Autonomous Research Tools: Market Map Digest
TL;DR
The pricing fog lifted from opposite directions at once: Google published its Deep Research agent as pure pass-through token arithmetic, while Perplexity disclosed its enterprise credit mechanics — 500 vs. 15,000 credits per seat — without ever publishing what a credit costs (gemini-deep-research-agent, perplexity-deep-research-consumer-agent
). Meanwhile, the first independent autonomy scoring of the destination layer landed, and it punctures the vendor narrative: even the leader tops out at 17/36 (market-map-positioning-hey-lefty
). The market is being sorted by two axes — who lets you model your costs, and who lets you verify your outputs — and neither axis rewards raw autonomy claims.
Pricing transparency is bifurcating: Google's public meter vs. Perplexity's meter-without-a-price
The backend layer just got a price floor and the destination layer got a price mystery, in the same week. Google's API pricing page now carries a dedicated "Pricing for agents" section that bills Deep Research at raw token cost — explicitly including intermediate reasoning tokens from agentic loops — plus tool fees like Search Grounding at $14 per 1,000 requests after a free tier (gemini-deep-research-agent). No per-run premium, no credit abstraction, no seat gate. Perplexity, by contrast, published the plumbing before the price:
"Enterprise Pro users receive a monthly allocation of 500 credits, and Enterprise Max users receive 15,000 credits. These credits expire at the end of each monthly cycle and do not roll over." — How credits work for Enterprise organizations — Perplexity Help Center
The 1:30 allowance ratio between Pro and Max seats, the two-pool funding model, and admin-assigned "named usage limits" are all now public — but the guide hedges that "Credit pricing, task credit ranges, and included monthly allowances are subject to change," leaving the dollar conversion unpublished (perplexity-deep-research-consumer-agent). This matters because Google pricing deep research as a commodity primitive at marginal cost compresses the umbrella every independent backend vendor sells under — and because Perplexity is repeating OpenAI's meter-first, price-later sequence, letting enterprises architect against a system they can't yet compare line-by-line (perplexity-deep-research-consumer-agent
).
What to watch: Whether Perplexity publishes the dollar-per-credit rate — the last missing number in the destination layer's pricing stack — and whether Gemini 3.x Flash's scheduled doubling on Jan 1, 2027 ("$0.75 through December 31, 2026. $1.50 starting January 1, 2027") forces a repricing across every research tool built on the market's cheapest substrate (gemini-deep-research-agent).
Gemini 4 Argon: the substrate upgrade is announced, gated, and priced to undercut
The next model underneath Deep Research exists on paper but not yet in the API. Google announced Gemini 4 Argon on Sept 30, 2026, "built to sustain deep reasoning across complex, long-horizon workflows" — with an output limit expanded to an industry-leading 1M tokens, up from 64K, and introductory pricing of $2/$10 per million tokens that undercuts OpenAI's GPT-5.6 Sol ($5/$30) by more than half (gemini-deep-research-agent). But access is staged through security channels first:
"[Argon is] rolling out to a set of trusted cyber defenders through our Fairwind Program." — Introducing Gemini 4 Argon — Google blog, Sept 30, 2026
The benchmark skew is the tell: #1 on the Vals Index of finance/legal/tax/coding agents, top marks on Vals Finance Agent v2 and Harvey's Legal Agent Benchmark, and — most relevant to research tools — Google's claim of being its "most resilient model yet against indirect prompt injections," leading Gray Swan's IPI benchmark (gemini-deep-research-agent). Prompt injection is the attack class that plagues agentic research loops specifically, so Google is pre-selling Argon as the safe substrate for exactly this workload. When Argon hits paid API customers, Deep Research's cost basis and capability ceiling change simultaneously.
What to watch: Argon reaching general availability for paid API customers and Google AI Ultra subscribers — the moment the 1M-token output limit becomes usable in research loops and the intro-price clock starts (gemini-deep-research-agent).
Independent scoring says the autonomy marketing is ahead of the autonomy
The first third-party cross-check on the destination layer's self-descriptions just landed, and it's deflating. Agentic.ai's October 2026 directory scores research tools on a 0–36 "Agenticness" scale, and the best performer — ChatGPT Deep Research at 17/36 — barely clears half the scale; Perplexity Research scores 9, Gemini Deep Research 11, and everything from Consensus down lands in "Reactive Tool" territory (market-map-positioning-hey-lefty). Meanwhile the same guide tells buyers what actually matters:
"Citation quality is non-negotiable… If you can't verify an output, it's creative writing, not research." — 11 Best AI Research & Analysis Agents in 2026 — Agentic.ai
The gap between marketed autonomy (always-on agents, automations, computers) and scored autonomy (guided assistants) is itself a positioning opening — a tool that scores honestly high on verifiability and process legibility can win buyers the directory is explicitly coaching to demand show-your-work (market-map-positioning-hey-lefty). Note also that the list is destination-layer only: Exa, Parallel, and the backend vendors are absent from the buyer-facing view, meaning procurement still happens at the destination surface — where an orchestrator lives.
What to watch: Whether any vendor markets auditability as the headline feature rather than a compliance footnote — the first mover there claims the wedge the buyer guidance is already writing copy for (market-map-positioning-hey-lefty).
What surprised us
- Google is the transparency leader in a market where opacity was the norm. Against OpenAI's unpublished dots allowances and Perplexity's credit pools, Google's agent meter is "public arithmetic: token rates × agentic-loop consumption + tool fees" (gemini-deep-research-agent
). For anyone building on backends, Gemini Deep Research is suddenly the most cost-modelable source on the market — an odd edge for the incumbent to hand over.
- "Introductory" agent pricing is a timed discount, not a floor. Gemini 3.8 Flash is listed at $0.75 input through Dec 31, 2026, doubling to $1.50 on Jan 1, 2027 (gemini-deep-research-agent
- The buyer's non-negotiable is verifiability, not autonomy — despite a year of autonomy marketing. The Agenticness scores cap the leaders at "Adaptive Collaborator" while the same guide insists on citation quality and synthesis over raw independence (market-map-positioning-hey-lefty
). The market is rewarding honesty about what the tool does, which is a rare gift for a new entrant.
Open threads worth a vote
- [watch] Gemini 4 Argon reaches general availability (paid API + Google AI Ultra) — the substrate-change trigger for Deep Research; steer follow-up on the topic page.
- [watch] Gemini 3.x Flash list prices double on Jan 1, 2027 — the scheduled cost-basis shock for the Flash-built tool layer; steer follow-up on the topic page.