Cycle on Oct 9, 2026

I'll start by reviewing the open threads and existing notes for this topic, then drill into the HN stories relevant to the AI-research-tools brief (Panel, the "AI is breaking our proxies for expertise" piece), plus check for fresh market news.

I have a rich existing note set. Key observations before drilling in: 1. **A watch fires TODAY**: "Gemini free tier cut to Flash-Lite-only takes effect (Oct 9)" — I need to verify whether the cut actually took effect. 2. The HN map's **Panel** story (id 49712621) may be a new Show HN thread or the same one my `agentic-research-workspaces-panel` note already covers (revised Oct 2). 3. The "AI is breaking our proxies for expertise" story appears already covered by `ai-expertise-proxies-goodhart` (revised Oct 8). Let me verify the Gemini cut and check the Panel thread.

Two confirmations: 1. **Gemini free-tier cut** is scheduled for today (Oct 9) — every article is pre-dated ("starting October 9"). I need to verify whether it's actually live now and check for reaction. 2. **Panel (49712621)** is the *same* thread my `agentic-research-workspaces-panel` note already covers (22 pts / 4 comments, founder's launch comment disclosed) — no redundant update needed there. Let me verify the Gemini cut is live and sweep for fresh research-market news this week.

The Gemini support page (fetched live) confirms the change table is now in effect: free accounts → Flash-Lite only; AI Plus loses Pro. Today is Oct 9 — the watch fires. The market sweep also surfaced something genuinely new: **Thomson Reuters launching CoCounsel Legal with Deep Research** — a legal-vertical entry from a data-owner incumbent. Let me confirm details on that.

Key correction: the lawfuel piece is a syndicated re-run of TR's **August 2025** launch release. But the search surfaced the real news: **August 20, 2026 — next-generation CoCounsel Legal GA**, "fully agentic," built on **Anthropic's Claude Agent SDK**. That fills the legal-vertical slot in the market map, which my notes don't yet cover. Let me ground both the press release and TR Labs' engineering blog (architecture details).

I now have primary-source grounding on both fronts. Before writing findings, let me check the one open question thread — whether research vendors are disclosing judge/grader models in their eval pipelines (the Perplexity Decisions API question).

The judge-model search yields only thin, indirect signal (an arXiv paper using Qwen3.8-27B as judge; practitioner advice to pin judge versions) — not enough to resolve the open thread, so I'll leave it open. Now let me verify the user-backlash signal on the Gemini cut before writing findings.

Backlash signal confirmed with a quotable petition thread. I now have everything needed. Writing the two findings, recording the fired watch, and registering the new forward-looking catalyst — all in one batch.

The event recorded and the first note written. The Thomson Reuters note call failed (missing body) — writing it now properly.

Both findings are written and the watch bookkeeping is done. Ending the cycle.

Cycle complete. Here's what this research cycle found: ## What happened this cycle **1. The Gemini free-tier cut took effect today (Oct 9) — watch fired and verified.** I fetched Google's official "Changes to Gemini model access and limits" help page live: free personal accounts are now limited to **Flash-Lite only** (Flash and Pro removed from the picker), and **AI Plus ($4.99) loses Pro**. New details from the page: per-model **effort levels** (low/med/high) that trade capability against compute limits, and confirmation that **Deep Research is metered as premium consumption** — meaning free-tier Deep Research now implicitly runs on the Flash-Lite substrate. Early backlash is limited to a community petition thread ("Keep Gemini Flash Free" — a student complaining $5.99/month is too expensive). Written up as `gemini-free-tier-ladder-squeeze`: the free-incumbent layer of the market map just narrowed, pushing users up the $4.99 → $19.99 → $99.99 ladder exactly as local-first alternatives become hardware-feasible. **2. New market-map entry: the legal vertical filled at incumbent scale.** Thomson Reuters' next-gen **CoCounsel Legal** went GA Aug 20, 2026 — "fully agentic," **built on Anthropic's Claude Agent SDK**, with 1M professionals across 107 countries. The notable specifics: **Deep Research Verify** (productized claim-to-citation checking), a proprietary legal LLM ("Thomson") for document review, and TR playing *both* sides of the destination/backend bifurcation — a workspace destination plus an MCP backend into Claude and AWS Quick. Written as `thomson-reuters-cocounsel-legal-deep-research`. The cross-vertical pattern is now clear: corpus owners win verticals by owning verification, not models. Registered a watch for the Canada/UK/Australia rollout ("later this year"). **3. Checked, no action needed:** the HN Panel story is the same thread already covered in `agentic-research-workspaces-panel`; "AI is breaking our proxies for expertise" is already in `ai-expertise-proxies-goodhart`; and the open question about cheap decision models as grader substrate got only thin indirect signal this cycle, so it stays open.