads. x
Other news
Models & releases
Anthropic Claude Code weekly limits — raised 50% for all paid plans through July 13, paired with the programmatic-credit change x
Mythos Preview — first model to solve both UK AISI cyber ranges end-to-end (3/10 → 6/10 pass rate vs prior version), per Glasswing x x
Gemini 3.2 Pro + Flash — outputs surface ahead of I/O, with leaks showing math, watermark removal, object swapping x x
Uncensored Opus 4.6 — community work to run locally on 128GB Macs within weeks x
GLM 5.1 — hits 245 tokens/sec on Cerebras Wafer Pass x
Grok Voice Think Fast 1.0 — multi-step workflow voice agent ships x
Funding & deals
Modal — $4.5B valuation, $300M ARR led by agent sandbox demand (covered above) x
Sierra Platform — $950M Series E at $15B from Tiger Global and GV x
IREN + NVIDIA — $3.4B, 5GW AI infrastructure deployment contract x
Recursive — Richard Socher launches new lab for self-improving systems x
Devtools & coding agents
Notion Developer Platform — CLI, Workers, sync, Agent Orchestration, Agents SDK (covered above) x
Cursor cloud agents — run inside fully configured dev environments with cloned repos, deps, credentials x
Cline SDK — open-sourced harness tops Terminal-Bench 2 (covered above) x
LangSmith Engine — background agent that reads traces, identifies issues, suggests evaluators and code changes x
Linear agent platform — multi-source agent integration ships x
CodeRabbit Project Atlas — AI-native code review interface x
Supabase Agent Skills — agents respect RLS by default when building x
Opera Browser CLI — local agents control the browser x
Lovable — first agentic coding tool to achieve safety certification x
Relace Jacq — cloud coding agent built on small trained models x
Industry & policy
Anthropic Claude for Small Business — bookkeeping, ad-campaign, and business-insight agents x
Meta — 14,000 effective headcount reduction (8,000 cuts on May 20 + 6,000 cancelled roles) under Chief AI Officer Alex Wang x
Codex switcher offer — two months free Codex for enterprise teams that migrate within 30 days x
ServiceNow — exhausted annual Anthropic coding budget early x
Vercel AI Gateway data — Google leads production scale, Anthropic dominates coding spend, OpenAI growing fast post-5.4 x
Microsoft — moves Copilot to token-based billing after high spend x
OpenAI / Helion — board fired after rejecting Altman's Helion deal x
Cursor + xAI — staff meetings amid xAI layoffs and departures x
Thinking Machines — a third of founding team has left Mira Murati's startup x
Robotics & physical AI
Figure Helix-02 — humanoid robots run full 8-hour autonomous shifts at human performance levels x
Unitree — launches app store for humanoid robot skills x
Infrastructure & platforms
PostgreSQL BM25 — new search extension claims 23x performance over baseline x
LiveKit Agents — answering-machine detection ships for outbound voice agents x
ClickHouse vs Elasticsearch — 2–6x faster full-text analytics x
Floci — open-sources 13MB Go binary emulating AWS with 40+ services x
Security
TanStack worm guide — DeRonin publishes 10-step hardening playbook (npm cooldown, SHA-pinned Actions, IDE extension audit, emergency rotation) x
Supply chain + AI — AI-enabled vuln discovery accelerating worm publishing pace x
Workato Seven Factors — open framework for safe AI agent governance x
GitHub vs Puppeteer — incorrectly flagged Puppeteer as malware, blocking new releases x
Composer 2.9.8 / 2.2.28 — patches GitHub Actions token leak x
Continuing threads
FDE expansion — Google adds an FDE org following the Palantir/OpenAI playbook; Anthropic posts senior FDE roles (May 12/13 Notable) x x
Anthropic + Stainless — The Information's $300M acquisition recap continues to circulate (May 13 Top dev) x
Anthropic Claude for Legal — TechCrunch and Box ride the rollout from May 13 x
Conductor — drops GPT-5.5 as default harness in response to Claude Code SDK pricing change (May 12 tail) x