Fugu Max & Ultra v2: Sakana AI Pricing, Benchmarks (2026)

Sakana AI announced two new models, Fugu Max and Fugu Ultra v2, on September 10-11, 2026. Fugu Max prices at $2 input / $6 output per 1M tokens, while Fugu Ultra v2 prices at $5 input / $30 output. Neither is a single giant model — both are trained orchestrators that route tasks to a pool of models behind one API. Updated September 2026, here is a fast rundown of pricing, benchmarks, and when to use which.

Sakana AI claims Fugu Max's output pricing is 40-60% cheaper than Sonnet 5, Kimi K3, and GPT-5.6 Terra. This is Sakana AI's own claim, and no independent third-party verification is available yet.

Fugu Max and Ultra v2 depart from the conventional single-network LLM design that answers in one forward pass. Behind a single API endpoint sits a pool of open-weight and task-specialized models, and a trained orchestrator routes each request to the model it judges best suited, recursively calling itself when needed to verify and combine outputs from lower-level models before returning a final answer. This originates with the original Fugu launched in June 2026; the key change in this update is the split into a cost-focused Max and a performance-focused Ultra v2.

Fugu Max's model pool is designed for cost efficiency, routing to "the cheapest model that can solve it," and includes a wide range of candidates including the NVIDIA Nemotron family. Fugu Ultra v2, by contrast, prioritizes hard, multi-step tasks and reportedly deliberately excludes Fable 5, Fable 5.1, and GPT-6-Astra from its pool. The technical reasoning behind that exclusion has not been disclosed.

According to Sakana AI, Fugu Max posted the top score on six benchmarks — Terminal Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench, and SWEFish — and claims to extend the cost-performance frontier (achieving the same accuracy more cheaply) on 7 of 10 benchmarks tested.

Fugu Ultra v2 reportedly scored best or tied-best on 5 of 8 benchmarks. Notably it scored 48.3 on Chartography (a chart/diagram comprehension benchmark), well above the 27.3 comparison score for Opus 5. On the software-engineering benchmark DeepSWE it scored 74.3, and Sakana AI claims this beats models priced 3-5x higher.

The original Fugu and Fugu Ultra, launched June 22, 2026, were a single orchestrator lineage. This update splits that into a cost-focused Max and a performance-focused Ultra v2, each with its own pricing structure, model pool composition, and exclusion policy. Details such as parameter counts or training data beyond what is listed here remain officially undisclosed.

Fugu Max and Ultra v2 are served through an OpenAI-compatible API. If you already use the OpenAI SDK or a compatible library, the typical setup is simply to swap the base URL to Sakana AI's endpoint and set your API key. Availability through aggregators such as OpenRouter is possible but not yet confirmed — check the latest official announcements for specific access channels. Since weights are not released, local execution is not possible; usage is always through the hosted API.

For broader context on managing spend, see this AI API cost optimization strategy guide. Here is how Fugu Max and Ultra v2 compare with other leading models on price and fit.

Comparison prices are the standard published API rates as of September 2026; batch APIs, caching and discounts change the effective cost, so check each vendor's official pricing page. Note that Fugu Max output at $6 is roughly 60% below Sonnet 5 output at $15, which is consistent with Sakana AI's claimed 40-60% saving.

Another orchestrator-style offering worth knowing is Sakana Marlin, a deep-research agent. It's built for a different job than the Fugu line, so it's worth checking the intended use case when deciding between them.

- Hosted API only: weights are not released, so no local or self-hosted deployment is possible
- Not available in the EU/EEA
- Output reproducibility: since routing can select a different model per request, output style may vary even for identical prompts
- Vendor lock-in: the model pool composition and routing logic are managed entirely by Sakana AI and cannot be controlled by users
- Hard to estimate cost precisely: effective cost depends on which model a given request is routed to, making upfront estimates difficult