Sakana AI releases Fugu Max and Fugu Ultra v2.0 efficiency-model updates

On Friday 2026-09-11 (2:15 AM JST), Sakana AI announced and shipped two new members of its Fugu orchestration family, both "available today via our standard OpenAI-compatible API":

1. Fugu Max (v1.0) — the cost-performance variant. Orchestrates "our largest pool of open and specialized models to date," explicitly including NVIDIA's Nemotron open-model family (via the Sakana–NVIDIA collaboration announced on NVIDIA/open-model work). Priced at $2 per 1M input tokens and $6 per 1M output tokens (cached input $0.25; web_search/web_fetch tools $0.007/call; flat rate regardless of context); 1M-token context window; 128K max output.
2. Fugu Ultra (v2.0) — the peak-performance flagship. Priced at $5 input / $30 output / $0.50 cached input per 1M tokens, rising to $10/$45/$1.00 above 272K context; 1M-token context; 128K max output. Training cutoff 2026-08-28. Sakana stresses that Fable 5, Fable 5.1 and GPT-6-Astra are NOT in Fugu Ultra v2's model pool — the system reaches its benchmark results by orchestrating a swappable pool of open and specialized models instead of relying on individual closed frontier models.

Headline claims (COMPANY CLAIM unless noted): Fugu Max achieves the best overall score on six benchmarks (Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, SWEFish — the last being Sakana's internal benchmark), expands the cost-performance Pareto frontier on seven of ten benchmarks, and its output pricing is 40-60% lower than Sonnet 5, GPT 5.6 Terra and Kimi K3. Fugu Ultra v2 achieves best-or-joint-best on five of eight benchmarks (GDP.pdf, Chartography, SWEFish, DeepSWE, Toolathon) and top-2 on seven of eight; Chartography 48.3 vs Opus 5's 27.3 and Fable 5's 29.5; DeepSWE 74.3, outperforming models that cost three-to-five times more per token (BigGo's independent recap puts DeepSWE 74.3 vs GPT-6 Astra 74.1 and Claude Fable 5.1 67.4).

The release page frames the two as "the same core orchestration architecture optimized for two distinct missions" — Fugu Max = best output at lowest cost; Fugu Ultra v2 = highest capability on complex multi-step tasks. "Both models are available today… upgrading to Max or Ultra v2 requires a single-line parameter change. No migration. New architecture, same API." Distribution: direct console (console.sakana.ai) plus OpenRouter, Vercel AI Gateway, opencode (models.dev), Creao and Merge; subscription tiers $20/$100/$200 per month also include Fugu Max and Ultra.

An in-window follow-up: on Sep 17, 2026, Sakana Chat (the consumer product) rolled Fugu Max out as its default model alongside a new cross-conversation memory feature — the largest live consumer exposure of the new orchestrator within the same window.

• Before: Sakana Fugu existed as Fugu (balanced latency/quality) and Fugu Ultra (v1.1-class, performance-optimized) plus Fugu Cyber, launched GA June 22, 2026 with an April beta; the July v1.1 release added Fugu-Cyber and a Claude Code interface; the NVIDIA/Nemotron partnership was announced July 16, 2026. There was a single "performance vs balanced" axis, and Fugu Ultra carried no versioned "Max"-style cost tier. Pricing war positioning was "frontier capability without export-control risk" (vs Anthropic's Fable/Mythos export-control revocation of June 12, 2026).
• Change (Sep 11): Sakana split the product line onto two explicit axes of the cost-performance plane: Fugu Max pushes the cost Pareto frontier outward (biggest pool, cheapest output class among frontier-adjacent products), and Fugu Ultra v2 pushes the performance frontier upward (explicitly without Fable 5/5.1 or GPT-6-Astra in the pool). Pricing became public and aggressive ($2/$6 for Max; $5/$30 for Ultra v2), the Nemotron pool integration landed, and "orchestration" was productized as a standardized, API-compatible category.
• After: As of the window's end, Sakana Fugu is a four-model lineup (Fugu, Fugu Ultra v2, Fugu Max, Fugu Cyber) behind one OpenAI-compatible API; Sakana Chat defaults to Fugu Max; independent commentary (Forkast, DataNorth) has begun analyzing "orchestration arbitrage" — value migrating from model providers to orchestrators; and the pricing war framing has moved from per-token model prices to per-task orchestration economics, in the same week DeepSeek's V4.1-Flash (S46) reset open-weight efficiency economics.

• Orchestration model, not a monolithic LLM: Fugu is itself a (comparatively small) language model trained to call other LLMs — "Fugu is itself a language model trained to call various LLMs in an agent pool, including instances of itself recursively." It decides when to answer directly and when to assemble a team of experts, handling model selection, delegation, verification and synthesis internally. The user hits one OpenAI-compatible endpoint.
• Research foundation: the learned-coordination approach builds on two ICLR 2026 papers — TRINITY ("An Evolved LLM Coordinator": a lightweight evolved coordinator assigning Thinker/Worker/Verifier roles; arXiv:2512.04695) and the Conductor (RL-trained natural-language coordination; arXiv:2512.04388) — plus the Fugu Technical Report (arXiv:2606.21228; submitted Jun 19, 2026, revised Jun 23, 2026), which describes large-scale fine-tuning, evolutionary algorithms and reinforcement learning behind the production system.
• Model-pool design: the pool is "swappable by design" — a resilience property: if a provider restricts access, Fugu routes around the disruption. Fugu Ultra v2's headline point is that it excludes the very frontier models it benchmarks against (Fable 5, Fable 5.1, GPT-6-Astra are not in the pool; training cutoff 2026-08-28), relying on open-weights and specialized models, including NVIDIA Nemotron (coding/tool-calling/instruction-following strengths) in Fugu Max's larger pool.
• Pool-configuration policy (from the product FAQ): Fugu Ultra's pool is fixed (it needs the full pool for peak performance); Fugu Max's pool is fixed (it must jointly optimize cost and performance); only base Fugu supports opting specific providers/models out for compliance reasons. Enterprise customers can get customized pool configurations via sales.
• Pricing mechanics: token-plan rates are blended per request — "we never stack model fees; you are charged a single rate based on the top tier model involved" — so orchestrating four models does not mean four bills. Tools (web_search, web_fetch) are metered separately at $0.007/call under Fugu Max. Token usage and cost are reported per request.
• What stays hidden: the FAQ states plainly (Q9) that the specific models selected and the coordination used are proprietary and not exposed by design — users cannot see which underlying models handled a given query. This is the central trade-off of the product (see Risks).
• Timeline of the "Fugu journey" (from the release page): April 2026 beta → June 2026 GA + Fugu Ultra v1 → July 2026 Fugu-Cyber & Claude Code interface → August 2026 Sakana Chat + NVIDIA Nemotron integration → September 2026 (today) Fugu Max + Fugu Ultra v2. (Nuance: the partnership page itself is dated July 16, 2026, while the release page's timeline puts "NVIDIA Partnership" under August; minor internal inconsistency, immaterial to the event.)

• The pricing war moved to "orchestration economics." Fugu Max's $2/$6 output pricing (40-60% below Sonnet 5 / GPT 5.6 Terra / Kimi K3 per Sakana; the per-token prices were independently confirmed via OpenRouter's live catalog during this research) is a different kind of price cut: instead of a single model getting cheaper, a routing layer gets cheaper per task. Forkast's same-day analysis ("The Orchestration Arbitrage") argues value migrates from model providers to orchestrators — if an orchestrator can deliver frontier-level output by routing to cheaper specialists, single-model providers lose premium-margin pricing power.
• Counter-narrative to monolithic scaling: Sakana's core argument — "the most capable AI will never come from a single monolithic model" — lands the same week the industry is debating pacing (S15 Amodei, S22 von der Leyen, S36 Trump): orchestration is offered as a scaling path of its own, competing with the "bigger model" axis without needing a trillion-parameter flagship.
• Export-control / AI-sovereignty positioning realized: born from the June 2026 Anthropic Fable/Mythos export-control disruption, Fugu is explicitly marketed as "the resilient, vendor-agnostic infrastructure required for true AI sovereignty." The Sep 11 release demonstrates the thesis with benchmarks that exclude the export-controlled frontier models entirely — a Japan-based answer to U.S. closed-frontier dependence.
• Open-model ecosystem leverage: Fugu Max's bet is that the Pareto frontier of the future will be "built out of many open, specialized models working in concert," with NVIDIA Nemotron folded in via the July 2026 partnership — giving NVIDIA a distribution story for its open models and Sakana a moat-able orchestration layer.
• Japanese AI leadership signal: a Tokyo lab shipping an enterprise-grade orchestration product with a fully bilingual release, consumer rollout (Sakana Chat), and a clear sovereignty pitch to Japanese enterprises/government is a significant marker for Japan's AI industrial policy this quarter.

• One-line migration to a cheaper/faster or stronger tier: existing Fugu API users change a single parameter to move to Fugu Max or Fugu Ultra v2 — no SDK migration, same OpenAI-compatible surface.
• Frontier-adjacent agent workloads at $6/1M output: Fugu Max's price point makes sustained agent loops (which burn output tokens across many sub-calls) materially cheaper, within a pool that includes open-weights workhorses rather than premium closed frontier APIs.
• Benchmark-competitive autonomous execution without the export-controlled frontier: Fugu Ultra v2 claims SWEFish/DeepSWE/Chartography-class results while excluding Fable 5/5.1 and GPT-6-Astra from its pool — so buyers in export-constrained or sovereignty-conscious environments can access frontier-adjacent capability.
• Vendor-agnostic orchestration as a buyable product: instead of wiring multi-agent frameworks themselves, teams can buy "multi-agent system as a model" with per-request cost visibility, subscription or token billing, and integrations through OpenRouter/Vercel/opencode/Merge.
• Consumer-scale demonstration: Sakana Chat defaulting to Fugu Max (Sep 17) puts the orchestration bet in front of everyday users, with memory features — evidence of orchestration running interactive product loads.

1. How large is Fugu Max's pool, exactly? "Largest pool to date, including NVIDIA Nemotron" — no count, no roster is published; pool size and composition are core product facts that remain undisclosed.
2. Do the headline benchmark numbers survive independent re-runs? Chartography 48.3 / DeepSWE 74.3 / Terminal Bench 2.1 leadership need third-party execution (the same standard applied to DeepSeek's V4.1-Flash tables this week).
3. When does the pool update next? Sakana's FAQ says ~two weeks to train/evaluate a new Fugu version after a major model release — with GPT-6 Astra and Fable 5.1 now mainstream, will they ever enter the pool, and does the "excluded frontier" claim degrade as those models age?
4. Does per-task orchestration spend actually beat single-model spend in production? DataNorth's "$1,100-volume" arithmetic and per-task-cost advice await real-world telemetry.
5. When does EU/EEA availability arrive under GDPR work, and how will EU AI Act GPAI attribution treat the hidden pool?
6. What will closed labs do in response — build their own orchestration layers, tighten API ecosystems, or keep cutting single-model prices (the Forkast "what to watch")?
7. How much margin does Sakana hold on $2/$6 blended rates when underlying open/specialized models carry their own token costs, and does the blend survive a pool-model price change?

• In-window (Sep 17, 2026): Sakana Chat defaults to Fugu Max with a new memory feature — first large consumer-scale rollout of the new orchestrator.
• Weeks ahead: per the FAQ's ~2-week pool-update cadence, a next Fugu iteration could land quickly as new models stabilize; first independent re-runs of the v2 benchmark claims; third-party per-task cost analyses (Forkast/DataNorth follow-ups); possible pricing responses from OpenAI/Anthropic/Google on single-model tiers.
• Quarter ahead: EU/EEA availability decision under GDPR compliance work; enterprise custom-pool deals; Japan public-sector adoption signals (MoD/SCSK/Sumitomo cluster already visible); whether major closed labs launch competitive orchestration layers or harden API ecosystems.

Fugu Max and Fugu Ultra v2 are not primarily a benchmark story — they are the first clean productization of "the routing layer is the product," with pricing ($2/$6, $5/$30) that reframes the AI cost conversation from per-token to per-finished-task. The technical claim that earns attention is Fugu Ultra v2 delivering frontier-adjacent results without Fable 5, Fable 5.1 or GPT-6-Astra in its pool: orchestration as an anti-lock-in scaling path, born from the June export-control shock and aimed squarely at "AI sovereignty." The discipline note for readers: every headline number is Sakana-published, the pool internals are deliberately invisible, and real-world per-task economics — not list prices — will decide whether the orchestration-arbitrage thesis holds. In a week dominated by calls to pace the frontier, Sakana's answer (with DeepSeek's V4.1-Flash alongside it) was to make frontier-class agent capability cheaper, swappable and less dependent on any single lab — a forceful market-level counterpoint to the pause narrative.

Evidence-status summary: CONFIRMED (event, date, availability, pricing — primary sources + OpenRouter catalog check); COMPANY CLAIM (Pareto-frontier performance, benchmark leadership, "40-60% cheaper," "surpasses GPT-6 Astra," sovereignty framing); INDEPENDENT EVIDENCE (OpenRouter live listing/prices verified during this research; Forkast, GIGAZINE, 窓の杜, BigGo, DataNorth recaps; June-launch VentureBeat/TechCrunch background); INTERPRETATION/PREDICTION as labeled inline.