Fugu Ultra v2 vs Kimi K3: due router, due altitudini

Both of these systems are routers, and that is the first thing most coverage misses. Kimi K3 is a 2.8-trillion-parameter Mixture-of-Experts model that activates roughly 104 billion parameters per token — which means that on every single token it generates, it makes a routing decision, selecting 16 of its 896 experts to do the work. Fugu Ultra v2, announced by Sakana AI on 11 September 2026, is a router at a different altitude entirely: it takes an incoming task, breaks it up, and dispatches the pieces across a pool of separate models. One routes inside a forward pass; the other routes across a fleet. Understanding that they are the same idea at two scales is the fastest way to see why their price tags differ by 2× on input and 5× on output.

The two routers, side by side

• Kimi K3 — Moonshot AI's flagship, released mid-July 2026 with open weights published on 27 July 2026. The architecture is unusual even by 2026 standards: 69 Kimi Delta Attention layers interleaved with 24 Gated MLA layers, Attention Residuals, a "Stable LatentMoE" design, and a 401M-parameter MoonViT-V2 vision encoder. Weights ship in MXFP4 with MXFP8 activations under quantisation-aware training. Its hosted API exposes reasoning effort with exactly one supported value — max.

• Fugu Ultra v2 — Sakana AI's peak-capability orchestration tier, and not a foundation model in the usual sense. It is a single OpenAI-compatible endpoint that decomposes tasks and dispatches them across a pool of open-weights and specialised models. Sakana states that Claude Fable 5, Claude Fable 5.1 and GPT-6 Astra are excluded from that pool, and lists a training cutoff of 20260828. There is no published parameter count, because there are no parameters to count in the way you would for K3 — the capability lives in the scheduling policy.

The consequence is that K3 is a component and Fugu Ultra v2 is an assembly. That makes this comparison unusual: these two are not only competitors, they are potential ingredients. A model like K3 is exactly the kind of open-weights, long-context, tool-calling system an orchestrator would plausibly rent. Sakana does not publish the pool's membership, so whether K3 is inside Fugu Ultra v2 is unknown — but if it is, some of what you pay Fugu's rate for is K3's tokens with a markup.

What the price gap actually is

• Input — Fugu Ultra v2 at a reported $5.00 per 1M (third-party listings; Sakana's announcement page does not publish its own rates) vs Kimi K3 at $3.00 per 1M, with cache hits at $0.30.

• Output — Fugu Ultra v2 at a reported $30.00 per 1M vs Kimi K3 at $15.00 per 1M. Half the price, for a model you can also download.

• Cached input — Fugu Ultra v2 $0.50 per 1M vs Kimi K3 $0.30 per 1M on a cache hit. In a long agent loop with a stable system prompt, that spread compounds on every turn.

• Context tiering — Kimi K3 is flat across its entire 1,048,576-token window with no long-context penalty. Fugu Ultra v2's reported tier above roughly 272,000 input tokens moves the whole request to about $10.00 / $45.00.

That last row is the one that decides real bills. A long-horizon agent run spends most of its life past 272K tokens, which is precisely where K3's flat rate stays flat and Fugu Ultra v2's doubles. If your workload looks like that, the effective gap on the input side is closer to 3.3× than 1.7×.

The one number they share

Both vendors report DeepSWE, and the comparison is less flattering to Fugu than its launch framing suggests. Sakana lists Fugu Ultra v2 at 74.3. Moonshot lists Kimi K3 at 67.5 — vendor-reported, from the model card. That is a 6.8-point gap on a coding-agent benchmark, which is real, but it is one test out of a much larger published set, and it is the same test that Sakana leads on by 1.6 points against GPT-5.6 Sol. A benchmark where three different vendors all cluster within seven points is measuring something real but is not separating them decisively.

Past that, the two publish on completely different shelves. Moonshot's K3 numbers include Terminal-Bench 2.1 at 88.3, GPQA Diamond at 93.5, HLE-Full at 43.5 (56.0 with tools), SWE-Marathon at 42.0, OSWorld-Verified at 84.8 and MCPMark-Verified at 94.5 — all vendor-reported, all on the model card. Sakana's eight for Fugu Ultra v2 include two internal benchmarks, SWEFish and Toolathon, which nobody outside Sakana can reproduce.

Independently, only one of them has been measured at all. Kimi K3 scores 44 on the Artificial Analysis Intelligence Index v4.3 — second among open-weight models, behind GLM-5.3 at 45 — and 93.40% on SWE-bench Verified under Vals AI's standardised agentic harness, behind Claude Opus 5 and DeepSeek V4 Pro but close. It also leads the Frontend Code Arena at 1679 Elo, ahead of Fable 5 at 1631 and GPT-5.6 Sol at 1618. If you have seen K3 quoted at 57 or 60, that is the superseded index scale and it should not be repeated.

Fugu Ultra v2 has no independent score of any kind. It was announced on 11 September 2026 and nobody outside Sakana has run it.

Speed: one measured, one not

Kimi K3 is slow, and this is measured rather than claimed: Artificial Analysis records 37.9 tokens per second with a time-to-first-token of 3.80 seconds, and flags it as notably verbose. For a model built around long-horizon agentic coding, throughput is a real constraint — a slow model in a long loop costs wall-clock, not just money.

Sakana publishes no latency or throughput figure for Fugu Ultra v2 at all. For an orchestrator that is arguably honest rather than evasive, because response time depends entirely on which models it decides to call and how many passes it makes. But it also means you cannot model the wall-clock cost of switching to it without measuring it yourself. For the previous Fugu Ultra, testers publicly reported runs stretching toward 30 minutes; that is the June-2026 model and not evidence about v2, but it is the number to beat when you benchmark.

Open weights, with a catch worth reading

Kimi K3's weights are downloadable, which is a genuine difference from Fugu Ultra v2's hosted-only model — but the licence is narrower than the phrase "open weights" usually implies. K3 ships under the Kimi K3 License, not an OSI-approved MIT or Apache grant, and the widely repeated claim that it is "modified MIT" is disputed. The terms require a separate agreement with Moonshot for model-as-a-service businesses above $20M in revenue over any twelve consecutive months, plus attribution for products above 100M monthly users or $20M monthly revenue. Internal use is exempt.

So the honest framing is: K3 gives you weights you can hold, inspect, fine-tune and self-host, subject to a revenue-gated commercial condition. Fugu Ultra v2 gives you none of that — no weights, no inspectable pool, no self-hosting — but also imposes no revenue condition, because there is nothing to license. Which of those is more permissive depends entirely on whether you are the kind of company the K3 licence is aimed at.

One practical caveat if self-hosting is the plan: K3's checkpoint is roughly 1.5 terabytes and the vendor recommends 64 or more accelerators. "Open weights" here means an inference-cluster decision, not a laptop one. For most teams the API is the sane path either way — which is why the day-0 vLLM and SGLang support matters more than the download does.

What our own traffic says

One data point that neither vendor publishes, and which is worth more than it looks: across the OrcaRouter gateway, Kimi K3 is the most heavily used model of any discussed in this article by a wide margin, moving roughly 3.27 billion tokens in the last seven days.

That is not a benchmark and it does not settle anything about quality. But it is a large sample of what developers actually choose when the decision costs them nothing but the token price, and it says that K3's combination of a 1M flat-rate window, open weights and strong coding-arena placement has already won a substantial share of real long-context agentic work. Fugu Ultra v2 has no comparable signal, because it launched today.

Getting to each of them

Kimi K3 is on OrcaRouter at Moonshot's own rate — $3.00 per million input, $0.30 on a cache hit, $15.00 per million output — passed through with zero markup, so a vendor price change lands here the same day rather than on a migration schedule. It is OpenAI-compatible on the same base URL as the rest of the catalogue, and because it sits behind the same gateway as everything else, you can put it in a failover chain or route to it conditionally rather than hard-coding a single provider into a production path. That matters more than usual here: a 2.8T-parameter model served at 37.9 tokens per second is exactly the kind of dependency worth having a fallback behind.

Fugu Ultra v2 is not routed by us. It is available from Sakana AI's own OpenAI-compatible API, and the company says existing Fugu integrations move to it with a single parameter change. Both models speak the same request format, so an evaluation that puts them behind one flag is a base-URL swap rather than a rewrite.

Quale scegliere

Kimi K3 is the more defensible buy for almost every workload. It is half Fugu Ultra v2's price on both input and output, flat across a 1M window with no long-context cliff, independently scored at 44 on the current Artificial Analysis index, self-hostable under a revenue-gated licence, and already the highest-traffic model in this comparison by a wide margin. Its costs are speed and verbosity: 37.9 tokens per second is genuinely slow for agentic loops, and the licence has teeth if you are a large model-as-a-service business.

Fugu Ultra v2 is the buy if you want the specific property Sakana is selling — a capability tier that decomposes hard multi-step work across a pool that structurally excludes the frontier vendors, so that no single upstream model can be revoked, repriced, or cut off underneath you. Its DeepSWE edge over K3 is real and vendor-reported; its top-two placement on seven of eight benchmarks is vendor-reported too, and on tests that partly do not exist anywhere else.

The thing to notice is that both of these are routing decisions, and you are being asked to pick an altitude. K3 routes tokens to experts inside a model you can download and control. Fugu Ultra v2 routes tasks to models you cannot see. The second is a bigger, more powerful idea. It is also the one you can only take on faith — and it costs five times as much per token to take that faith.