# Fugu Ultra ![Organization](https://img.shields.io/badge/Org-Sakana%20AI-e86e3a) ![Type](https://img.shields.io/badge/Type-Multi--Agent%20Orchestration-555) ![Context](https://img.shields.io/badge/Context-272K%2B%20tokens-1f6feb) ![Pricing](https://img.shields.io/badge/Price-%245%20%2F%20%2430%20per%201M-2da44e) ![License](https://img.shields.io/badge/License-Proprietary-d29922) ![Modality](https://img.shields.io/badge/Modality-Text-8957e5) ![Released](https://img.shields.io/badge/Released-Jun%202026-555) > **Quick answer:** Fugu Ultra v1.1 is Sakana AI's maximum-performance orchestration model, officially launched 24 July 2026. It coordinates a deep pool of frontier LLMs to maximise answer quality on hard, multi-step problems — improving up to 7.9 points over v1.0 with particularly strong gains on TerminalBench and SWE-bench Pro. Fugu Ultra v1.1 reports 95.5% on GPQA Diamond, 93.2% on LiveCodeBench, 90.8% on LiveCodeBench Pro, and 50.0% on Humanity's Last Exam — shoulder-to-shoulder with Anthropic's Fable 5 and Mythos Preview — at $5/$30 per million tokens. ## At a Glance **Where Fugu Ultra leads** - Frontier-tier reasoning and science: 95.5% on GPQA Diamond, 50.0% on Humanity's Last Exam. - Best-in-class coding: 93.2% on LiveCodeBench, 90.8% on LiveCodeBench Pro, 73.7% on SWE-bench Pro. - No export-control risk — delivers frontier capability without single-vendor dependency. - Strong long-context performance: 93.6% on MRCRv2, 73.3% on Long Context Reasoning. **Where it lags** - Higher latency than Fugu (standard tier) due to deeper agent coordination. - Agent pool is fixed — cannot opt out providers the way Fugu allows. - Not available in EU/EEA markets. - Architecture and routing decisions are proprietary and not inspectable via the API. **Best for:** AI research, paper reproduction, Kaggle competitions, cybersecurity analysis, literature and patent investigation, and any hard multi-step reasoning task where answer quality matters more than response speed. ## What Fugu Ultra Is Fugu Ultra is the high-performance tier of Sakana AI's Sakana Fugu product. Where the standard Fugu variant is tuned for balanced latency, Fugu Ultra coordinates a deeper, fixed pool of expert agents and invests more compute in verification and synthesis before returning an answer. This makes it significantly stronger on benchmarks that require sustained multi-step reasoning — and in real-world workflows where a single missed step cascades into errors. Like Fugu, Fugu Ultra is a language model trained to orchestrate other LLMs rather than a traditional fixed-weights model. Its coordination strategies are learned, grounded in Sakana AI's ICLR 2026 research: [TRINITY](https://arxiv.org/abs/2512.04695) (an evolved coordinator that assigns Thinker, Worker, and Verifier roles) and [The Conductor](https://arxiv.org/abs/2512.04388) (reinforcement learning for natural-language coordination patterns). In practice, Fugu Ultra can call instances of itself recursively as one of its own agents. Fugu Ultra v1.1, released 24 July 2026, improves over v1.0 (fugu-ultra-20260615) by up to 7.9 points, with the largest gains on TerminalBench 2.1 (+7.5 pts, 82.1%) and SWE-bench Pro (+4.5 pts, 73.7%). The v1.1 update incorporates the latest frontier models into the orchestration pool without any price change. Benchmark results show Fugu Ultra v1.1 matching or exceeding Anthropic's Fable 5 and Mythos Preview — currently among the highest-scoring closed models — on engineering, scientific, and reasoning benchmarks. Sakana AI notes these models are not in Fugu's agent pool (they are not publicly accessible), making the comparison an independent capability validation. ## Specifications | Field | Value | |---|---| | Organization | Sakana AI | | Model type | Multi-agent orchestration system (max quality tier) | | API identifier | `fugu-ultra-v1.1` (also `fugu-ultra-v1.0` / `fugu-ultra-20260615` for the prior version) | | Access | Closed, API only (OpenAI-compatible) | | Context window | 272K tokens (standard rate); extended context at $10/$45 per 1M tokens | | Architecture | Learned model orchestration (TRINITY + Conductor, ICLR 2026) | | License | Proprietary | | Release date | 24 July 2026 (v1.1); 22 June 2026 (v1.0 / fugu-ultra-20260615) | | Modality | Text | | Availability | Global (excluding EU/EEA) | ## Pricing | | Input (per 1M tokens) | Output (per 1M tokens) | Cached input (per 1M) | |---|---|---|---| | Standard context (≤272K) | $5.00 | $30.00 | $0.50 | | Extended context (>272K) | $10.00 | $45.00 | $1.00 | Fugu Ultra is available on both pay-as-you-go and subscription plans. Every subscription tier (Standard $20/mo, Pro $100/mo, Max $200/mo) includes access to Fugu Ultra. Source: [Sakana Fugu pricing](https://sakana.ai/fugu/). ## Context Window Fugu Ultra supports up to **272K tokens** at standard rates, with extended context at a premium. At 272K tokens this covers roughly **400+ pages** of text — full codebases, long research papers, or extended multi-document analysis. The MRCRv2 score of 93.6% confirms strong multi-document retrieval performance within this window. ## Fugu Ultra vs Alternatives | Model | GPQA-D | LiveCodeBench | HLE | Price (in/out per 1M) | |---|---|---|---|---| | **Fugu Ultra** | 95.5% | 93.2% | 50.0% | $5 / $30 | | [Fugu](/models/sakana-ai/fugu) | 95.5% | 92.9% | 47.2% | blended agent rate | | Fable 5 (Anthropic) | ~92.0% | ~87.8% | ~49.8% | Restricted (export controls) | | Mythos Preview (Anthropic) | ~92.0% | ~87.8% | ~49.8% | Restricted (export controls) | | [Gemini 3.1 Pro](/models/google/gemini-3-1-pro) | ~93.6% | ~85.3% | ~44.4% | $2 / $12 | Fugu Ultra's benchmark profile is competitive with the most capable closed models while being publicly accessible via API. The key differentiator vs Fable 5 and Mythos Preview is availability: those models are subject to export controls that restrict access, while Fugu Ultra is available globally (excluding EU/EEA). *(Rival scores sourced from [Sakana AI technical report](https://github.com/SakanaAI/fugu/blob/main/Fugu_technical_report.pdf), reported as max of Fable 5 and Mythos Preview where both scores exist.)* ## How Fugu Ultra Performs on Real Agent Tasks Sakana AI's beta programme (nearly 500 users) documented several compelling patterns for Fugu Ultra. In AutoResearch — an agent loop that iteratively edits training code, runs experiments, and keeps only changes that improve validation loss — Fugu Ultra ran 123 experiments on a single H100 over 14 hours and finished with the best mean bits-per-byte (0.9774 ± 0.0019), ahead of all tested frontier baselines. In a security assessment, a single scoped instruction drove a full recon-to-report pipeline (recon, XSS/SQLi checks, auth review, final report with evidence and retest steps) without exceeding scope or destructive actions. User testimonials from the beta emphasise three recurring themes: comprehensiveness (finding 20+ code review issues where single-model tools found 3), persona stability across long multi-turn sessions, and autonomous progress on open-ended research over hours — all signals of the sustained, multi-step coordination that distinguishes Fugu Ultra from single-call models. ## Use Fugu Ultra via API ```python from openai import OpenAI # Fugu Ultra exposes an OpenAI-compatible API client = OpenAI( api_key="YOUR_SAKANA_API_KEY", base_url="https://api.sakana.ai/v1", # replace with actual endpoint ) response = client.chat.completions.create( model="fugu-ultra-v1.1", messages=[{"role": "user", "content": "Reproduce the training setup from this paper and identify gaps..."}], ) print(response.choices[0].message.content) ``` ## Related Models & Benchmarks - Compare: [Fugu](/models/sakana-ai/fugu) · [Claude Fable 5](/models/anthropic/claude-fable-5) · [Claude Opus 4.8](/models/anthropic/claude-opus-4-8) · [Gemini 3.1 Pro](/models/google/gemini-3-1-pro) - Benchmarks: [SWE-bench Pro](/benchmarks/swe-bench-pro) · [LiveCodeBench](/benchmarks/livecodebench) · [GPQA Diamond](/benchmarks/gpqa-diamond) · [Humanity's Last Exam](/benchmarks/humanitys-last-exam) · [MRCRv2](/benchmarks/mrcr-v2) ## Frequently Asked Questions

What is Fugu Ultra?

Fugu Ultra is Sakana AI's maximum-performance AI model. The current version is v1.1, released 24 July 2026 (v1.0 was released 22 June 2026 as `fugu-ultra-20260615`). It is a multi-agent orchestration system that coordinates a deep pool of frontier LLMs to maximise answer quality on hard, multi-step problems. It is accessed via the API identifier `fugu-ultra-v1.1` through an OpenAI-compatible endpoint.

How does Fugu Ultra compare to GPT-5 or Claude?

Sakana AI's benchmark data shows Fugu Ultra at 95.5% on GPQA Diamond and 93.2% on LiveCodeBench, matching or exceeding Anthropic's Fable 5 and Mythos Preview. The key practical difference is availability: Fugu Ultra is publicly accessible, while Fable 5 and Mythos Preview have export control restrictions. Note that Fugu Ultra's scores come from multi-agent coordination rather than a single set of model weights.

How much does Fugu Ultra cost?

$5 per million input tokens and $30 per million output tokens for standard context (up to 272K tokens). Contexts above 272K are billed at $10/$45 per million. Cached input is $0.50/$1.00 per million. Source: Sakana AI pricing page.

What is Fugu Ultra's context window?

Fugu Ultra supports 272K tokens at standard rates, with extended context available at a higher rate ($10/$45 per 1M tokens).

Is Fugu Ultra open source?

No. It is a proprietary, closed model accessible only through Sakana AI's API. There are no downloadable weights.

What is the difference between Fugu and Fugu Ultra?

Fugu balances performance and latency for everyday coding, code review, and chatbot use cases. Fugu Ultra coordinates a deeper, fixed pool of expert agents to maximise quality on hard, multi-step problems — at the cost of higher latency. Fugu also allows opting out specific providers; Fugu Ultra's pool is fixed.

--- *Specs and scores sourced from [Sakana AI's Fugu v1.1 announcement](https://sakana.ai/fugu-1-1-claude-code-interface/) and the [Fugu product page](https://sakana.ai/fugu/). Last updated 2026-07-25.*

Fugu Ultra — Evaluations

Evaluations