Autonomous Research Tools: Market Map Digest — The Free Rung Gets Priced Up, and the Vertical Layer Fills at Incumbent Scale
TL;DR
Two moves define this cycle. Google's free-tier cut took effect on schedule October 9, converting Gemini's free tier into a demo for the $4.99 → $19.99 → $99.99 ladder and quietly demoting free Deep Research to a Flash-Lite substrate (gemini-free-tier-ladder-squeeze). Meanwhile the legal vertical's data-owner incumbent — Thomson Reuters — shipped fully agentic CoCounsel Legal on Anthropic's Claude Agent SDK, with a productized verification layer and a million-seat installed base (thomson-reuters-cocounsel-legal-deep-research
). Together they confirm the bifurcation: hosted general-purpose research gets more expensive at the bottom, while vertical research gets won by whoever owns the corpus and the citations.
The free-incumbent ladder is being pulled up on purpose
The general-purpose free tier is no longer a growth strategy — it's a pricing ladder. Google's official help page, "Changes to Gemini model access and limits," now live, confirms the post-cut state: free accounts get Flash-Lite only, and the $4.99 AI Plus tier loses Pro, the flagship model (gemini-free-tier-ladder-squeeze). The metering granularity is increasing alongside the gating:
"For each available model, you can also select the model's effort level from low, medium, and high. This will increase the model's ability to complete tasks... but will also use more of your limit." — Changes to Gemini model access and limits — Gemini Apps Help
Deep Research is explicitly listed among premium features that "may cause you to reach your limit faster" — and with Flash removed from the free picker, free-tier Deep Research runs implicitly on Flash-Lite under standard limits (gemini-free-tier-ladder-squeeze). The timing is the tell: the cut took effect October 9, a week before the Surface Laptop Ultra's October 16 ship date puts serious local inference in consumer hands — the hosted free rungs are being priced up exactly as local-first alternatives become hardware-feasible. For the market map, this widens the wedge between "free but degraded" and "local and yours," which is precisely the seam a player like Hey, Lefty should occupy.
What to watch: Whether the "Keep Gemini Flash Free" petition gains traction or the backlash stays at student grumbling — and whether any Q2 deep-research startup publicly positions against the downgrade.
The vertical layer fills — and it's the corpus owner, not the startup, who wins it
The legal slot in the market map, previously occupied by startup copilots, is now filled at incumbent scale. Thomson Reuters announced general availability of the next-generation CoCounsel Legal on August 20 — "a fully agentic AI experience" engineered to "reason, plan, and execute at the level of a senior associate" (thomson-reuters-cocounsel-legal-deep-research). The architecture disclosure is unusually candid about the model strategy:
"Deep Research is model agnostic. We work with multiple frontier model providers and open-source options so we can select the best model for each workflow. At launch, a core part of the system uses Claude 4 for its tool orchestration." — Deep Research in Westlaw and CoCounsel: Building Agents That Research Like Lawyers — Thomson Reuters Labs
The pattern is exactly what the health-vertical watch predicted, arriving from the data owner rather than a startup: TR owns the corpus (Westlaw, Practical Law), the verification infrastructure (KeyCite), and the relationships (one million professionals across 107 countries) — and rents frontier models for orchestration (thomson-reuters-cocounsel-legal-deep-research). Most instructive for the map is Deep Research Verify, a named feature that "checks whether cited Westlaw and Practical Law authority supports specific legal assertions" — the audit-grade "show your work" motion from the trust thesis, productized. TR is also playing both sides of the destination-vs-backend frame, shipping an MCP backend that lets lawyers call CoCounsel from Claude directly. Pricing remains undisclosed and sales-led, the mirror image of the consumer ladder above.
What to watch: The thread on next-gen CoCounsel's international GA — Canada, UK, and Australia are slated for "later this year," and whether TR publishes seat pricing at that point will tell us how opaque the vertical layer intends to stay (thomson-reuters-cocounsel-legal-deep-research).
What surprised us
- The incumbent built the verification product, not just the claim. Everyone says "grounded in citations"; TR shipped a discrete verification pass — Deep Research Verify — as a sellable feature (thomson-reuters-cocounsel-legal-deep-research
). That's the strongest signal yet that provenance is a product surface, not a footnote, and it's a feature Hey, Lefty could generalize beyond legal.
- The squeeze is coordinated with hardware, not against it. Google's free-tier cut lands a week before Microsoft's local-inference hardware ships (gemini-free-tier-ladder-squeeze
) — whether or not anyone is coordinating, the hosted-free window and the local-feasible window are closing and opening at the same moment.
- TR's agent takes ~10 minutes per comprehensive analysis and is fine with that. The Labs post describes runs that take "about 10 minutes to complete and verify" with evaluator models calibrated against human judgment (thomson-reuters-cocounsel-legal-deep-research
- Model-agnosticism keeps showing up on the incumbent side. TR explicitly touts multi-provider orchestration on Claude's own agent SDK (thomson-reuters-cocounsel-legal-deep-research
Open threads worth a vote
- [watch] CoCounsel Legal next-gen international GA (Canada/UK/Australia "later this year") — steer follow-up priority and whether we treat published seat pricing as the trigger signal.