|

Sakana AI Launches Fugu Max and Fugu Ultra v2 for Cheaper, Stronger Multi-Agent Orchestration

ByRicardo

Sakana AI has launched Fugu Max and Fugu Ultra v2, 2 new fashions in its Sakana Fugu household. Fugu isn’t a single basis mannequin. It is a discovered orchestrator that routes work throughout a pool of different fashions behind 1 API. The new launch tunes that structure for 2 missions. Fugu Max targets one of the best output per greenback. Fugu Ultra v2 targets the very best functionality on laborious, multi-step duties.

Is it deployable? Yes, as a hosted API. Both fashions are stay in the present day by means of Sakana’s OpenAI-compatible API. There are not any open weights to self-host, and Sakana doesn’t supply the service within the EU/EEA.

Why Sakana Frames This as a 2-Axis Problem

Sakana’s argument is direct. Real workloads are judged on functionality and value collectively. Sending a easy information lookup to a multi-trillion-parameter mannequin wastes cash. A greater system picks the most cost effective equipment that may nonetheless remedy the duty.

Sakana group describes this with the Pareto frontier. On that frontier, gaining high quality prices extra, and slicing value loses high quality. Fugu Max and Fugu Ultra v2 share 1 core orchestration structure. Only the optimization goal differs.

The launch follows a quick cadence. Fugu entered beta in April, reached general availability in June, and added Fugu-Cyber and a Claude Code interface in July.

How Fugu Orchestration Works

The Sakana Fugu’s Technical Report describes Fugu fashions as language fashions in their very own proper. They learn a question and construct an agentic scaffold for it on the fly. Training combines large-scale fine-tuning, evolutionary algorithms, and reinforcement studying.

The system builds on 2 ICLR 2026 papers. TRINITY makes use of a light-weight developed coordinator that assigns Thinker, Worker, or Verifier roles throughout turns. The Conductor is skilled with reinforcement studying to find natural-language coordination methods and targeted prompts.

Fugu Max: More Models, Less Cost

Fugu Max widens the pool of fashions Fugu can orchestrate. It provides a big set of open-weights and specialised fashions. That consists of the NVIDIA Nemotron household, by means of Sakana’s collaboration with NVIDIA. Fugu Max routes every process to the leanest mannequin able to fixing it.

Sakana group experiences the next:

  • Pricing: $2 per 1M enter tokens and $6 per 1M output tokens.
  • Output value: 40% to 60% decrease than Sonnet 5, GPT 5.6 Terra, and Kimi K3.
  • Performance: Best total rating on 6 benchmarks: Terminal Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench, and SWEFish.
  • Efficiency: Expands the cost-performance Pareto frontier on 7 of 10 benchmarks.

Sakana locations Fugu Max inside hanging distance of elite fashions at 2x to 6x decrease value. SWEFish is an inside Sakana benchmark constructed from its personal coding challenges. Treat that outcome as a vendor sign.

Fugu Ultra v2: Raising the Ceiling

Fugu Ultra v2 targets complicated reasoning, autonomous analysis, and full-stack software program growth. Its largest beneficial properties seem on sustained reasoning over visible and structured information.

  • Chartography (visible reasoning and information interpretation): 48.3, versus 27.3 for Opus 5 and 29.5 for Fable 5.
  • DeepSWE (real-world software program engineering): 74.3, forward of fashions priced 3x to 5x increased per token.
  • Breadth: Best or joint-best on 5 of 8 benchmarks: GDP.pdf, Chartography, SWEFish, DeepSWE, and Toolathon.
  • Consistency: Top 2 on 7 of 8 benchmarks.

Fable 5, Fable 5.1, and GPT-6-Astra should not in Fugu Ultra v2’s agent pool. The mannequin’s coaching cutoff is August 28, 2026. Sakana’s predominant message is frontier output with out dependence on any 1 proprietary mannequin. The analysis group states that this reduces publicity to vendor lock-in, API revocations, and sudden service cutoffs.

Interactive Explainer

Key Takeaways

  • Fugu Max prices $2/$6 per 1M enter/output tokens and targets output per greenback.
  • Fugu Max posts one of the best total rating on 6 benchmarks and expands the frontier on 7 of 10.
  • Fugu Ultra v2 scores 48.3 on Chartography and 74.3 on DeepSWE.
  • Ultra v2 reaches these scores with out Fable 5, Fable 5.1, or GPT-6-Astra in its pool.
  • Both ship in the present day by way of an OpenAI-compatible API, with a 1-line swap for current customers.

Check out the Technical details and Project page. Also, be at liberty to comply with us on Twitter and don’t neglect to hitch our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Need to companion with us for selling your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar and so on.? Connect with us

The put up Sakana AI Launches Fugu Max and Fugu Ultra v2 for Cheaper, Stronger Multi-Agent Orchestration appeared first on MarkTechPost.

Similar Posts