Overview
Fugu Ultra v2 is the capability-first tier of Sakana AI's Fugu family, released on 11 September 2026 alongside the cost-first Fugu Max. Fugu is not a single trained network: Sakana describes it as "a Multi-Agent System, Delivered as One Model" that reaches its results "by dynamically coordinating and orchestrating a diverse pool of powerful models" behind one OpenAI-compatible endpoint. Which models it picks, and how it coordinates them, are proprietary and deliberately not exposed by the API.
The launch claim that distinguishes v2 is about independence rather than raw score. Sakana writes that "Fugu Ultra v2 achieves these scores without Fable 5, Fable 5.1, or GPT-6-Astra in its agent pool" and that it "does not rely on individual proprietary frontier models to deliver frontier output" — that is, the orchestration layer, not a borrowed flagship, is doing the work. The two published figures are 48.3 on Chartography, "which tests visual reasoning and data interpretation", against Opus 5 at 27.3 and Fable 5 at 29.5; and 74.3 on DeepSWE, "a benchmark for real-world software engineering". Sakana also says it "tops SWEFish".
Pricing is published on the Fugu product page against the model id fugu-ultra-v2.0: $5 per million input tokens, $30 per million output and $0.50 per million cached input, rising to $10 / $45 / $1.00 for contexts above 272K tokens. The line began as a beta in April 2026 ("Sakana Fugu: A Multi-Agent Orchestration System as a Foundation Model"), shipped as Fugu Ultra in June 2026, and picked up a v1.1 plus a Claude Code interface in July 2026. Sakana publishes no parameter count or context-window figure for the orchestrator itself, so none is listed here.
| Released | 2026-09-11 |
|---|---|
| License | Proprietary |
| Weights | API only |
| Architecture | Multi-agent orchestration — a pool of separately-trained models coordinated behind one endpoint rather than a single network. Sakana states the pool membership and routing policy are proprietary and not exposed. |
| Modalities | Text |
| Status | Generally available — served through Sakana AI's OpenAI-compatible API as fugu-ultra-v2.0. |
Benchmarks
Chartography, as reported by Sakana AI at launch
| Benchmark | Fugu Ultra v2 | Opus 5 | Fable 5 |
|---|---|---|---|
| Chartography | 48.3 | 27.3 | 29.5 |
This model's scores
- Chartography48.3
- DeepSWE74.3
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Pricing
| Input | $5.00 / 1M tokens |
|---|---|
| Cached input | $0.50 / 1M tokens |
| Output | $30.00 / 1M tokens |
Above 272K tokens of context the published rates rise to $10.00 input, $45.00 output and $1.00 cached input.
Strengths
- 48.3 on Chartography against Opus 5 at 27.3 and Fable 5 at 29.5, as reported by Sakana
- 74.3 on DeepSWE, Sakana's cited benchmark for real-world software engineering
- Reaches its published figures without Fable 5, Fable 5.1 or GPT-6-Astra in the agent pool, per Sakana
- Drop-in adoption: one OpenAI-compatible endpoint, with Sakana describing a single-line parameter change for existing Fugu users
- Explicit pricing above 272K tokens, so long-context cost is knowable in advance rather than unsupported
Best for
- Complex reasoning and analysis work where the capability tier matters more than the per-token price
- Real-world software-engineering agents, the workload DeepSWE and SWEFish are meant to measure
- Teams that want frontier-level output without pinning their stack to one vendor's flagship model
- Chart and data-interpretation tasks, the ground Chartography covers
How to access
| Provider | Model ID |
|---|---|
| Sakana AI API ↗ | fugu-ultra-v2.0 |
FAQ
Is Fugu Ultra v2 a single model?
No. Sakana AI describes Fugu as a multi-agent system delivered as one model: it coordinates a pool of separately-trained models behind a single OpenAI-compatible endpoint. Sakana states that which models it selects, and how it coordinates them, are proprietary and not exposed by the API.
What does Fugu Ultra v2 cost?
Sakana's Fugu product page lists fugu-ultra-v2.0 at $5 per million input tokens, $30 per million output tokens and $0.50 per million cached input tokens, rising to $10 / $45 / $1.00 for contexts above 272K tokens.
How does it compare with Fugu Max?
They share the same orchestration architecture aimed at two different missions: Fugu Ultra v2 is the capability tier for complex reasoning, while Fugu Max is the cost tier at $2 / $6 per million tokens. Both shipped on 11 September 2026.
Does it depend on a frontier model from another lab?
Sakana states that "Fugu Ultra v2 achieves these scores without Fable 5, Fable 5.1, or GPT-6-Astra in its agent pool" and that it "does not rely on individual proprietary frontier models to deliver frontier output".
Are the weights available?
No. Fugu is an API product; Sakana publishes no weights for the orchestrator, and the composition of its model pool is proprietary.