Tested and written up.
New
Added Sep 20, 2026
Fugu Ultra v2 is the high-capability sibling to Fugu Max, released alongside it September 11, 2026. Where Fugu Max optimizes for cost, Fugu Ultra v2 asks 'what is the absolute highest capability achievable on complex, multi-step tasks,' pricing at $5 per million input tokens and $30 per million output tokens at standard context lengths. Sakana reports Fugu Ultra v2 scored 48.3 on the Chartography benchmark against 27.3 for Anthropic's Claude Opus 5 and 29.5 for Claude Fable 5, without including either closed model anywhere in its orchestration pool.
Why: The Chartography gap over Opus 5 and Fable 5, this site's own leaderboard leaders, is large enough to be worth flagging even though it's a single Sakana-run benchmark. It's a real signal that learned orchestration over open models can compete with frontier closed models on specific task types, not a general claim that it beats them everywhere.
New
Added Sep 20, 2026
Fugu Max is one of two orchestration models Sakana AI released September 11, 2026. Fugu is not a single monolithic model but a learned orchestrator: it receives one API request, decides which models in its pool should handle the task, and stitches their answers back together. Fugu Max is tuned for the question 'what is the best possible output at the lowest possible cost,' pricing at $2 per million input tokens and $6 per million output tokens, which Sakana claims runs 40-60% below Claude Sonnet 5, GPT-5.6 Terra and Kimi K3.
Why: Orchestration-as-a-model is a distinct approach from either a single frontier model or a manual multi-agent pipeline, and the published pricing gap is large enough to matter for high-volume workloads. The tradeoff is that output quality depends on which models are in Sakana's pool at request time, an opaque, moving dependency the site can't fully audit from the outside.
New
Added Sep 20, 2026
Atria Dawn Preview is an open-weight preview release from Shanghai AI Laboratory, released around September 11, 2026 under the MIT license. Built on the 744B-parameter MoE GLM-5.2 foundation model, it accepts text input only with a 256K-token context window and is purpose-built for long-horizon research and engineering agents: combining task objectives with environmental feedback to carry a scientific method through problem analysis, tool use, code implementation, experiment execution, result analysis and failure recovery. Across 16 published benchmarks spanning real-world research, engineering and digital work, Shanghai AI Lab reports it is competitive with frontier agents and achieves the top reported score on five of them.
Why: It's a rare open-weight, MIT-licensed model aimed specifically at the full research-agent loop rather than chat or coding alone, and it's self-hostable, which none of the closed frontier agent products are. Its benchmark suite is the lab's own and hasn't been widely replicated yet, and running a 744B MoE at full precision is out of reach for most individual users.
New
Added Sep 20, 2026
Jev is TypeSafe AI's first System One Model, published September 18, 2026 by former OpenAI researcher Diogo Almeida after two years in stealth. It is not an LLM and does not generate text: it takes application state and a structured question and returns a typed, calibrated probability that software can act on directly, what the company calls 'a frontier-intelligence function call.' It uses parallel sampling to produce the whole output at once instead of token-by-token, trained with a method TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD). Because the output space is fixed in advance, TypeSafe says it cannot hallucinate. Published workflow evals claim 70-500ms response times and up to 193.6x faster, 444.6x cheaper results than equivalent-intelligence LLM calls on...
Why: It's a genuinely new model class rather than another chat model: built for the automation gap the founder points at directly, 'models have been superhuman at chat for years, so where is all the automation?' The catch is that every multiple above is TypeSafe's own published eval, it's gated behind an early-access waitlist with no public self-serve signup yet, and it only helps if your problem is a bounded decision, not open-ended generation.