Sakana Fugu Ultra: An AI Orchestrator, Not a Model You Can Download

Sakana Fugu Ultra is not another LLM with weights for Ollama or MLX. Sakana AI sells it as an AI orchestrator: behind one OpenAI-compatible API, it selects and coordinates several specialized models for a request. Fugu Ultra is the performance-focused version for difficult, multi-step work where answer quality matters more than a quick response.

That distinction changes almost everything for Mac users. There is no GGUF download, no MLX build, and no local Apple Silicon benchmark. Rather than running a model on a Mac, the user sends work to a cloud service whose internal model selection is not exposed by Sakana.

Fugu Ultra productizes model routing and agent coordination rather than offering a new local foundation model. For U.S. readers, API token tiers and internal orchestration costs matter most; Sakana also explicitly excludes direct service in the EU/EEA.

What Sakana Fugu Ultra actually is

A single language model has to plan, reason, code, and check its own work for each task. Fugu Ultra takes a different approach: a smaller coordinator breaks down the task, assigns suitable agents from a model pool, and combines their work into one answer.

Sakana positions Fugu Ultra as the quality-first option alongside its faster standard Fugu model. The company says Ultra coordinates a deeper pool of specialist agents for demanding coding, research, and reasoning tasks. Sakana names paper reproduction, Kaggle competitions, cybersecurity analysis, and literature or patent research as early examples.

The core idea is not new. Multi-agent systems and model routing have existed for years. Sakana’s approach is to learn the coordination rather than require a customer to hand-design a workflow for every use case. It is based on the TRINITY and Conductor research projects:

1. A coordinator considers the request and the work completed so far.
2. It selects an agent and a role such as Thinker, Worker, or Verifier.
3. Agents work over multiple steps; a verifier can validate a result or trigger more work.
4. The system stops when an answer appears sufficiently checked or a defined budget is exhausted.

TRINITY describes a compact coordinator with a small decision head. Conductor studies how a model can learn communication patterns and focused instructions for other LLMs with reinforcement learning. This is a more accurate technical description than calling Fugu Ultra a new frontier model in its own right.

Fugu Ultra is neither open source nor local Mac AI

Anyone building local AI workflows on a Mac with Ollama, LM Studio, or MLX should keep the product categories separate:

Question	Answer
Can I download Fugu Ultra as GGUF?	No. Sakana does not publish local Fugu Ultra weights.
Does it run through MLX on Apple Silicon?	No. It is a hosted API service, not an MLX model.
Can a Mac still be part of the workflow?	Yes, as a development machine or API client. Inference and orchestration do not run locally.
Can I see the underlying models for each request?	No. Sakana does not expose routing or specific model selection.
Is Fugu Ultra open source?	No. Its research papers and technical report are public, but the commercial product is not.

Publishing research does not make the service an open model. For privacy or compliance reviews, an OpenAI-compatible endpoint is not enough either: Sakana does not disclose which models actually executed a request. Sakana’s FAQ allows provider opt-outs for standard Fugu, not for Ultra or Max, whose pools are fixed; custom enterprise arrangements require a separate discussion.

Historical launch benchmarks (June 2026, not Ultra v2)

At launch, Sakana published coding, reasoning, and long-context benchmark results. The table below shows a useful selection. Higher is better.

Benchmark	Fugu Ultra	Opus 4.8	Gemini 3.1 Pro	GPT-5.5	What it measures
SWE-bench Pro	73.7	69.2	54.2	58.6	software-engineering tasks using mini-swe-agent
Terminal-Bench 2.1	82.1	74.6	70.3	78.2	terminal and tool-use work
LiveCodeBench Pro	90.8	84.8	82.9	88.4	current programming tasks
Humanity’s Last Exam	50.0	49.8	44.4	41.4	difficult knowledge and reasoning questions
CharXiv Reasoning	86.6	84.2	83.3	84.1	scientific reasoning with figures
GPQA-Diamond	95.5	92.0	94.3	93.6	graduate-level specialist questions
SciCode	58.7	53.5	58.9	56.1	scientific programming
Long Context Reasoning	73.3	67.7	72.7	74.3	reasoning over long contexts
MRCRv2	93.6	87.9	84.9	94.8	retrieval in long contexts

The figures are vendor-reported, and they do not justify the shortcut claim that Fugu Ultra is best at everything.

The Fugu results are reported by Sakana. Sakana also uses provider-reported scores for some baseline values. Agent systems are sensitive to their tool scaffold, turn limit, prompts, and cost budget. For SWE-bench Pro, Sakana uses mini-swe-agent as the scaffold. A fair decision still needs reproducible settings and real tasks from the buyer’s own workflow.

The table itself explains why that caution matters: Fugu Ultra does not lead on SciCode, Long Context Reasoning, or MRCRv2. Orchestration can help with complex work, but it is not a universal guarantee for every benchmark or production task.

October 9, 2026: Fugu Ultra v2 and Fugu Max API pricing

Sakana launched Ultra v2 and Fugu Max on September 11, 2026. Ultra prioritizes complex tasks; Max focuses on cost-performance using a broader pool that includes NVIDIA Nemotron models. Both are hosted orchestration services, not models you can install on a Mac. Sakana’s release note.

Sakana’s price sheet names versioned billing IDs fugu-ultra-v2.0 and fugu-max-v1.0. The official API examples instead use the request aliases fugu-ultra (Ultra v2) and fugu-max. OpenRouter slugs sakana/fugu-ultra-v2 and sakana/fugu-max belong to a separate gateway with its own terms. A listed million-token context window does not mean a single flat price.

API list rate per 1M tokens (USD)	Ultra v2: up to 272K context	Ultra v2: over 272K context	Fugu Max: any context
Input	$5.00	$10.00	$2.00
Cached input	$0.50	$1.00	$0.25
Output	$30.00	$45.00	$6.00

The long-context surcharge still applies to Ultra v2. Do not assume the $5/$30 headline rate covers every request within its 1M-token context window. Sakana also bills the internal orchestration tokens at standard input/output rates; visible answer tokens alone understate real usage. At Sakana, web_search and web_fetch cost an additional $0.007 per call. As a base-token-only illustration, 100,000 input plus 10,000 output tokens cost $0.80 with Ultra v2 at standard context rates or $0.26 with Fugu Max; orchestration, cache, and tools can change the invoice. Sakana’s official pricing.

For U.S. customers, Sakana’s API platform also lists Standard $20/month, Pro $100/month, and Max $200/month subscriptions with fixed usage allowances (neither unlimited usage nor Sakana Chat subscriptions). These are USD list prices; applicable sales tax can be additional. OpenRouter is a separate billing provider: check the selected route, cache accounting, and current terms instead of assuming direct-Sakana invoices match.

Geography and privacy: Sakana currently says its direct service is not available in the EU/EEA. U.S. readers should still verify their account’s country eligibility, data-use settings, subprocessors, and contract before uploading private repositories. OpenRouter access does not establish that requests stay in the United States or that another provider shares Sakana’s terms. Sakana’s product and FAQ.

Which agents run? Sakana’s general site mentions provider opt-outs, but the detailed FAQ states that Ultra and Max use fixed pools; routine provider/model opt-outs apply to standard Fugu. Custom enterprise configurations may be negotiated. Do not assume the Ultra v2 model pool is configurable per request.

Using Fugu Ultra from a Mac: official API request and actual billed tokens

Sakana’s direct Fugu service is not available in the EU/EEA. This request example is for eligible accounts outside the EU/EEA, including approved U.S. customers. It is not a regional-workaround guide. Third-party gateways have separate access, processing and contract terms. A successful API request can incur charges.

Sakana’s official quickstart specifies https://api.sakana.ai/v1 and the alias fugu-ultra for Ultra v2. On a Mac, curl can check connectivity. Keep your key in the shell environment; never commit it to Git or print it in shared logs:

# Approved Sakana API accounts outside the EU/EEA only.
read -rs SAKANA_API_KEY
export SAKANA_API_KEY
curl --fail-with-body -sS https://api.sakana.ai/v1/chat/completions \
  -H "Authorization: Bearer $SAKANA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"fugu-ultra","messages":[{"role":"user","content":"Reply only with OK."}],"max_tokens":64}'

This is a vendor-documented request pattern, not an independently executed live test. Coding agents need timeouts, rate limits and a firm spending budget; the Mac is the client, not the inference host.

Why the visible reply is not the entire invoice

The official usage-field documentation distinguishes user-visible input/output from internal orchestration. Orchestration input and output tokens are separately billable. The cached_tokens and orchestration_input_cached_tokens fields are subsets of their respective input counts, not extra full-price units. Do not add total_tokens to already itemized charges a second time.

For pay-as-you-go rates per million tokens:

1. Uncached input = input_tokens − cached_tokens + orchestration_input_tokens − orchestration_input_cached_tokens.
2. Cached input = cached_tokens + orchestration_input_cached_tokens.
3. Output = output_tokens + orchestration_output_tokens.
4. USD bill = (uncached input × input rate + cached input × cache rate + output × output rate) / 1,000,000 + $0.007 per web_search/web_fetch call.

Two fully specified hypothetical cases (not benchmark measurements): The short request uses 100,000 input tokens (20,000 cached), 10,000 visible output tokens, 40,000 orchestration input tokens (10,000 cached), 20,000 orchestration output tokens and two paid web calls. That classifies as 110,000 uncached input, 30,000 cached input and 30,000 output tokens. The long request uses 300,000 input (60,000 cached), 12,000 visible output, 70,000 orchestration input (20,000 cached), 25,000 orchestration output and two web calls.

Example including 2 paid web calls	Ultra v2	Fugu Max
Short context up to 272K	$1.4790	$0.4215
Long context over 272K	$4.6590	$0.8360

Ultra’s over-272K pricing applies to the long-context request, not just the tokens above the threshold. Fugu Max has no long-context rate increase. Real agent calls can use more or fewer internal tokens. This example does not represent subscription allowances or an OpenRouter invoice.

Before integrating: check country eligibility; inspect the full usage breakdown and billable tools; apply a USD budget and timeout; and assess downstream providers and data policies. Sakana also offers Fugu Cyber, which requires separate approval and pay-as-you-go billing; it is not included as a free extra in regular monthly plans. Model catalog.

What does this mean for Mac users?

Fugu Ultra belongs in a Mac AI discussion, but not in the same way as a local model:

• As an API tool: A Mac can handle the editor, terminal, agent client, and data preparation. Fugu Ultra would be the remote reasoning and orchestration service.
• Not as private local AI: Prompts and working context leave the machine. Anyone intentionally processing confidential material locally with Ollama or MLX does not get an equivalent replacement.
• Not as a hardware benchmark: Unified memory, RAM, and Metal performance of an M1 through M4 Mac say nothing about Fugu Ultra’s model performance. Network quality, API cost, data handling, latency, and workflow design matter instead.
• As a product idea: Fugu points to a possible future where users do not choose a single frontier model for each task, but pay a service to select and coordinate them.

For local workflows, Ollama, MLX, LM Studio, and open-weight models remain the appropriate category. Fugu Ultra competes more directly with advanced cloud agents and premium APIs than with a Qwen or Llama download.

Who Fugu Ultra may suit, and who it does not

Potentially suitable: U.S. teams with verified account access that automate long, multi-step work and still review results professionally. Examples include reproducible research experiments, code analysis with clear test coverage, or structured literature reviews.

Not suitable: local-first workflows, organizations that require full transparency about every model provider involved, and tasks where a plausible but incorrect answer is unacceptable.

The same applies to security work. Models should only be used in clearly scoped, authorized environments. A provider example is not a security approval and does not replace human responsibility, testing, or release controls.

An orchestrator that removes model guessing

Sakana Fugu Ultra addresses a direct product question: why should users have to guess which model and agent workflow is best for every task? Sakana moves that decision into a proprietary orchestrator and reports strong, but not yet independently validated, benchmark results.

For a local Mac AI stack, Fugu Ultra is not a new model to install and test. It is a cloud service with an opaque model chain, material cost and privacy questions, and a direct EU/EEA access restriction. Its product idea is model routing and agent coordination, not local inference.