Gergely Orosz flagged it plainly: a developer at a large AI-bullish enterprise told him they're going "HARD on GPT-5.6 Sol" because Anthropic hasn't updated Fable's data-retention policy, and their inputs would be stored. Not a model-quality call — a policy call. That's the same week Apple filed a court complaint alleging OpenAI ran a coordinated campaign to recruit 400-plus Apple employees and have candidates bring physical hardware to interviews for show-and-tell. Two frontier labs, two different liability problems. The legal and compliance surface is now where enterprise deals break.
The underlying split is between labs racing on capability and enterprises making decisions on paperwork. Anthropic is posting 70%-plus gross margins and annualizing past $50B ARR — and losing named accounts to a retention-policy gap. OpenAI is defaulting into Microsoft 365 Copilot across Word, Excel, and Teams while simultaneously defending a trade-secret lawsuit from the platform it needs to distribute hardware through. Both are compounding fast. Both are accumulating exposure at the same rate. The moat everyone is pricing is model quality; the bind that actually moves enterprise contracts is legal and policy posture, and almost no frontier lab has caught up to that.
Top developments
Apple — sued OpenAI, accusing it and its hardware chief of running a coordinated trade-secret theft campaign to build its upcoming AI-devices suite; court filing alleges 400+ Apple employees recruited and candidates directed to bring physical Apple parts to interviews for "show and tell." First open legal front between a hyperscaler platform owner and a frontier lab as OpenAI's hardware roadmap collides with Apple's, and Cupertino litigating rather than partnering signals the AirPods-camera bypass is now the strategy. wsj.com
Claude Code — shipped an in-app browser on desktop; Claude can pull up docs, designs, and any site, read them, click through, and interact the same way it does with local dev servers, with sandboxed configurable sessions. The agent-plus-browser primitive Anthropic had been shipping around (Cowork, Chrome extension) folds into the coding surface itself — one fewer context handoff between the tool and the site. x
OpenAI GPT-5.6 — becomes the preferred model in Microsoft 365 Copilot, per Sam Altman. Microsoft's flagship enterprise AI surface now defaults to OpenAI's newest family across Word, Excel, PowerPoint, Teams, and Outlook — running in parallel to Anthropic's Claude for Office GA from May. Both frontier labs are now inside the same vendor's productivity stack, and the model-picker is the new user default, not the platform. x
Notable discussions
GPT-5.6 Sol destroying files in agent mode — Matt Shumer's Sol ran
rm -rf /Users/mattsdevboxand wiped his Mac, saying he trusts Fable 1000x more; Crémieux independently hit the same "delete then panic" behavior and noted the model card had documented it; levelsio said it's why he refuses to run coding agents on his laptop, and doodlestein/morganlinton pushed pre-tool-use hooks as the fix. The first public safety cluster around Sol's agentic mode lands the same week Ultra went GA — the harness, not the model, is what saves your filesystem. x x x
Dan Shipper — reads the race as a strategy fork: Anthropic scales Fable as big and expensive as needed to reach RSI and break away, while OpenAI trains a smaller, efficient Sol they can serve cheaply at scale. Sol is the better daily driver for most today, but the compounding-benefit prize sits with whoever gets to recursive self-improvement first. x
Chamath Palihapitiya — says his company's AI token costs are doubling every 45 days while downstream productivity gains sit at ~5% max, and that scaling has effectively asymptoted; his advice is to consider getting liquidity now before the reckoning shows up in prices. First blunt operator saying the token curve is not cashing out into the productivity curve. x
Han Xiao — position paper argues LLMs can't take the Kuhnian "jump": feed one every paper up to 1905 and it still doesn't invent general relativity. Induction and deduction don't reach the third move — inventing the frame itself — and nothing about scaling changes that. Timely counter to the "Sol Ultra proved a 50-year math conjecture" cycle happening the same day. x
Alex Veremeyenko — pxpipe renders your Claude Code context as PNGs before it leaves the machine, dropping end-to-end Fable bills 59–70% by exploiting that images are billed by pixel dimensions, not by text density. Honest about the tradeoff — it's lossy, so hashes and IDs stay as text. First serious cost-cut hack that doesn't require model-swapping. github.com
Gergely Orosz — flags a dev at a large AI-bullish enterprise saying Anthropic hasn't updated Fable's data-retention policy, meaning the vendor would store their inputs, so they can't use it and are going "HARD on GPT-5.6 Sol" as a result. A concrete example of Anthropic losing meaningful enterprise share on policy, not model quality. x
Amjad Masad — while AI makes coding less rigid, infrastructure is going the other way: infra teams are writing formal specs for the first time, and the faster the surface layer moves, the more deterministic the runtime beneath has to be. The taste split of the year — soft on top, hard underneath. x
Other news
Models & releases
GPT-5.6 Sol Ultra — proved the 50-year-old Cycle Double Cover Conjecture via 64 subagents in under an hour, per OpenAI research cdn.openai.com
Alibaba Wan-Streamer — real-time omni-modal understanding and generation x
Cartesia — new realtime transcription model for production voice agents x
Kyutai MuScriptor — open audio-to-MIDI model, multi-instrument, multi-genre x
Netflix — releases video datasets and models on Hugging Face x
Microsoft SkillOpt — boosts ChatGPT accuracy from 41% to 80% without parameter changes arxiv.org
GitHub Copilot GPT-5.6 — three capability variants now available in Copilot github.blog
Devtools & coding agents
Cursor side chats — durable parallel agent conversations you can @-mention back into the main thread x
shadcn/typeset — reusable styled component that fixes unstyled markdown HTML rendering x
OpenCode 2.0 beta — new API plus hot-context feature that avoids cache busting x
Google AI Studio pretty URLs — free custom URLs for deployed apps x
OpenSWE — open-source alternative to proprietary coding-agent systems github.com
Toolport for teams — unified MCP server stack across Claude/Cursor/Codex, cuts tool-context tokens 300k→800 x
finance-skills — open-source Claude Code plugins for DCF, LBO, KYC, and equity research crossed 3k GitHub stars github.com
Tencent Agent Memory — open-source local long-term memory layer for AI agents github.com
AWS AgentCore runtimes — new agent infrastructure primitive x
Sierra Pinecone — internal agent automates 90% of Sierra's coding, analytics, and busywork on a model-agnostic harness sierra.ai
Every "Tend" — open-source loop for tending inbox, hiring pipeline, or CS queue inside ChatGPT Work github.com
Infrastructure & platforms
pgrust — clean-room PostgreSQL rewrite in Rust passes the full 46,066-query Postgres 18.3 regression suite, disk-compatible, with a WIP branch claiming 50%/300x speedups on OLTP/OLAP github.com
Google Open Knowledge Format — open spec formalizing the LLM-wiki pattern into a portable interoperable format cloud.google.com
npm v12 — blocks lifecycle scripts and remote dependencies by default github.blog
Node.js 26 — 18 security vulnerability fixes x
Blacksmith — one-line YAML swap for GitHub-hosted runners, faster and cheaper CI x
Funding & deals
Industry & policy
Anthropic $16.6M invoice bug — Korean free-plan user billed with no API usage x
Enterprise AI credential-sharing — VentureBeat survey finds 69% of orgs still share credentials across AI agents venturebeat.com
Claude Code moral-character logs — thetreygoff flags user-invisible telemetry the model keeps x
Continuing threads
Anthropic profitability + $50B ARR (cont. from 07-09 / 07-10) — Dylan Patel's podcast reports April and May free-cash-flow positive at 70%+ gross margins, with Anthropic annualized past $50B x
Bun rewrite methodology (cont. from 07-09 Notable) — mitchellh calls $165K in API pricing for 11 days of frontier work an incredible deal, and pulls the clean-context takeaway: don't use one long auto-compacting session for both build and verify x
Meta Muse Spark 1.1 pricing (cont. from 07-10 Top dev) — _MaxBlade benchmarks Muse Spark output tokens at 86% cheaper than GPT-5.6 Sol and 91% cheaper than Fable 5 x
Karpathy second-brain / LLM-wikis (cont. from 07-09) — hwchase17 ships LangChain's OpenWiki personal-brain build with Nemotron 3 Ultra + Deep Agents; BraceSproul releases OpenWiki Brains with local config langchain.com youtube.com
Multi-model orchestration (cont. from 07-10 Notable) — Fable-orchestrates-Sol, Sol-in-Claude-Code, Fable+Sol-in-same-loop combos posted across the day; the harness-is-the-runtime frame is where working teams have settled github.com x