Fugu Ultra v2

Sakana AI announced Fugu Ultra v2 on September 11, 2026, alongside Fugu Max, as part of its orchestration-based model family. Both share the same core orchestration architecture but are optimized for different missions: Fugu Max targets the lowest possible cost per task, while Fugu Ultra v2 targets the absolute highest The announcement explicitly states that Fugu Ultra v2 achieves its results "without Fable 5, Fable 5.1, or GPT-6-Astra in its agent pool," meaning it does not depend on those frontier models to deliver high-capability output. Sakana frames this as evidence that an orchestration layer can match or surpass closed ecosyst

Contra Collective reports that Sakana AI shipped Fugu Ultra v2 on September 11, 2026, positioning it as an orchestration model rather than a single generator. The article explains that behind a single endpoint, a coordinator routes and combines output from a pool of specialized language models, with Sakana grounding th On reported benchmark numbers, Sakana claims Fugu Ultra v2 is best or joint-best on five of eight benchmarks and top-two on seven of eight, including 74.3 percent on DeepSWE and 48.3 percent on Chartography, where the article notes it clears Claude Opus 5 (27.3) and Claude Fable 5 (29.5) by wide margins. Pricing is $5

BigGo Finance's September 14 coverage restates Sakana's September 11 announcement, confirming that Fugu Ultra v2's 74.3 DeepSWE score edges out GPT-6-Astra's 74.1 and Claude Fable 5.1's 67.4, and its 48.3 Chartography score substantially exceeds Claude Opus 5's 27.3. The piece notes that existing users can migrate to t API pricing is reported as $5 input / $30 output per million tokens for Ultra v2 and $2 / $6 for the cost-optimized Fugu Max, with Fugu Max claiming 40–60% lower per-output costs versus Sonnet 5, GPT 5.6 Terra, and Kimi K3. The article references Sakana's June technical report describing Fugu itself as a language model

Yahoo Finance / Forkast News reports that on September 11, 2026, Sakana AI launched Fugu Max v1.0 and Fugu Ultra v2.0, framing both as orchestration engines rather than monolithic models built on learned model coordination from ICLR 2026 papers TRINITY and Conductor. Fugu Max is priced at $2 per million input tokens an On benchmarks, the piece reports Fugu Max achieves the best overall score on six evaluations (Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, SWEFish), while Fugu Ultra achieves best or joint-best on five of eight benchmarks including GDP.pdf, Chartography, DeepSWE, Toolathon, and SWEFish. The article argu

The Unwind AI newsletter for September 11, 2026 covers OpenAI's Codex harness becoming an Agents API, then shifts to Sakana AI's launch of Fugu Max and Fugu Ultra v2 as orchestration models that route work across a swappable pool of open and specialized models. It states Fugu Max approaches elite-model performance at 2 The newsletter also covers Edge0 running a 35B sparse mixture-of-experts model on Apple Silicon by streaming weights from SSD, achieving 14.9-17.7 tokens per second on an M4 Pro Mac mini with about 2.9 GB of active memory from a 23 GB on-disk footprint, and DeepSeek's V4.1-Flash, the smallest model in a new architectur

DataNorth's write-up details that both Fugu Max and Fugu Ultra v2 are released as hosted API only on September 11, 2026, and are currently unavailable in the EU or EEA. Fugu Max costs $2 per million input tokens and $6 per million output tokens, which Sakana claims undercuts Sonnet 5, GPT 5.6 Terra, and Kimi K3 by 40–6 For Fugu Ultra v2 specifically, DataNorth reports base pricing at $5/$30 per million input/output tokens with a long-context surcharge kicking in above 272,000 tokens (where pricing rises to a higher tier), and no published cached input rate. The piece reinforces the architectural framing that Fugu is an orchestrator m

There's An AI For That (TAAFT) documents Fugu Ultra v2 as the capability-focused member of Sakana AI's second-generation Fugu family, released alongside the cost-optimized Fugu Max and announced September 10, 2026. The directory entry describes it as a learned orchestrator built on Sakana's ICLR 2026 TRINITY and Conduc On benchmarks, the entry reports Fugu Ultra v2 achieves the best or joint-best score on five of eight benchmarks tested, including a Chartography score of 48.3 versus 27.3-29.5 for competing models and a DeepSWE score of 74.3, placing in the top two on seven of eight benchmarks overall. The model's agent pool deliberat

AI Weekly's alert reports that Sakana AI shipped Fugu Max on September 11, 2026, priced at $2 per million input tokens and $6 per million output tokens, which the company claims runs 40-60% below the output cost of Sonnet 5, GPT 5.6 Terra, and Kimi K3. The system routes each task across a swappable pool of open-weight On Fugu Ultra v2 specifically, the piece pairs it with Fugu Max as a higher-capability variant on the same routing architecture, quoting 48.3 on Chartography against Opus 5's 27.3 and 74.3 on DeepSWE. Sakana frames Ultra v2 as "What is the absolute highest capability we can achieve on complex, multi-step tasks?" versus

LumiChats confirms Sakana AI released Fugu Max v1.0 and Fugu Ultra v2.0 on September 11, 2026, as the second generation of its Fugu orchestration line. Both are coordinator models behind one OpenAI-compatible endpoint, routing prompts across open-weight and specialized models using Thinker, Worker, and Verifier roles. The design descends from Sakana's TRINITY and Conductor research presented at ICLR 2026. The piece reports Fugu Ultra v2 claims best-or-joint-best on five of eight benchmarks and top-two on seven, scoring 48.3 on Chartography and 74.3 on DeepSWE. Sakana's training and eval cutoff is August 28, 2026, and the pool explicitly excludes Fable 5, Fable 5.1, and GPT-6 Astra. The editorial raises skepticism about benchmarking systems that exclude the models they surpass on paper.

AI/TLDR's model profile for Fugu Ultra v2 explicitly names the model id fugu-ultra-v2.0 and characterizes it as the capability-first tier of Sakana AI's Fugu family, released on September 11, 2026 alongside the cost-first Fugu Max. It confirms Fugu is "a Multi-Agent System, Delivered as One Model" that coordinates a di The profile reports published benchmarks of 48.3 on Chartography (versus Opus 5 at 27.3 and Fable 5 at 29.5) and 74.3 on DeepSWE, plus leadership on SWEFish, all sourced from Sakana. Pricing is $5 input / $30 output / $0.50 cached input per million tokens at base tier, rising to $10/$45/$1.00 for contexts above 272K to

Sakana AI shipped Fugu Max and Fugu Ultra v2 on September 11, 2026, per this editorial analysis. Fugu Ultra v2 is priced at $5 per million input and $30 per million output tokens, rising to $10/$45 above 272K context, while Fugu Max costs $2/$6. Both expose a single API surface and orchestrate a pool of open-weight and specialized models including NVIDIA's Nemotron family. The article reports Fugu Ultra v2 placed in the top two on seven of eight Sakana benchmarks, including a 48.3 on Chartography versus Opus 5's 27.3 and Fable 5's 29.5, and 74.3 on DeepSWE. Crucially, Sakana states Fable 5, Fable 5.1, and GPT-6 Astra are excluded from the agent pool. The result is framed as a benchmarked argument that orchestration now beats raw model capability.

Sakana AI's official product page positions Fugu as "a Multi-Agent System, Delivered as One Model" — a learned coordinator that dynamically assembles agents from a pool of models and routes work through non-obvious collaboration patterns, exposed via a single OpenAI-compatible API. The page frames Fugu as outperforming single-vendor setups on complex, multi-step tasks while reducing API complexity for callers. It notes current EU/EEA unavailability pending GDPR compliance. The official page also lists partner gateways where Fugu is reachable, including OpenRouter and Vercel's AI Gateway, signaling a multi-channel distribution strategy. Marketing emphasizes frontier-level performance without dependence on any single vendor's model. Quantitative benchmarks and version-specific v2 details are not present in the scraped excerpt, so version-level claims must be sourced from derivative coverage.