An Orchestration API Posing as a Model
Sakana AI released Fugu Max and Fugu Ultra v2 on September 11, 2026, both running behind the same OpenAI-compatible endpoint as the existing Fugu and Fugu Cyber models. Neither is a single neural network. Each is an orchestration layer that picks from a swappable pool of open and specialized models — including NVIDIA Nemotron checkpoints through Sakana's collaboration with NVIDIA — to answer a request. The pricing is per million tokens, the API is OpenAI-compatible, and a single key covers all four models.
This article covers what the four-model lineup is, what the console pricing actually looks like for each tier, where the orchestration layer beats single-model providers on benchmarks, and where the cost model has traps that a budget should not miss.
The Four-Model Lineup
Sakana positions the Fugu family as four tuned configurations of the same orchestration architecture, picked by workload rather than by benchmark.
| Model | Position | Tuning goal | Self-serve pricing? |
|---|---|---|---|
| Fugu | Default | Balanced performance and latency | Yes |
| Fugu Max | Cost-efficient Pareto | Best $/output across multi-step workloads | Yes |
| Fugu Ultra v2 | Peak performance | Hard benchmarks, agentic coding, visual reasoning | Yes |
| Fugu Cyber | Security specialist | Cybersecurity domain workflows | Sales-led |
The September 11 release post frames the design choice directly. The company's thesis is that the cost-performance frontier is two-dimensional, not one: capability on one axis, cost on the other. Fugu Max pushes the frontier northwest (more capability per dollar); Fugu Ultra v2 pushes it upward (higher peak capability at a higher price). Both rely on the same underlying orchestration machinery, which means existing Fugu customers can switch to Max or Ultra v2 with a single-line parameter change. There is no separate API surface to integrate.
Pricing on the Console
The console at console.sakana.ai/pricing lists token rates for three of the four models. Fugu Cyber is sales-led and does not appear in the public rate card. All rates are per million tokens; web_search and web_fetch are billed at $0.007 per call in addition to token charges.
| Model | Context | Input /1M | Output /1M | Cached input /1M |
|---|---|---|---|---|
| Fugu Max | Any | $2.00 | $6.00 | $0.25 |
| Fugu Ultra v2 (standard) | ≤ 272K | $5.00 | $30.00 | $0.50 |
| Fugu Ultra v2 (long context) | > 272K | $10.00 | $45.00 | $1.00 |
| Fugu | Per console | Pay-as-you-go | Pay-as-you-go | Per console |
Two structural details stand out. First, Fugu Max uses fixed pricing regardless of context length — there is no doubled tier above 272K, unlike Fugu Ultra v2. For a workload that routinely pushes long context, that difference makes Fugu Max materially cheaper than Ultra v2 once the boundary is crossed. Second, the cached-input ratio on Fugu Max is roughly one-eighth of the input rate ($0.25 vs $2.00), and on Fugu Ultra v2 it is one-tenth ($0.50 vs $5.00). For an agent that carries a long system prompt or skill file, the cached tier is the single biggest lever on the bill.
The console page also exposes the orchestration pricing model. Sakana states that when a request activates a single underlying model, you pay only the standard rate for that model. When multiple agents are activated, Sakana does not stack model fees — only the top-tier model involved is billed at its standard rate. The Responses API usage fields break out user-visible model work from orchestration work, so a cost dashboard can track them separately. Standard OpenAI responses show a single input/output aggregate; Sakana exposes the split.
Subscription and Token Plans
Two ways to pay for the API Platform. Sakana Chat has no subscription and is unaffected by either.
| Plan | Price | Best for | Notes |
|---|---|---|---|
| Standard subscription | $20/month | Lightweight daily usage, occasional agent sessions | Monthly-plan tokens served at standard priority |
| Token Plan | Pay-as-you-go | Heavy production workloads needing maximum reliability | Consumption tokens served at higher priority than monthly-plan tokens |
The priority detail matters for production. Even when a subscription is active, pay-as-you-go tokens are served at higher priority than monthly-plan tokens, so a heavy workload should top up with Token Plan credits rather than relying on subscription tokens alone. Sakana is direct about this on the pricing page: pay-as-you-go is positioned as the production tier.
Where the Orchestration Wins
The benchmark story is the reason the orchestration layer is worth evaluating at all. Fugu Max wins or ties the top score on six benchmarks in Sakana's published charts; Fugu Ultra v2 wins or ties the top score on five and places top two on seven of eight. The Pareto framing is that Fugu Max expands the cost-performance frontier (the northwest push), and Fugu Ultra v2 pushes the peak (the upward push). Both achieve these scores without Fable 5, Fable 5.1, or GPT-6-Astra in their agent pool.
| Benchmark | Fugu Max | Fugu Ultra v2 | Notes |
|---|---|---|---|
| Terminal Bench 2.1 | Best overall | — | Agentic coding terminal tasks |
| GPQAD | Best overall | — | Graduate-level QA |
| AA-LCR | Best overall | — | Long-context reasoning |
| GDP.pdf | Best overall | Best or joint-best | Document processing |
| AutomationBench | Best overall | — | Browser/automation agent tasks |
| SWEFish (internal) | Best overall | Best or joint-best | Sakana's internal coding benchmark |
| Chartography | — | 48.3 (Opus 5: 27.3, Fable 5: 29.5) | Visual reasoning over charts |
| DeepSWE | — | 74.3 | Real-world software engineering |
| Toolathon | — | Best or joint-best | Tool-use agent benchmark |
Three readings of the chart matter. First, Fugu Max sits at a Pareto point single-model providers cannot reach — performance within striking distance of frontier closed models at two to six times lower token spend. Second, Fugu Ultra v2's headline figures (Chartography 48.3, DeepSWE 74.3) are competitive with frontier closed models that cost three to five times more per token. Third, both beat their benchmarks without proprietary frontier models in the agent pool, so the architecture is the differentiator rather than the underlying weights.
The API Surface
Both new models and the original Fugu and Fugu Cyber all sit behind one OpenAI-compatible endpoint. Sakana's console exposes the standard Chat Completions path with streaming, plus a Responses API that breaks out orchestration tokens separately. Existing Fugu customers upgrade by changing one parameter in their request payload — there is no migration step.
The architectural implication is worth keeping in mind. Because Fugu is orchestration, the response time is bounded below by the latency of the slowest agent in the pool. The September release notes that Fugu Max and Fugu Ultra v2 use the same architecture optimized for two distinct missions, so they share the same orchestration latency floor. For latency-sensitive interactive workloads, this is the structural cost of using an ensemble rather than a single-model provider. The cached-input ratios partially offset it for prompt-heavy workloads, but they do not change the round-trip floor.
The Trade-Offs to Weigh
The orchestration pricing model has a soft ceiling. When a single underlying model is selected, you pay that model's standard rate. When multiple agents fire, only the top-tier model in the chain is billed at its standard rate — the orchestration layer does not stack fees. That is generous in steady-state, but it also means the cost dashboard sees only one rate per request even when several models contributed. If your usage profile is skewed toward requests that activate long chains, you are getting more compute than the line item shows.
Fugu Cyber is not self-serve. The console pricing page sends interested users to contact Sakana. If cybersecurity is the workflow, the price is whatever Sakana quotes — there is no published rate to budget against.
Context pricing tiers apply to Ultra v2 only. Fugu Max uses fixed pricing regardless of context length. Fugu Ultra v2 doubles input, output, and cached-input rates above 272K context. A workload that routinely exceeds that boundary should budget against the long-context tier; a workload that stays under it pays the standard tier.
The subscription is a discount on light usage, not on production. The console page is explicit that pay-as-you-go tokens are served at higher priority than monthly-plan tokens. If production reliability matters, top up with Token Plan credits even when a subscription is active. The subscription is the right answer for a developer exploring the API; Token Plan is the right answer for a production workload.
Web tools are billed per call, not per token. web_search and web_fetch are charged at $0.007 per call on top of token charges. An agent that issues many web searches per task can drive a meaningful share of the bill from the per-call line alone. Build your cost model to track search calls separately from tokens.
The Bottom Line
Sakana's Fugu family is structurally different from the usual per-model API catalog: four configurations of the same orchestration architecture, behind one OpenAI-compatible endpoint, with per-million-token pricing that the company publishes on a public console. Fugu Max is the cost-efficient tier at $2/$6, Fugu Ultra v2 is the peak tier at $5/$30 with a doubled long-context tier, and the cached-input ratios are aggressive enough to matter for prompt-heavy workloads.
The headline benchmarks are real — Fugu Ultra v2's 48.3 on Chartography and 74.3 on DeepSWE — and they were achieved without Fable 5, Fable 5.1, or GPT-6-Astra in the agent pool. The trade-offs are the ones the architecture implies: a soft cost ceiling that does not always reflect the work done, a sales-led Fugu Cyber tier, context-tiered pricing on Ultra v2 that you have to plan around, and a per-call web-tools bill that can dominate the token cost on search-heavy workloads.
- Fugu Max: $2 input / $6 output / $0.25 cached per 1M, fixed regardless of context length
- Fugu Ultra v2: $5 / $30 / $0.50 standard (≤ 272K); $10 / $45 / $1.00 long context
- OpenAI-compatible endpoint; Responses API exposes user-visible vs orchestration tokens
- Subscription $20/month (Standard, lightweight); Token Plan (pay-as-you-go, higher priority)
- Fugu Cyber: enterprise pricing via sales contact, no public rate card
- Web tools billed at $0.007 per call on top of token charges