AI orchestration is moving beyond simple model selection. Sakana AI has released Fugu Max and Fugu Ultra v2, while Microsoft and GitHub have introduced HydraFusion, giving developers new ways to route work across multiple models, control costs, and tackle difficult multi-step tasks.
The shift matters because these systems do not treat every request the same way. Instead of sending every task to one model, they decide how different models should work together, bringing model capability and spending into the same equation.
Sakana AI Pushes Fugu Into Its Next Phase
Sakana AI released Fugu Max and Fugu Ultra v2 on September 10, 2026. Fugu is a learned orchestrator that routes work across a pool of other models behind one API, giving developers access to a model-selection layer without managing each model separately.
Fugu Max targets the best output per dollar. Its pricing is $2 per 1M input tokens and $6 per 1M output tokens, with an output price that is 40% to 60% lower than Sonnet 5, GPT 5.6 Terra, and Kimi K3. The model also adds a large set of open-weights and specialized models, including the NVIDIA Nemotron family.
Those cost claims arrive alongside strong benchmark results. Fugu Max has the best overall score on 6 benchmarks and expands the frontier on 7 of 10 benchmarks, according to Sakana AI’s published results.
Fugu Ultra v2 takes a different aim: the highest capability on hard, multi-step tasks. It scores 48.3 on Chartography and 74.3 on DeepSWE, reaching those scores without Fable 5, Fable 5.1, or GPT-6-Astra in its pool.
Fugu Ultra v2 has a training cutoff of August 28, 2026. Both Fugu models are live today through Sakana AI’s OpenAI-compatible API, but Sakana AI offers no open weights to self-host, and the service is not available in the EU/EEA.
Fugu has moved through several stages this year. It entered beta in April, reached general availability in June, and added Fugu-Cyber and a Claude Code interface in July.
HydraFusion Turns Routing Into Workflow Design
GitHub announced HydraFusion on September 4, 2026. Microsoft developed the model routing system, which is available through GitHub as a research preview in Copilot CLI.
HydraFusion routes each coding request across models in real time instead of sending every task to one model. It evaluates the request and assigns one of three execution patterns: Single, Cascade, or Critique.
That difference separates HydraFusion from GitHub’s auto mode, a model routing feature that launched earlier this year. Auto model selection chooses a single model, while HydraFusion orchestrates a workflow that can use different models and stages.
Mario Rodriguez, chief product officer at GitHub, described the distinction this way: “routing to the right model is quickly becoming table stakes, but where HydraFusion is different is that it is addressing the ‘what’s the best way to solve this task’ rather than ‘which model should handle this task?’”
He also said, “HydraFusion doesn’t just prompt a model, it dynamically constructs an execution strategy — whether a task is best handled by a single model, should start with a faster model and escalate, or if the work would benefit from an independent model reviewing and improving the result.”
The system can cut estimated cost by up to 67% compared with using Claude Opus 5 alone. GitHub’s offline evaluations show HydraFusion scored 4.9 percentage points higher on verified task quality than the Opus 5 baseline on TerminalBench 2.1, while costing 67% less.
On DeepSWE, HydraFusion scored 1.5 percentage points below Opus 5 on quality, with 36% lower estimated cost. On CheckpointBench, it measured 0.1 percentage points below Opus 5 on quality, while cutting estimated cost by 65%.
GitHub’s own numbers show HydraFusion matched or exceeded Opus 5 in one of three benchmarks. Awan Farz, a developer who analyzed the published results, explained the cost mechanism: “You pay the cheap model on every cascade request. The more expensive model only runs on the subset of tasks that fail the quality gate.”
More Routing Systems Enter the Field
HydraFusion joins a growing set of systems built around model routing. Nvidia’s NeMo Switchyard, released in August, is marketed as maintaining frontier-level accuracy while cutting task cost to roughly a third of running Claude Opus 4.8 alone.
Nvidia’s published benchmark shows that routing 7% of calls to a frontier model cut spending 74%, while giving up a measurable slice of accuracy. That result highlights the central tradeoff facing every orchestrator: savings depend on deciding which requests need the strongest model and which can use a cheaper path.
OpenRouter’s new Auto router also launched in August. It claims to outperform its predecessor across a wide spectrum of tasks and cost levels, although its benchmark scores below the old router in two of five test categories.
These releases point toward a new layer in AI software. The question is no longer only which model has the highest score, or which API costs the least. Fugu Max, Fugu Ultra v2, HydraFusion, NeMo Switchyard, and OpenRouter’s Auto router all put the orchestration strategy at the center of the product.
As these systems develop, developers will compare more than raw model capability. They will also measure how each orchestrator handles difficult workflows, how much it saves, and when it decides that a task deserves a stronger model.
Based on
