Morning Brief: Wednesday, September 23

One hundred seventeen feeds. Two weeks. 5,035 items reduced to what follows. (what we track, how we crawl, subscribe)

Tuesday's model-launch pileup didn't resolve into a winner, it doubled: Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Sol/Luna both shipped within hours of each other, both cutting prices, and Simon Willison's same-day roundup treats the coincidence itself as the story. Meta's Muse, still riding Monday's download numbers, now has to answer for TechCrunch reporting that its resemblance to OpenClaw "isn't a coincidence."

A quieter thread carries more weight by volume: agent memory is the single most common subject in today's arXiv cs.AI batch, with half a dozen independent papers (DolphinBench, MemCalib, RPMem, SAGE, AkasicMEM, TARL) converging on the same open problem the same week InfoWorld ran a practitioner piece titled, plainly, "Fixing agent memory."

Top (5-7 min)

Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war: Simon Willison, 2026-09-22. Same-day synthesis of both labs' releases and the price cuts that came with them.
Anthropic releases Opus 5.5 with lower prices and Fable-level performance: TechCrunch, 2026-09-22. Anthropic's release, benchmarked against its own Fable line.
OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes: TechCrunch, 2026-09-22. OpenAI's same-day answer, pitched on cost and error rate rather than raw capability.
JetBrains goes big on agentic development — and bets the IDE still matters: The New Stack, 2026-09-22. JetBrains frames its agent push as a 26-year-history-defining bet that the IDE, not the chat window, remains the unit of work.
Meta admits Muse's likeness to OpenClaw isn't a coincidence: TechCrunch, 2026-09-22. Follows Monday's download and Amazon-block stories with an admission of direct imitation.
Pentagon Investigators Say Overreliance on Palantir AI Contributed to US Strike That Killed 123 Iranian Children: Slashdot, 2026-09-22. A single-sourced but serious claim about an AI-targeting failure with a fatal outcome.
[AINews] Claude Opus 5.5, the new default model for AINews — and everybody cuts prices 40-50%: Latent Space, 2026-09-23. Next-day wrap treating Opus 5.5's price cut as an industry-wide signal, not a one-off.

Themes this week

Frontier-lab launches now arrive in pairs, not sequence: Simon Willison: Opus 5.5, Sol, Luna, and a new price war (Tue), Vercel: Claude Opus 5.5 now on AI Gateway (Tue), Vercel: GPT-6 Sol and Luna now on AI Gateway (Tue), The New Stack: Anthropic cuts Opus 5.5 pricing by 20%, agent calls may route to an older model (Tue).
Agent memory becomes the week's dominant arXiv thread: arXiv: DolphinBench, mapping the Pareto frontier of agent memory (Wed), arXiv: MemCalib, benchmarking memory use in LLM agents (Wed), arXiv: RPMem, long-term recurrent parametric memory across sessions (Wed), arXiv: AkasicMEM, governed enterprise memory for agents (Wed), InfoWorld: Fixing agent memory (Tue).
Jev shifts from viral clones to formal evaluation: HN: Jev in 25 Lines of Python (Wed), arXiv: JEV-as-a-Judge, accept when confident, escalate when unsure (Wed), arXiv: REFLEX with Jev for efficient selective control in LLM agents (Wed), HN: Show HN, JevBench, a reproducible benchmark for typed decision models (Tue).
Meta's Muse compounds adoption with an imitation admission: TechCrunch: Meta admits Muse's likeness to OpenClaw isn't a coincidence (Tue), 404 Media: Meta tests Muse AI agent calls that are actually made by humans in a call center (Tue).

Scan (15 min)

Wednesday feeds
- Visual Studio Code 1.138 brings agent sessions to Dev Containers, InfoWorld, 09-23
- Software dependencies are running away from us, InfoWorld, 09-23
- Microsoft Helps Take Down Massive Automated, AI-Powered Phishing-as-a-Service Platform, Slashdot, 09-23
- Trump Denounces Attempts to Control AI, Wants It Renamed 'Super Intelligence' in US Documents, Slashdot, 09-23
- SF October 14th: A Birds of a Feather Session on Agentic Engineering, Simon Willison, 09-23
- Sandboxing with minimal effort, Lobsters, 09-23
- Smart glasses are already causing havoc in India, MIT Technology Review, 09-23
Tuesday carryover
- The agent coordination protocol hiding in plain sight: GitHub issues, InfoWorld, 09-22
- AWS launches CloudWatch Omni to unify observability for AI agents and applications, InfoWorld, 09-22
- Z.ai disables coding assistant feature after flaw exposed enterprise code upload risk, InfoWorld, 09-22
- The Genie One MCP is now Generally Available, Databricks, 09-22
- Hacking group ShinyHunters claims it breached the FBI, stole agents' and applicants' data, TechCrunch, 09-22
- GitHub Actions leaking secrets when Miri output is cached, Lobsters, 09-22
- GPT-6 Astra Breaks an Old Enigma Message, Schneier on Security, 09-22

Tail

The Palantir/Pentagon strike report is single-sourced so far: Only Slashdot (via Bloomberg) has surfaced in the warehouse; no independent corroboration, Pentagon statement, or Palantir response has appeared yet for a claim this serious.
Z.ai's coding-assistant flaw has no vendor postmortem yet: InfoWorld reported the feature was disabled after an enterprise code-upload exposure risk; no technical writeup of the flaw itself is in the warehouse.
GPT-6 Astra's Enigma break is notable but unreviewed: Schneier linked to the cryptocellar writeup same-day; no independent cryptographic assessment of the claim has surfaced yet.

Feed silences (>72h since last item)

Sources that publish frequently but have gone quiet:

Neel Nanda (400 days) — last item 2025-08-19.
Brendan Gregg (228 days) — last item 2026-02-07.
Spritely Institute (133 days) — last item 2026-05-13.
Andy Wingo (130 days) — last item 2026-05-16.
Aphyr/Jepsen (103 days) — last item 2026-06-12.
Typst (100 days) — last item 2026-06-15.
Eugene Yan (94 days) — last item 2026-06-21.
Lilian Weng (81 days) — last item 2026-07-04.
Andrej Bauer (74 days) — last item 2026-07-11.
Julia Evans (64 days) — last item 2026-07-21.
Stephen Wolfram (64 days) — last item 2026-07-21.
Marc Brooker (56 days) — last item 2026-07-29.
Antithesis (36 days) — last item 2026-08-18.
Bunnie Studios (24 days) — last item 2026-08-30.
Vicki Boykis (22 days) — last item 2026-09-01.
deepmind-blog (22 days) — last item 2026-09-01.
GitHub Engineering (21 days) — last item 2026-09-02.
Fly.io (20 days) — last item 2026-09-03.

Build provenance