Autonomous Research Tools: Market Map Digest — The Ultra Deep Research Extreme
TL;DR
The deep-research market just got its first "Virtual CSO": Sakana AI's Marlin runs autonomous loops for up to eight hours and returns hundred-page cited reports plus executive slides, at the highest per-run price published anywhere in the category (sakana-marlin-ultra-deep-research). That matters for the market map because Marlin is the third vendor to publish a per-run meter — and because its own FAQ explicitly declines the private-data, local-first lane, which is exactly where Hey, Lefty sits. The extreme end of the market is now defined; the gap it leaves open is defined too.
The long-horizon extreme: an 8-hour agent sold as strategy headcount
The top of the deep-research market just split off from the consumer tier entirely. Where OpenAI, Google, and Perplexity return 5–10 page briefings in roughly 3–30 minutes, Sakana's launch post pitches a different unit of work:
"Give it a research topic, and Marlin works autonomously for up to roughly eight hours, crafting a detailed strategy report up to a hundred pages long, along with executive summary slides." — Sakana AI Launches Its First Commercial Product, Sakana Marlin
The buyer is strictly B2B — "corporate strategy and business-planning teams at financial institutions and operating companies, consulting firms, think tanks, and research houses" — with sample deliverables like Strait of Hormuz blockade scenarios and the global AI-regulation patchwork shipped as cited PDF plus PPTX deck (sakana-marlin-ultra-deep-research). This is a deliberate reframe of what "deep research" is sold as: not a chat answer but the output of a strategy team, priced accordingly. A closed beta of roughly 300 professionals ran from April 2026 before launch, so the demand signal isn't hypothetical.
What to watch: Whether other labs follow the "Virtual CSO" framing — deliverable-out, hours-long runs — or whether the consumer tools' 3–30 minute tier absorbs the mid-market first (sakana-marlin-ultra-deep-research).
Pricing: the third published per-run meter, and the most expensive one
Per-run pricing is becoming the legible alternative to credit-pool opacity. Marlin charges ¥98 (~$0.61) per credit at 100 credits per run — about $61 pay-as-you-go — dropping to an effective ~$47/run on the ¥150,000/month Pro plan and ~$42/run on Team, with Enterprise as a custom quote (sakana-marlin-ultra-deep-research). That's the third published per-run meter in the market alongside Exa Agent Ultra's budgeted subagent runs and Parallel's per-request CPM — and the highest of the three, consistent with selling weeks of analyst work rather than minutes of compute (sakana-marlin-ultra-deep-research
).
The contrast with last cycle's picture is stark: Google prices deep research as raw token arithmetic at marginal cost, Perplexity publishes credit mechanics without a dollar conversion, and Sakana publishes a clean per-run price with no abstraction at all. Three vendors, three pricing philosophies — and the buyer-facing segmentation now has a genuine enterprise band with a real anchor price in it.
What to watch: Whether Marlin's effective $42–$61 per-run band becomes the reference point enterprise buyers use to sanity-check quotes from every other research vendor.
Core technology: multi-model orchestration gets commercial validation
The orchestration thesis — that the winning research agent routes across frontier models rather than committing to one — just got a $2.6B lab behind it. Marlin's engine is Adaptive Branching Monte Carlo Tree Search (AB-MCTS, a NeurIPS 2025 Spotlight), extended into Multi-LLM AB-MCTS: the search tree dynamically picks which frontier model to invoke per sub-task, with a 7B "RL Conductor" trained via reinforcement learning to delegate among GPT/Claude/Gemini-class workers (sakana-marlin-ultra-deep-research). Sakana states the conviction outright:
"Each grows from the same conviction that runs through our research: that the most capable AI comes not from a single model, but from systems that reason over time and work together." — Sakana AI Launches Its First Commercial Product, Sakana Marlin
With Nvidia, Google, Citi, and MUFG among its investors, Sakana commercially validating model-agnostic orchestration legitimizes the architecture — but note how they sell it: as a closed, cloud-hosted service. The orchestration layer is the product; the customer's own models and hardware are nowhere in the loop.
What to watch: Whether "fugu-beta" and Sakana Chat extend the orchestration stack toward interactive use, or whether the lab stays pinned to the long-horizon deliverable niche (sakana-marlin-ultra-deep-research).
The gap Marlin leaves open: private data and local-first
The most expensive research agent in the market deliberately won't touch your corpus. Marlin's own FAQ scopes it narrowly:
"Sakana Marlin excels at hypothesis validation using aggregated public data... Conversely, it is currently not suited for (a) niche sectors with minimal public data, (b) use cases requiring split-second real-time tracking, or (c) research that relies entirely on internal, private data." — Sakana Marlin — Your Virtual CSO
Enterprise data policy is no-training-without-opt-in, but the deeper boundary stands: public web data in, cited deliverable out, nothing local (sakana-marlin-ultra-deep-research). This sharpens the map considerably. The orchestration-over-frontier-models thesis is now validated at the cloud B2B extreme — while local-first orchestration over the user's own hardware, models, and private data remains unclaimed at any price point. Hey, Lefty's lane isn't just adjacent to Marlin; it's the mirror image of the same architecture.
What to watch: Whether any competitor ships multi-model orchestration that runs against a customer's private corpus — the first mover there collapses the distance between Sakana's extreme and the local-first gap.
What surprised us
- The per-run price is the message, not the cost. At ~$42–$61 per run, Marlin isn't priced against Perplexity credits or Gemini tokens — it's priced against consulting hours (sakana-marlin-ultra-deep-research
). Publishing that number does more to define the enterprise band than any feature list.
- A lab with Nvidia and Google money validated multi-model orchestration — and left the local-first version of it on the table. The AB-MCTS/RL-Conductor stack is exactly the architecture a local-first tool would want; Sakana just only ships it cloud-hosted over public data (sakana-marlin-ultra-deep-research
- The FAQ is the best competitive intel in the launch. Vendors usually bury their non-fit cases; Sakana's explicit "not suited for internal, private data" is a free positioning document for anyone selling the opposite.
Open threads worth a vote
- [watch] Does any competitor ship orchestration over private/local data? — the mirror-image gap Marlin's FAQ just defined; steer follow-up on the topic page.
- [watch] Does the "Virtual CSO" deliverable-out framing spread? — whether hours-long runs become a category or stay a Sakana niche; steer follow-up on the topic page.