Sakana Marlin: The "Ultra Deep Research" Extreme — 8-Hour Runs, 100-Page Reports, Sold as a Virtual CSO
Sakana AI's first commercial product, Sakana Marlin (launched June 15, 2026, per David Ha's launch post), defines the long-horizon extreme of the deep-research market Market Map & Positioning: Where Does Hey, Lefty Fit?. Where consumer deep-research tools (OpenAI, Google, Perplexity) return 5–10 page briefings in roughly 3–30 minutes, Marlin runs autonomous reasoning loops for up to eight hours and returns a report of up to a hundred pages plus presentation-ready executive slides. The launch announcement leads with the temporal pitch: "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."1 (sakana.ai/marlin-release)
Positioning and buyers. Billed as a "Virtual CSO" (Chief Strategy Officer), it targets "corporate strategy and business-planning teams at financial institutions and operating companies, consulting firms, think tanks, and research houses" — strictly B2B, no consumer tier. The workflow is topic-in, deliverable-out: a brief scoping exchange, then the human steps away while Marlin cycles hypothesis formation, web gathering, and verification; the output is framed as "structured strategic options ready for executive review" rather than a summary. Public sample reports include Strait of Hormuz blockade resolution scenarios, the global AI-regulation patchwork, and the return of the bond vigilantes — each delivered as a cited PDF plus a PPTX deck. A closed beta ran from April 2026 with ~300 professionals from financial institutions, operating companies, consulting firms, and think tanks.
Pricing (product page, yen-denominated; USD conversions via VentureBeat):
- Pay-as-you-go: add-on credits at ¥98 (~$0.61) each, 100 credits per run →
¥9,800 ($61) per run2 - Pro: ¥150,000/month (
$936) including 2,000 credits → 20 runs/month ($47/run effective) - Team: ¥400,000/month (
$2,495) including 6,000 credits → 60 runs/month ($42/run effective) - Enterprise: custom quote
This is the third published per-run meter in the market, alongside Exa Agent Ultra's budgeted subagent runs Exa Agent Ultra: Deep Research Becomes a Metered Subagent-Swarm API — With Self-Graded Benchmarks and Parallel's per-request CPM Parallel: Deep Research Sold as Per-Request Backend Infrastructure — with Audit-Grade Provenance and Shapley-Value Payouts to Publishers — and the highest per-run price of the three, consistent with selling the equivalent of weeks of CSO-team work rather than minutes of chat. It slots into the enterprise band of Research-Tool Market Segmentation & Price Points, Mid-2026 — and the Verified Capability Limits.
Core technology. The engine is Adaptive Branching Monte Carlo Tree Search (AB-MCTS, NeurIPS 2025 Spotlight, arXiv:2503.04412, "Wider or Deeper?"), extended into Multi-LLM AB-MCTS: the search tree dynamically chooses which frontier model to invoke per sub-task — e.g. one model for ideation, a reasoning-heavy model to audit and correct earlier nodes — treating frontier models as a plug-and-play network instead of committing to one3. The lineage is Sakana's research stack: The AI Scientist (published in Nature), ALE-Agent (winner of the AHC058 combinatorial optimization contest), and the 7B "RL Conductor" orchestrator trained via reinforcement learning to delegate among GPT/Claude/Gemini-class workers. The release notes the roadmap continues in the same direction, pointing to Sakana Chat and "work that coordinates frontier models" (fugu-beta).
Data boundaries — and the gap this leaves open. 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.ai/marlin) Enterprise data policy: no training or fine-tuning on customer inputs without explicit opt-in. So the most expensive research agent in the market deliberately does not touch the user's private corpus or local machine — exactly the local-first, private-data lane Hey, Lefty occupies per Market Map & Positioning: Where Does Hey, Lefty Fit?.
Why it matters for Hey, Lefty's positioning. Sakana states the orchestration thesis 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/marlin-release) A lab valued at $2.6B (Series B, Nov 2025) with Nvidia, Google, Citi, and MUFG among its investors is now commercially validating model-agnostic multi-model orchestration — but selling it as a closed, cloud-hosted, public-data-only B2B service at $47–$61 per run. That both legitimates the orchestration thesis and sharpens the contrast: closed cloud B2B versus local-first orchestration over the user's own hardware, models, and private data.
-
An instance of Automating deep research requires extended reasoning compute, not instant data retrieval. — The deep-research extreme is sold as hours of extended reasoning compute rather than instant retrieval. ↩︎
-
An instance of Long-horizon agentic loops collapse under the unpredictable cost of reasoning tokens. — Eight-hour agentic runs force per-run metered pricing rather than a flat subscription feature. ↩︎
-
An instance of Reliable agentic workflows must now be routed across hot-swappable, model-agnostic architectures. — A $2.6B lab makes hot-swappable, model-agnostic multi-model routing its core commercial architecture rather than a single-provider dependency. ↩︎