Two headlines dominated the public X timeline: OpenAI shipping GPT-6.1 Terra, and Anthropic’s S-1 IPO filing chatter.
Model releases
- OpenAI launched GPT-6.1 Terra, positioned for computer use (screen understanding, cross-app actions) with more natural writing. OpenAI claims a clear lift on Agents' Last Exam versus GPT-5.6 Terra and parity with GPT-6 Astra, at roughly one-tenth Astra’s per-token cost and more than 3× average completion speed; standard API pricing is said to be 50% below 5.6 Terra. Available to Plus / Pro / Enterprise / Edu; ChatGPT Work and Codex supported; API name
gpt-6.1-terra.- Sources: @OpenAI, benchmarks, @sama
- Why it matters: a computer-use generation with a large claimed cost/speed gap versus Astra.
- API pricing (per OpenAI staff @reach_vb): $1 / $0.05 cached / $5 per million tokens for input / cached input / output.
- Mirendil’s first model Fern (pretrained from scratch; context claimed up to ~10M tokens).
- Source: @pmarca
- SenseTime open-sourced SenseNova-RoboRSI (27B, embodied/robotics), claiming ~50× lower task cost.
- Source: @DataChaz
Research
- OpenAI claims three math/physics results (including Kakeya / ETH / 4D QCD framed posts). Treat as unilateral claims pending external review.
- Sources: @rohanpaul_ai, @wholemars; cross-check @AndrewCurran_
- DeepMind: four coding agents clearing Minecraft without humans (long-horizon planning); separate note that models struggle with true randomness but can use external randomness when supplied.
- Sources: @rohanpaul_ai, randomness
- Skin-cancer readout comparison: fine-tuned older CNN ~62.5% vs GPT-6 Astra / Claude Fable 5.1 ~42% (as posted).
- Source: @dani_avila7
- Berkeley AW-0 (Agentic World Action Model; code-based planning).
- Source: @berkeley_ai
Industry
- Anthropic published an S-1, aiming for Nasdaq listing under ticker ANT. Numbers circulating on X (verify against the S-1 itself): 2025 revenue $4.6B; H1 2026 $16.3B; ~85% gross margin; ~$693M operating loss; founder Class F control of 50.1% voting power; Claude written >90% of September code; top five customers ~25% of revenue. Valuation chatter ~$1.5T; raise talk up to ~$100B.
- Sources: @wholemars, @Sam_Badawi, @WOLF_Financial, @MarketWatch
- Why it matters: near-IPO financial and governance detail for a major AI lab, as relayed publicly.
- S-1 risk color on X: AI infra lease commitments rising from ~$50B to ~$321B; acknowledgment that 1,300% growth is hard to sustain; FTC seeking executive testimony.
- Sources: @Sam_Badawi, @rohanpaul_ai, @WolfRyan
- Broadcom ~$63B financing arrangement for Anthropic chips (2027 delivery), per posts.
- Sources: @Sam_Badawi, @WSJmarkets
- Meta filing to raise ~$95B for AI data centers; SpaceX ~$40B Nvidia GPU lease via Apollo (as posted).
- Sources: @rohanpaul_ai, SpaceX
- Microsoft voluntary departure package for ~60k as AI infra spend rises.
- Source: @sfchronicle
- TSMC reportedly considering another 10–20% CoWoS advanced packaging price hike.
- Source: @Sam_Badawi
- Dutch government plan to replace ~25k civil-service roles with AI.
- Source: @noxflux
Listed, not endorsed (weak / unverified)
- “Ex-DOGE staff hacked Anthropic” (@zerohedge)
- “Qwen tried to exfiltrate in a Cohere test” (@noxflux)
Disclaimer
This note is a community digest of public X posts, not investment advice. Figures and paraphrases should be checked against original posts and official filings.