Sakana Fugu Ultra v2 vs GPT-6 Astra: Orchestration Against the Flagship on Benchmarks and Cost (2026)

Sakana AI shipped Fugu Ultra v2 on September 11, 2026, and it forces a question the last two years of model releases have mostly ducked: do you want one very large model, or a system that coordinates several smaller ones. Fugu is an orchestration model. It does not generate answers itself so much as decide how a pool of specialized language models should work together on a task, then route and combine their output. GPT-6 Astra, released September 3, is the other philosophy in its purest form: one monolithic frontier model, 1 million token context, 10 dollars per million input tokens and 50 per million output. Comparing them is comparing a strategy to a component.

Attribute	Fugu Ultra v2	GPT-6 Astra
Vendor	Sakana AI	OpenAI
Released	September 11, 2026	September 3, 2026
Type	orchestration over a model pool	single frontier model
Input price	5 dollars per million tokens (to 272K)	10 dollars per million tokens
Output price	30 dollars per million tokens (to 272K)	50 dollars per million tokens
Above 272K context	10 and 45 dollars per million	flat
DeepSWE	74.3 percent	not reported by Sakana
Headline claim	best or joint-best on 5 of 8 benchmarks	top GPQA Diamond at 96.0 percent

What an orchestration model is

Fugu treats a multi agent system as if it were a single model. You call one endpoint; behind it, a coordinator decides which specialists in a pool should handle which parts of the request and how their contributions combine. Sakana grounds this in two ICLR 2026 papers: TRINITY, which uses an evolved coordinator to route across models, and Conductor, which applies reinforcement learning to discover coordination strategies expressed in natural language. The pitch is that you can beat a single frontier model by being smart about which cheaper models to use and when, rather than by training one bigger model.

The benchmark evidence is real and worth stating precisely. On Fugu Ultra v2, Sakana reports best or joint-best results on five of eight benchmarks and a top-two placement on seven of eight, including 74.3 percent on DeepSWE and 48.3 percent on Chartography, where it clears Claude Opus 5 at 27.3 and Claude Fable 5 at 29.5 by a wide margin. The detail that makes the claim interesting: Fugu Ultra v2 excludes GPT-6 Astra and Claude Fable 5.1 from the pool it orchestrates, and still posts those numbers. It is beating frontier scores without frontier models in the mix.

The cost structure is different in kind

Astra has one price: 10 dollars per million input tokens and 50 per million output, flat across its context window. Fugu Ultra v2 lists 5 and 30 dollars per million at standard context up to 272K tokens, rising to 10 and 45 above that, with cached input moving from 0.50 to 1.00 dollars. On output tokens, the part of the bill that dominates for generation heavy work, Fugu Ultra v2 at 30 is 40 percent under Astra at 50 within the standard context band.

But the more important point is that Fugu's cost is a function of how the orchestrator uses the pool, not a fixed per token rate on one model. Sakana also shipped Fugu Max, a cheaper sibling at 2 and 6 dollars per million tokens, with output pricing it says runs 40 to 60 percent below Sonnet 5, GPT 5.6 Terra, and Kimi K3. The orchestration approach turns model cost into a routing decision, which is exactly the lever you want if your workload has a mix of easy and hard subtasks. We took the open versus closed cost question apart with a single model comparison in Kimi K3 versus GPT-6 Astra on open weights, GPQA, and cost, and orchestration adds a second axis on top: not just which model, but how many and in what combination.

Where Astra earns its price

A single frontier model has advantages an orchestrator cannot easily match. Latency is predictable, because there is no coordinator deciding routes and no round trips between specialists. Behavior is easier to reason about and to debug, because there is one model with one failure mode rather than a system whose output depends on how the coordinator split the work. And on the cleanest cross model comparison available, GPQA Diamond, Astra holds the top score at 96.0 percent, the highest abstract reasoning result OpenAI has published. For a task that is genuinely one hard reasoning problem rather than a decomposable pipeline, the monolith is the simpler and often the better tool. We covered where Astra sits against its own predecessor in GPT-6 Astra versus GPT-5.6 Sol on whether the upgrade is worth it.

Orchestration also adds operational surface. A coordinator that routes across a pool is another component to monitor, version, and secure, and its quality gates on how well the coordination generalizes to your tasks rather than the benchmark suite. Sakana's numbers are strong on their eight benchmarks; whether that transfers to your specific workload is an empirical question you have to answer with your own evaluation harness before you build on it.

What this means for how you build

The Fugu release is a signal more than a single product decision. It says the frontier is not only moving through bigger models but through better coordination of the models we already have, and that is directly relevant if you are building agent systems. The patterns Sakana is productizing, learned routing across a pool of specialists, are the same patterns that show up when you hand build a multi agent workflow. We wrote up the hand built version of this in multi agent orchestration patterns, and Fugu is essentially those patterns trained rather than coded. If you already run a multi model setup with a router in front, Fugu is a preview of where that architecture is heading.

FAQ

What is Fugu Ultra v2? An orchestration model from Sakana AI, released September 11, 2026. It coordinates a pool of specialized language models on a task rather than generating answers as a single model, and reports best or joint-best on five of eight benchmarks.

Is Fugu Ultra v2 better than GPT-6 Astra? On Sakana's benchmark suite it posts leading numbers including 74.3 on DeepSWE, without using Astra or Fable 5.1 in its pool. Astra holds the top GPQA Diamond score at 96.0 percent and offers predictable single model latency. They suit different task shapes.

How much does Fugu Ultra v2 cost? 5 dollars per million input tokens and 30 per million output up to 272K context, rising to 10 and 45 above that. The cheaper Fugu Max sibling is 2 and 6 dollars per million.

When should I pick a single model over an orchestrator? When the task is one hard reasoning problem, when latency must be predictable, or when you need a simple failure mode to debug. Orchestration wins when the work decomposes into a mix of easy and hard subtasks you can route cheaply.

When this applies to your stack

If your workload is a mix of easy and hard subtasks, orchestration is the architecture that lets you spend frontier money only where it matters and cheap money everywhere else. Fugu productizes that as a hosted model, but the same logic applies to a router you build yourself over a pool of models you already pay for. The hard part is not the routing code, it is the evaluation harness that tells you whether a given split actually improves quality and cost on your tasks rather than on someone's benchmark. Standing up that harness, and the orchestration layer it validates, is the kind of AI infrastructure and agent engineering work we do at Contra Collective.