Executive assessment

OpenRouter is the stronger starting point for broad model access and explicit operational controls. Sakana AI’s Fugu is a candidate for managed, learned orchestration on difficult tasks. Neither should win a procurement decision on benchmark marketing alone. This is our assessment, not a measured head-to-head result.

The relevant comparison is Sakana Fugu versus OpenRouter’s inference platform, including its Auto Router and Fusion capabilities. Sakana AI is the company; Fugu is its multi-agent service. An API is a software interface; a coding harness is the application that lets a model work with files, tools, and a development environment.

This is not an exclusive choice: OpenRouter already distributes Sakana Fugu. Buyers can evaluate direct Sakana access, Fugu through OpenRouter, and other OpenRouter models or orchestration options. Buying through OpenRouter does not reveal Fugu’s private internal model decisions. Sakana catalog on OpenRouter

Fugu learns how to delegate, check, and combine work across models behind a single interface. OpenRouter unifies access to many providers, while Fusion adds multi-model deliberation. Consequently, “Sakana coordinates; OpenRouter only routes” is no longer an adequate distinction. Sakana launch explanation · OpenRouter Fusion documentation

Research date: October 1, 2026. This report is a light review of official documentation plus the specified video segment, not a performance benchmark or contractual audit. Prices are in US dollars and may change.

The Sakana segment in the video

The source is AI Revolution’s “GPT 6 SOL Leak, Gemini 4.0, DeepSeek V4.1 and More AI News,” uploaded September 12, 2026. This report analyzes only the Sakana discussion, approximately 13:35–15:36; the next story starts during the final overlapping caption cue. Watch the relevant section

The presenter describes Fugu Max and Ultra as orchestration systems sold through a familiar API, rather than standalone monolithic models. He highlights their pricing, research roots, swappable expert pools, claimed benchmark results, and the specialized Cyber variant. His central argument is economic: value could move from owning a particular model to coordinating models and controlling demand. He also names OpenRouter as a distribution integration.

Our assessment: the product description is consistent with the official documentation below, but the claim that model suppliers will lose premium margins is a forecast, not an established outcome. The benchmark discussion repeats supplier-reported results; this review did not independently reproduce them. “Not a model” is shorthand: the orchestration system itself uses a learned model as coordinator. Fugu technical explanation

Feature matrix

Decision dimension Sakana AI / Fugu OpenRouter
Primary role Managed orchestration: delegate reasoning and execution across expert models. [S1] Unified model access, provider routing, billing, and optional deliberation. [O1, O2]
Relationship Available directly and through OpenRouter. Distributes Fugu alongside alternatives; the products can be combined. [O8]
Product choices Fugu balances latency/quality; Ultra emphasizes quality; Max cost/performance; Cyber security tasks. [S2] Explicit model selection, Auto Router, and Fusion, among other options. [O2, O3]
Multi-model method Learned coordinator can delegate, verify, synthesize, and recurse. [S1] Fusion uses parallel panel responses, analyst comparison, then an outer model’s final answer; recursion is blocked. [O2]
Choice and visibility Standard Fugu allows model exclusions; Ultra/Max pools are fixed. Per-query underlying models are not disclosed. [S2] Paid catalog advertises 500+ models and 80+ providers. Auto Router returns the selected model; Fusion activity identifies participating models. [O3, O4, O5]
Integration OpenAI-compatible interface; Chat Completions, Responses, and Anthropic Messages; documented Claude Code/Codex launchers. [S3] OpenAI-compatible API and SDK ecosystem, with provider fallback options. [O1]
Privacy controls Training-use opt-out; enterprise pool customization by discussion. This is not evidence of universal zero retention. [S2] Zero-data-retention routing controls; enabled tools/plugins require separate policy review. [O6]
Regional availability Currently unavailable in EU/EEA. [S2] Business/Enterprise offer EU and US in-region inference routing. [O7]
Enterprise procurement Custom requirements require sales discussion; reviewed pages do not establish equivalent SSO/SLA terms. [S2] Enterprise lists SSO/SAML, managed policy enforcement, and contractual SLAs. [O4]

Sources: S1: Fugu launch, S2: Fugu product and FAQ, S3: Sakana integration guide, O1: OpenRouter FAQ, O2: Fusion, O3: Auto Router, O4: plans, O5: Fusion catalog, O6: retention controls, O7: Business, O8: Sakana catalog.

OpenRouter’s platform-level regional and privacy capabilities do not guarantee that every listed model is eligible. Confirm the exact Fugu endpoint’s availability and terms; do not treat an intermediary as a way around Sakana’s restrictions.

What the matrix means

Delegation is not the same as provider selection. Auto Router classifies tasks and uses recent market-spend rankings within cost bands; its cost tier is not a hard spending ceiling. Fugu instead delegates the internal solution strategy to its coordinator. Fusion provides a more explicitly configurable deliberation pattern. Auto Router mechanics · Fugu architecture · Fusion configuration

Transparency is a purchasing criterion. Fugu’s proprietary internal routing may be acceptable when output quality is the priority. Where a buyer needs model-level attribution, the visibility difference deserves explicit evaluation. Neither a panel’s agreement nor an internal verifier proves correctness; acceptance tests still matter.

Service tiers and pricing

Sakana Fugu

Offering Published price or billing basis
API subscriptions Standard $20/month; Pro $100/month (10× baseline usage); Max $200/month (20×). Absolute baseline allowance is not specified on the pricing page.
Fugu pay-as-you-go One underlying model: its standard rate. Multiple agents: one top-tier rate, rather than stacked model rates.
Ultra v2.0, per million tokens Input $5; output $30; cached input $0.50. Above 272K context: $10 / $45 / $1 respectively.
Max v1.0, per million tokens Input $2; output $6; cached input $0.25. Web search/fetch: $0.007 per call.
Cyber Contact sales.

Ultra’s orchestration tokens are separately billable even when reported inside usage-detail fields. API subscriptions do not change Sakana Chat limits. Pay-as-you-go traffic has higher priority than subscription traffic. Do not estimate a bill from visible answer tokens alone. Sakana pricing

OpenRouter

Offering Published price or billing basis
Free Limited free catalog; pricing grid lists 50 requests/day.
Standard Provider inference rates plus 5.5% fee on credit purchases.
Business Provider inference rates plus 8% fee on credit purchases; regional routing and expanded organization controls.
Enterprise Custom commercial terms; fee discounts available.
Bring your own key (BYOK) Standard/Business: first $25,000/month of list-price inference has no BYOK fee, then 5%; Enterprise allowance $200,000/month, then 5%.

The platform fee is on credit purchases, not a per-token markup. Standard and Business have no monthly minimum or seat license. BYOK also involves the customer’s upstream provider bill. Plan comparison · Fee basis

Fusion is not free inference. Its detailed documentation describes additional panel and analyst calls beyond the normal request, estimating roughly 4–5× a single completion for the default panel. The catalog’s “free” FAQ conflicts with its own paid-call description; budget using actual underlying usage, not that FAQ label. Fusion cost documentation · Conflicting catalog wording

These prices are not an apples-to-apples cost comparison. Different systems may consume different tokens, tools, retries, and review time to complete the same assignment.

Direct versus intermediary pricing also matters. OpenRouter lists Fugu Max at the same $2/$6 input/output rates, but lists web search at $10 per 1,000 calls versus Sakana’s direct $7 per 1,000. Its listing also identifies orchestration tokens as billable. Add the applicable OpenRouter platform fee and verify the chosen endpoint rather than assuming every charge is identical. OpenRouter Fugu Max listing

Recommendation for executives

Our recommended decision rule is:

  1. Start with OpenRouter when control and breadth dominate: choosing providers, comparing models, enforcing routing preferences, or managing regional requirements.
  2. Evaluate Fugu when outcome quality dominates: difficult coding, investigation, or analytical assignments where managed orchestration could reduce rework. Confirm pool, privacy, and geographic constraints first.
  3. Compare Fugu against both a strong single model and Fusion. Omitting Fusion would understate OpenRouter’s relevant capability.
  4. Use cost per accepted outcome, not token price, as the deciding metric. Include orchestration, tools, retries, elapsed time, and human review. Repeat representative tasks and score outputs without revealing the supplier.

No winner on quality or total cost is established by this research. A small, controlled evaluation is more defensible than choosing from incompatible vendor benchmark charts.

Implications for ELO

Octocore’s Engagement Lifecycle Orchestrator (ELO) addresses a different architectural question: preserving the work’s state, decisions, history, and artifacts as agents and harnesses change. That is distinct from selecting or coordinating the models that generate the next response.

Our proposed architecture keeps the durable engagement record outside either inference service. A provider can then be evaluated or replaced without making its account the sole home of the work’s history. This is a design recommendation—not a claim that a Fugu or OpenRouter integration with ELO has been validated in this report.

The market signal is encouraging but also competitive: multi-model access is becoming ordinary infrastructure. ELO’s strongest case should therefore be demonstrated continuity and traceability of the work, rather than simply access to more models.