Thursday 7 May 2026
Both Anthropic and OpenAI announced services companies this week, signaling the AI labs’ pivot from model builders to enterprise deployment partners. Meanwhile, OpenAI shipped GPT-5.5 Instant as ChatGPT’s new default and a Miami startup claims a 12-million-token context window.
🤖 Models and Launches
GPT-5.5 Instant becomes ChatGPT’s new default — OpenAI rolled out GPT-5.5 Instant with improved factuality, reduced hallucinations, image understanding, and stronger personalization via saved memories, past chats, and connected Gmail.
Subquadratic debuts a 12-million-token context window — The Miami startup claims ~1,000x efficiency gains over standard attention, with compute scaling linearly rather than quadratically. A 50M-token window is planned. Researchers are demanding independent verification.
Google prepares upgrades for Gemini Flash — A candidate on LM Arena is performing competitively against Gemini 3.1 Pro, with transitions from Flash 2 to 3/3.1-Lite imminent. Signs also point to Flash 3.2.
Gemma 4 gets 3x faster inference via multi-token prediction drafters — Speculative decoding delivers up to 3x speedup without quality degradation by predicting multiple future tokens in parallel.
AI2 releases MolmoAct 2 — An upgraded action reasoning model that improves real-world robot task performance, paired with a large open bimanual manipulation dataset.
🏢 Industry
Anthropic + Blackstone JV ($1.5B) and OpenAI’s Deployment Company ($4B) — Anthropic’s unnamed JV with Blackstone, Hellman & Friedman, and Goldman Sachs ($300M each) builds Claude-powered systems for enterprises. OpenAI’s Deployment Company, backed by TPG, Brookfield, and Bain Capital, raised $4B at a $10B premoney valuation. The AI labs are becoming services companies.
Anthropic committed to spending $200B on Google Cloud over 5 years — Alphabet stock jumped on the report. Google plans to invest up to $40B in Anthropic in return. The symbiosis deepens as Anthropic hits compute constraints.
Anthropic launches higher usage limits and a compute deal with SpaceX — The company is expanding capacity across multiple fronts, including a new compute partnership.
Coinbase lays off 14% of staff — CEO Brian Armstrong cited quarterly volatility and a down market, positioning the cuts as a way to emerge “leaner, faster, and more efficient.”
WiseTech staff in limbo as AI touted as better than humans — The software company announced 2,000 job cuts in February but workers are still waiting to learn which roles will go, even as the company promotes its AI capabilities.
Apple to let users choose rival AI models in iOS 27 — Users will be able to select third-party AI providers (Google, Anthropic, etc.) to power Siri and writing tools across iOS 27, iPadOS 27, and macOS 27 this fall.
🛠️ Agents and Tools
Agents can now create Cloudflare accounts, buy domains, and deploy — Cloudflare enabled autonomous agent workflows for account creation, domain purchasing, and deployment, generating significant HN discussion (350 comments).
OpenAI ships Agents SDK for TypeScript — Includes sandbox agents and an open-source harness. Codex also got Auto Review for lower-friction approvals and task progress UI.
Computer use is 45x more expensive than structured APIs — Vision agents require detailed prompts and thousands of input tokens per screenshot. The cost gap makes building MCP/REST surfaces worthwhile despite higher upfront effort.
Agents for financial services — 10 ready-to-run Anthropic templates — Finance is Anthropic’s second-highest revenue segment. Templates cover pitchbooks, KYC screening, and month-end close.
Cursor agents that monitor GitHub and auto-fix CI failures — The CI monitoring agent launches as coding automation extends into CI/security workflows.
🔬 Research and Analysis
Google rethinks hallucinations through uncertainty — The paper reframes hallucinations as failures to express uncertainty rather than gaps in knowledge, proposing “faithful uncertainty” for aligning confidence with reliability.
In search of wasted bits: how much information do LLM weights carry? — Deep analysis of the memory-vs-compute tradeoff in LLM inference. Quantization transfers less data but doubles computation — the real bottleneck is memory bandwidth.
ProgramBench: top accuracy is 0% on whole-repo generation — Meta’s 200-task benchmark asks models to generate substantial software (SQLite, FFmpeg, a PHP compiler) from scratch. Models can pass >50% of individual tests but none complete full tasks.
Harness quality is the real differentiator in coding agents — Model-harness-task fit matters more than abstract benchmarks. Base models without productized harnesses expose how much agents depend on instructions, tools, and measurement loops.
📊 Culture and Trends
Vibe coding and agentic engineering are getting closer than I’d like — Simon Willison on the uncomfortable convergence. 287 HN comments.
Claude Code is not making your product better — At the frontier, spending on tokens may not produce economic value. The taste to delete, compress, and refuse is more valuable than ever.
Cognitive surrender vs cognitive offloading — Addy Osmani on the line between delegating to AI but still owning the answer vs. when AI output quietly becomes your output.
How to work and compound with AI — Eugene Yan’s practical guide: provide good context, encode taste as config, make verification easy, and let every artifact become context for the next session.
🔥 Hacker News
Agents can now create Cloudflare accounts, buy domains, and deploy (350 comments) — Cloudflare’s autonomous agent deployment story dominated HN.
StarFighter 16-Inch (348 comments) — New 16-inch laptop from StarLabs generating significant buzz.
Vibe coding and agentic engineering are getting closer than I’d like (287 comments) — Willison’s take on the convergence of vibing and engineering.
Valve releases Steam Controller CAD files under Creative Commons (278 comments) — Valve opens up its hardware designs.
Higher usage limits for Claude and a compute deal with SpaceX (225 comments) — Anthropic expanding capacity with SpaceX compute partnership.
245TB Micron 6600 ION Data Center SSD now shipping (178 comments) — A quarter-petabyte in a single SSD form factor.
🌍 World and Australia
US fires on Iranian-flagged oil tanker as Trump gives Tehran fresh ultimatum — Iran’s top negotiator says the US aims to force Tehran’s “surrender.” Israel also struck Beirut for the first time since the ceasefire began.
Israeli intelligence officers charged with bribery over Polymarket bets on Iran strikes — An air force major and accomplice allegedly used classified information to bet on the timing of military operations.
Ted Turner dies — The CNN founder and media mogul has died.
Coalition’s roadmap to slash immigration leaked — The opposition’s immigration policy document surfaces ahead of budget season.
TikTok’s algorithm favored Republican content in 2024 elections — For You pages prioritized pro-Republican content in three states studied, researchers say.
Single dose of psilocybin can cause anatomical brain changes — Participants took 25mg of psilocybin and reported deeper psychological insight and better wellbeing a month later.
Sources: TLDR (General + AI), AINews/Latent Space, The Guardian, Hacker News — 3 newsletters processed, ~30 stories distilled