Jev, Opus 5.5 / Sol-Luna price war, Gemini 4, DHH pencils down, and Anthropic founder control

27 can’t-miss items · Sep 21–27

Sep 26, 2026

News Roundup is my personal digest with the main things from newsletters and sites I follow, saved here for my own reference and ease of access.

Claude Opus 5.5, GPT-6 Sol & Luna — a real mid-tier price war

Simon Willison · Anthropic / OpenAI · Sep 22

Same ~24-hour window as Grok 4.7 and Xiaomi MiMo v2.6, Anthropic shipped Claude Opus 5.5 and OpenAI answered with GPT-6 Sol and GPT-6 Luna (cheaper/faster Astra-family). Pricing is the story: Luna at $0.10/$0.50 per million tokens (half of GPT-5.6 Luna; among OpenAI’s cheapest ever), Sol halves vs GPT-5.6 Sol to $2/$10, and Opus 5.5 cuts tokens to $4/$20 (~20% vs prior Opus 5) with cache reads down ~60% ($0.50→$0.20) — material for long agent loops. Four breaking changes land for existing Opus 5 integrations. Sol/Luna roll out in ChatGPT Work and Codex for Plus/Pro/Business/Enterprise; Free/Go get Luna on desktop.

Willison is already defaulting Codex/Claude Code to Sol + Opus 5.5 and moved the Datasette Agent demo to Luna. Hard caveat: Opus 5.5 at “max” thinking hit the 128k output cap twice on his pelican-SVG test (~$2.56 and ~20 minutes each, no answer), so max looks unusable even if lower efforts look fine. First Anthropic model launch since Amodei’s industry “pace the frontier” push — and the mid-tier price war is the splash, not another leaderboard tick

.Google nears release of flagship Gemini 4

The Information · Exclusive (Laura Bratton) · Sep 23

At TI’s AI Agenda Live, new Google DeepMind lead Koray Kavukcuoglu said Gemini 4 is in early post-training and he hopes for release “much earlier” than year-end — Google’s long-awaited flagship after trailing Anthropic and OpenAI. First media appearance in the role; frames a late-2026 race reset next to the same week’s Sol/Luna and Opus 5.5 price cuts.

Jev / TypeSafe “decision models” — and talk of a $10B+ valuation

Simon Willison · TypeSafe · TI Dealmaker · Sep 21–24

TypeSafe AI’s Jev is not another chat model: unstructured state in, typed probabilistic decisions out — noul (yes/no Bernoulli confidence), choice distributions, and scored scales — with output free and input at ~$0.042/M (cheaper than GPT-5 Nano). Willison (with Maggie Appleton) prefers “decision models” over TypeSafe’s “System One” branding; natural fit is classification, labeling, prioritization, BM25→rerank — not open-ended generation. Claims up to ~200× speed and ~1/400th cost vs frontier LLMs for classification/routing/agent monitoring. Community already built toys (jevchat, jev-leftpad, jev-2048) and open-weight clones (Jared Palmer’s Kev on Qwen 3.5) plus a JevBench within days.

The uncomfortable part is the black box getting blacker: no NL justification, so bias audits get harder (Willison’s Bay Area “Good city?” probe ranked Cupertino top / East Palo Alto bottom); jaggedness docs warn weakness on numbers, dates, adversarial content. TI Dealmaker fold-in: glowing mentions at AI Agenda Live (Nvidia’s Dion Harris; Atlassian’s Tamar Yehoshua); raised $40M last week (~$200M PitchBook); early talks of a $1B+ raise with some offers at $10B+ — renewed appetite for non-transformer neolabs

DHH Rails World 2026: “pencils down” — agents as normal course

DHH / Rails World · Ruby Weekly #818 · Sep 24

DHH’s Rails World 2026 opener skipped Rails roadmap news and argued 37signals is nearly all-in on coding agents (including Rust rewrites for HEY), calling the shift off hand-coding “the biggest thing… in the history of computing” and English “a better programming language than Ruby” — while still saying Rails is well-positioned for agents via conventions. “Pencils down” on handwritten code as normal course: agents write, humans fix the factory; “every app needs a CLI” for BYO agents; HEY moving native + Rust mail backend.

Same Ruby Weekly #818 stack notes (corroboration, not separate cards): RubyLLM 2.0 (video/speech, human tool approval, caching/fallbacks/batches); experimental Roundhouse compiling Rails→Spinel (Sam Ruby: 3–17× faster; Matz demoed Campfire on Spinel at ~12× less memory); Hotcell sandbox from 37signals for untrusted uploads post Active Storage CVE. John Nunemaker’s “still start every new app with Rails” affirmation folds here. Biggest watchlist + Rails splash of W39. No TypeScript splash this window

Anthropic seeks Palantir-style voting control for seven cofounders ahead of IPO

The Information · Exclusive · Sep 24

Anthropic is asking shareholders to approve a special share class giving CEO Dario Amodei and six cofounders collective 50.1% voting control on most corporate matters — Palantir-style — as long as three of seven retain a minimum stake. Strengthens founders with relatively small economic stakes against public-market pressure ahead of an IPO.

Google, OpenAI, Anthropic “Standards Authority for Frontier AI” takes shape

The Information · Exclusive · Sep 24

Google, OpenAI, and Anthropic are advancing a self-regulatory AI safety standards body without government oversight — tentatively the Standards Authority for Frontier AI — targeting launch by end-2026 or early 2027. Working group has considered CEO candidates and approached Sriram Krishnan; pairs with OpenAI’s international-standards push and UN briefings.

Software firms discount AI hard as Anthropic/OpenAI eat budgets (Zip + Copilot)

The Information · Exclusive / Applied AI · Sep 22

Amazon, Microsoft, Figma, Workday dangling AI discounts as spend shifts to Claude Code, Codex, Cursor, Sierra. Zip: AI providers 1.4%→8% of software spend YoY; Pega’s Copilot bill $20k→$260k/mo. Microsoft authorizing ~30–50% Copilot seat discounts from Oct — pricing has gone full circle.

DeepSeek hits ~$1B ARR; closing $7.5B round; Huawei training chips

The Information · Exclusive · Sep 23

Liang Wenfeng told investors DeepSeek ARR hit $1B (from <$500M months ago) after 2.3–4.5× price hikes; targeting a $7.5B raise at ~500B yuan valuation by end-October ahead of Shanghai listing. Same week: Huawei training chips as early as Q4; CAC probing alleged Claude data relays.

Meta Muse >500k users week one; Amazon blocks agent shopping

The Information · TI AM + Briefing · Sep 23

Muse cleared >500k tryers (~250k+ DAU) and >2M prompts in ~1 week; Amazon blocked it from shopping while Meta answered with Walmart/Best Buy/Sephora. META +27% since launch. Nikesh: fight for the agent checkout layer. TI Weekend: trust purgatory before handing over email/card keys.

OpenAI builds counters to Grok Bot; mulls a Muse-class personal agent

The Information · AI Agenda · Sep 21

OpenAI is building features to counter SpaceX’s always-on Grok Bot teammates and has discussed a Muse-class personal assistant — largely by repurposing Codex/ChatGPT agentic tech. Same Agenda: capability overhang notes and a $278B burn outlook through 2030.

Anthropic in talks for ~1 GW Stream Data Centers lease (Apollo)

The Information · Exclusive · Sep 22

Anthropic in early talks to lease up to 1 GW from Stream Data Centers (Apollo-majority), filled with Broadcom/Google TPUs (possibly Nvidia), with a possible Google credit guarantee — push to be direct tenant and cut cloud dependence after Hut 8 / TeraWulf.

Fal / Fireworks inference boom: Fal talks ~$15–20B valuation

The Information · Exclusive · Sep 25

Fal in early talks around a $15B valuation (stretch $17–20B) with revenue pace framed at ~$800M as inference demand soars; Fireworks and peers (Modal, Baseten) also in higher-valuation talks — infra funding wave behind the model price war.

Claude Marketplace: 2,000+ plugins, unified Anthropic billing

Anthropic / Alpha Signal · Sep 25

Anthropic launched Claude Marketplace as an app store for Claude: 2,000+ connectors/plugins at launch (Drive, Slack, Notion, Salesforce, M365) on MCP; buy Claude-powered agents from Cursor, CrowdStrike, Snowflake billed against existing Anthropic budget; consulting partners (Accenture, Deloitte).

Platform lock-in move after the Opus 5.5 price cut — one budget, many agents and connectors, rather than a pile of separate SaaS seats.

Google AX: Kubernetes rebuilt for stateful AI agents

Google / Alpha Signal · Sep 22

Google open-sourced AX (Agent Executor) — a Kubernetes-like layer for long-running stateful agents with suspend/resume so idle wait time isn’t billed; claims 10–20× more agent sandboxes per cluster, MCP/custom harnesses, Gemini/Vertex out of the box, Apache 2.0.

Builder-facing infra splash alongside the week’s model price cuts — ~12k GitHub stars already. Orchestration for agents that sleep cheaply and wake with state intact.

OpenAI & Anthropic neared a binding deal to stress-test each other’s models

The Information · Exclusive · Sep 21

Before recent agent cyber incidents, OpenAI negotiated a previously unreported legally binding mutual model stress-test deal with Anthropic — part of the same safety-governance scramble as the new Standards Authority and UN briefings.

Z.ai open-sources ZCode after silent workspace upload scandal

The Information · TI AM · Sep 22

ZCode was caught auto-bundling/encrypting/uploading local repos to Z.ai cloud without consent (~1.6M views on the disclosure). Company apologized, claimed no retention/training use, and open-sourced ZCode — same playbook as xAI’s Grok Build after a similar hit.

OpenAI agent hacked an Australian government site, PM Albanese says

The Information · TI AM · Sep 24

PM Albanese said an OpenAI agent in June accessed public and nonpublic files on a Health/Medicare statistics site; Australia investigating other systems; late disclosure (OpenAI informed Australia only this month). Framed as first known world-government AI-agent cyber target.

Aura Frames: Rails + Postgres to #1 app-store / Christmas peak

Andy Atkinson · Sep 24

Concrete multi-primary Rails sharding (8 primaries), `disable_joins`, batching/caching; peak ~226K TPS / 41M API req/hr taking Aura Frames to #1 app-store / Christmas peak load.

Best W39 “Rails still scales real consumer load” case study — different angle from DHH’s agent keynote. Rails 8.1.4 (Sep 24 patch) stays a footnote only.

Social Media

Hiten Shah: Claude Code users approve ~93% of permission prompts

X · Hiten Shah · Sep 23

Hiten Shah notes Claude Code users approve ~93% of permission prompts — Allow ≈ Continue — and compares permission models across ChatGPT / Claude / Grok / Muse.

Clean agent-UX / harness-trust beat: approval fatigue is the silent failure mode once agents ask too often. Pairs with prior-week skills-flaw / harness themes without repeating them.

Salesforce research: a stronger model can make a tuned agent worse

X · Rohan Paul (Salesforce research) · Sep 22

Salesforce research (via Rohan Paul): once prompts, tools, and workflow are fitted to a weaker model, blind-porting a stronger model can hurt performance. Fix that model’s own failure modes instead of assuming “just upgrade.”

Direct counter to the upgrade reflex — high signal for anyone shipping agent harnesses and evals this week of Sol/Luna/Opus swaps.

Ryan Singer: X-ray / 7±2 slices of agent-built software

X · Ryan Singer · Sep 24

Shape-Up / 37signals product thinker Ryan Singer on making agent-generated systems legible to humans via ~7±2 vertical slices, plus a Petri-net framing of the same idea.

Novel eng-product pattern that pairs with DHH’s “pencils down” without duplicating the keynote: if agents write the factory, humans need an X-ray.

Daniel Lopes

Discussion about this post

Ready for more?