Sakana AI Launches Fugu Max and Fugu Ultra v2 on 11 September 2026
On 11 September 2026, Sakana AI released two new models: Fugu Max v1.0 and Fugu Ultra v2.0, available via an OpenAI-compatible API endpoint. Fugu Ultra v2.0 achieves 74.3 on DeepSWE v1.1 — the autonomous software engineering benchmark that uses real pull-request tasks from open-source repositories — placing it above Anthropic Claude Opus 5.0, OpenAI GPT-5.6 Sol, and several commercial agentic coding products on the same leaderboard at a fraction of their token cost. On the Chartography benchmark, which measures visual reasoning and the interpretation of charts and structured data, Fugu Ultra v2.0 scores 48.3, compared with 27.3 for Opus 5 and 29.5 for Fable 5. The 11 September release is the first time Sakana has offered Fugu as a split product family, with Ultra targeting maximum benchmark performance and Max targeting cost efficiency for high-volume workloads.
The Architecture: TRINITY and Conductor
Fugu is not a single trained neural network. Sakana describes it as a Multi-Agent System delivered through one OpenAI-compatible endpoint. The system routes each incoming task across a dynamically selected pool of constituent models, assigning each a role — Thinker, Worker, or Verifier — before synthesising their outputs into a single response. Two coordination mechanisms underpin the system. The first, TRINITY, uses a lightweight evolved coordinator to assign roles across turns based on the content and structure of the task. The second, Conductor, is trained with reinforcement learning to discover natural-language coordination strategies: it generates focused sub-prompts for each constituent model and adapts them based on intermediate outputs. Both architectures are grounded in research published at ICLR 2026. The external API exposes the orchestration layer as a single model endpoint, so callers interact with Fugu the same way they would interact with any hosted language model — there is no visible multi-agent complexity from the caller's perspective.
What Fugu Ultra v2 Achieves on Key Benchmarks
Fugu Ultra v2.0 scores best or joint-best on five of the eight benchmarks Sakana published at launch: GDP.pdf (a financial document reasoning benchmark), Chartography, SWEFish (a software engineering evaluation), DeepSWE v1.1, and Toolathon (a tool-use benchmark). On DeepSWE v1.1, the 74.3 result places it above every single-model competitor listed in Sakana's published comparison table. Fugu Ultra v2 supports a 1,000,000-token context window and up to 128,000 completion tokens per call. The expanded completion token limit is directly relevant to autonomous software engineering tasks, where generating large code patches or long technical documents in a single pass requires sustained generation capacity that many frontier models cap at 8,000 to 32,000 tokens.
Fugu Max: The Cost-First Tier
Fugu Max v1.0 is Sakana's cost-optimised tier in the Fugu family. It achieves best overall scores on six benchmarks at launch, with wins on Terminal Bench 2.1, GPQAD, AA-LCR, and AutomationBench among those Sakana named. Sakana positions Fugu Max for teams running high-volume agent tasks where per-call cost is the primary constraint, rather than achieving maximum accuracy on any single difficult evaluation. Both Max and Ultra share the same TRINITY and Conductor coordination architecture but draw from different constituent model pools and apply different resource budgets per call. The split gives engineering teams a direct cost-versus-accuracy lever: route simpler or higher-volume tasks through Max and reserve Ultra for complex reasoning and code tasks.
Pricing and Access
Fugu Ultra v2.0 is priced at $5.00 per million input tokens and $30.00 per million output tokens, with cache reads at $0.50 per million tokens and web search at $10.00 per 1,000 calls. Sakana reports that Ultra is priced roughly three to five times lower per token than frontier single models that score below it on DeepSWE, meaning the performance-per-dollar improvement is substantial. Fugu Max pricing was not published separately at launch. Both models are available via the Sakana API and through OpenRouter with the same OpenAI-compatible interface, making drop-in integration straightforward for teams already using LiteLLM, LangChain, or any SDK that targets the OpenAI chat completions endpoint.
What This Means for AI Product Teams in India
Indian engineering teams building agents, coding assistants, and retrieval-augmented workflows face a recurring trade-off between the capability of large frontier models and the per-call cost of running them at scale. Fugu's architecture — dynamically coordinating a pool of models behind a single OpenAI-compatible endpoint — is a reference design for how product teams can assemble higher-capability systems without routing every query to the most expensive single model. For teams building on AWS India or Azure India, Fugu Max offers a compelling profile for high-frequency agent tasks: better performance on agentic and tool-use benchmarks than individual frontier models, at costs competitive with mid-tier single models. The DeepSWE 74.3 result on Fugu Ultra v2 also raises the floor for what Indian software companies can expect from AI-assisted code review, automated pull-request resolution, and test generation running as background agents in their CI/CD pipelines.
The Bottom Line
On 11 September 2026, Sakana AI released Fugu Max v1.0 and Fugu Ultra v2.0 through an OpenAI-compatible API. Fugu Ultra v2.0 scores 74.3 on DeepSWE v1.1 and 48.3 on Chartography, placing it above Opus 5 and GPT-5.6 Sol on both benchmarks. The system routes tasks across a coordinated pool of constituent models using TRINITY and Conductor, two architectures from ICLR 2026, behind a single endpoint. A 1M-token context window and 128K completion token limit make Ultra viable for long-horizon software engineering and document reasoning tasks. Fugu Max targets high-volume cost-sensitive workloads. Ultra is priced at $5.00 per million input tokens and $30.00 per million output tokens. For product teams in India building AI agents and coding assistants, Fugu demonstrates how multi-model orchestration behind a single endpoint can deliver frontier benchmark performance at sub-frontier cost.
Frequently Asked Questions
What are Fugu Max and Fugu Ultra v2, and when did Sakana AI launch them?+
Fugu Max v1.0 and Fugu Ultra v2.0 are two AI models launched by Sakana AI on 11 September 2026, available through an OpenAI-compatible API endpoint and on OpenRouter. Both are multi-agent orchestration systems, not single trained neural networks: Sakana describes them as Multi-Agent Systems delivered as one model, coordinating a pool of specialist constituent models behind a single endpoint. Fugu Ultra v2.0 targets maximum benchmark performance, scoring 74.3 on DeepSWE v1.1 and 48.3 on Chartography, above Anthropic Claude Opus 5.0 and OpenAI GPT-5.6 Sol on both evaluations. It supports a 1,000,000-token context window and up to 128,000 completion tokens per call. Fugu Max v1.0 targets cost-efficiency for high-volume agent workloads, achieving best scores on six benchmarks including Terminal Bench 2.1. Both models use the same TRINITY and Conductor coordination architecture.
How does Sakana Fugu's TRINITY and Conductor architecture work?+
Fugu routes each task across a dynamically selected pool of constituent models using two coordination mechanisms developed for ICLR 2026. TRINITY uses a lightweight evolved coordinator that assigns one of three roles — Thinker, Worker, or Verifier — to each constituent model across the turns of a conversation, structuring the multi-agent exchange around the specific demands of the task. Conductor is trained using reinforcement learning to discover natural-language coordination strategies: it generates focused sub-prompts for each constituent model and adapts them based on intermediate outputs, effectively learning which framing and sequencing of instructions produces the best final answer. From the caller's perspective, both mechanisms are invisible: the orchestration layer presents as a single OpenAI-compatible model endpoint, and the caller sends a standard chat completion request without specifying how the underlying models are coordinated.
What benchmark scores does Fugu Ultra v2.0 achieve?+
Fugu Ultra v2.0 achieves 74.3 on DeepSWE v1.1, the autonomous software engineering benchmark that evaluates models on real pull-request tasks from open-source repositories. This result places it above Anthropic Claude Opus 5.0 and OpenAI GPT-5.6 Sol in Sakana's published comparison table. On Chartography, a benchmark measuring visual reasoning and structured data interpretation, Fugu Ultra v2.0 scores 48.3, compared with 27.3 for Opus 5 and 29.5 for Fable 5. The model scores best or joint-best on five of the eight benchmarks Sakana published at launch: GDP.pdf, Chartography, SWEFish, DeepSWE v1.1, and Toolathon. The model is priced at $5.00 per million input tokens and $30.00 per million output tokens, which Sakana reports is three to five times lower per token than frontier single models that score below it on DeepSWE.
What does Sakana Fugu mean for AI product teams building agents in India?+
For Indian engineering teams building AI agents, coding assistants, and automation pipelines, Sakana Fugu's multi-model orchestration architecture offers two practical benefits. First, the OpenAI-compatible endpoint means teams already using LangChain, LiteLLM, or any SDK targeting the OpenAI chat completions API can integrate Fugu with minimal code changes. Second, the split between Fugu Max and Fugu Ultra gives teams a direct cost-versus-accuracy lever: route simpler or higher-volume tasks through Max to control costs, and use Ultra for complex reasoning and agentic coding tasks where benchmark quality matters. For teams running on AWS India or Azure India, Fugu Max is particularly relevant for high-frequency agent tasks — automated code review, CI pipeline integration, and document processing — where per-call cost determines whether the economics of running agents at scale are viable.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.
