Step 5 Preview vs Fugu Ultra v2: One of These Is a Model, the Other Is a Router

The comparison most people want between Step 5 Preview and Sakana AI's Fugu Ultra v2 is a benchmark table. Neither side can supply it honestly. Step 5 Preview has one independent measurement, taken under a rubric that was replaced last week. Fugu Ultra v2 has no independent measurement at all, because the evaluators that publish an intelligence index have never put a page up for it. What the two do have is a price ratio of exactly five to one on input tokens and a difference in kind that makes the ratio almost meaningless: one is a 600-billion-parameter mixture-of-experts checkpoint you can download and run on your own hardware, and the other is an orchestration layer that answers your request by dispatching it to a pool of other people's models.

That difference is not a detail. It decides what you are actually buying, what happens to your results when the vendor changes its mind, and whether a long-horizon agent built on top of either of them can be re-run next quarter for the same output.

Two products carrying the same marketing label

StepFun shipped Step 5 Preview to its API on 20 September 2026, then published BF16 and FP8 weights on Hugging Face and ModelScope on 15 October. It is a single model: a sparse mixture of experts with 600 billion total parameters and 27 billion active per token, routed 8-of-352 plus a shared expert, in a 92-layer stack the vendor describes as "narrow-deep". The context window is 1M tokens, the input can be text, images or video, and the output is text.

Fugu Ultra v2 is a point release, not a new architecture — Sakana AI shipped it on 11 September 2026 as a refresh of Fugu Ultra from 22 June, with a model ID of fugu-ultra-v2.0 and a single-line parameter migration from the prior version. It is a conductor: a request goes in, the system decides which model or models in its pool should handle it, runs them, and synthesises the answer. The pool deliberately excludes Claude Fable 5, Claude Fable 5.1 and GPT-6 Astra. Sakana's argument for the smaller pool is that weaker members were dragging the aggregate down; the second-order effect, which Sakana does not advertise, is that the composition of the thing you are buying is a vendor decision that can change without a version number.

The spec lines, both sides

• What it is — Step 5 Preview: one downloadable 600B/27B mixture-of-experts model. Fugu Ultra v2: an orchestrated pool of models with a synthesising layer on top.

• Release date — Step 5 Preview: API on 20 September 2026, weights on 15 October 2026. Fugu Ultra v2: 11 September 2026, a refresh of a June 2026 original.

• Context window — Step 5 Preview: 1M tokens. Fugu Ultra v2: 1M tokens, with a price step above 272k tokens.

• Input modalities — Step 5 Preview: text, images and video. Fugu Ultra v2: whatever the members of its pool accept, which Sakana does not enumerate per-modality.

• Input price per million tokens — Step 5 Preview: $1.00. Fugu Ultra v2: $5.00, rising to $10.00 above 272k tokens of context.

• Output price per million tokens — Step 5 Preview: $2.70. Fugu Ultra v2: $30.00, rising to $45.00 above 272k tokens.

• Cached input per million tokens — Step 5 Preview: a 90% discount on $1.00. Fugu Ultra v2: $0.50, rising to $1.00 above 272k tokens.

• Independent score — Step 5 Preview: 29 on Artificial Analysis Intelligence Index v5.0. Fugu Ultra v2: no Artificial Analysis page exists, so no score.

• Weights — Step 5 Preview: downloadable under StepFun Model License 1.0. Fugu Ultra v2: none. The pool it draws on is other vendors' weights supplied under other vendors' terms.

What five dollars of input actually buys

The headline ratio is five to one on input and just over eleven to one on output, and both numbers flatter Step 5 Preview in a way that the per-task bill may not confirm. Fugu Ultra v2 is priced like a system that expects to spend tokens: it dispatches a request, several models answer, and a synthesiser writes over the results. Sakana's own pitch that adding more agents does not multiply your bill is about the subscription tiers — $20 Standard, $100 Pro and $200 Max per month, all of which include Fugu, Fugu Ultra and Fugu Max — and not about the metered rate card. On the metered card, orchestration is a verbose business, and output is billed at $30 per million.

Step 5 Preview's economics run the other way. Artificial Analysis measures 179.1 output tokens per second on StepFun's API against a class median of 74.1, $2.35 per full index task, and a blended $0.54 per million tokens at a seven-to-two-to-one cache-hit ratio. Against a class median of $1.40 per task on the same page, that is a cheap model that is also fast.

So the honest framing is not "Fugu is five times the price". It is that Fugu charges a premium for a capability Step 5 Preview does not claim — arbitrating between several models per request — and that whether the premium is worth paying depends entirely on whether you believe that arbitration is better than a single strong model at doing the job you have. Sakana's published evidence for that belief is vendor-run: Chartography 48.3 against 27.3 for Claude Opus 5 and 29.5 for Claude Fable 5, DeepSWE 74.3, best or joint-best on five of eight benchmarks and top-two on seven of eight. Those are real numbers produced by the company selling the product, and they are the only numbers there are.

The portability test, which is the real difference

Here is the question that separates the two more cleanly than any score. Suppose you build an agent on Fugu Ultra v2 in September 2026, and it works. In November, Sakana adds a model to the pool, or removes one — it has already removed three frontier models once — or reweights the synthesiser. Your prompts are the same, your inputs are the same, and your outputs are different, with no version bump you can pin and no changelog entry that tells you which behaviour moved. You are not buying a model; you are buying a policy over models, and the policy is a dial the vendor turns.

That is not a criticism of Sakana. It is the nature of the product, and for plenty of workloads it is exactly what you want — a vendor whose full-time job is picking the right model per request will often beat your hand-rolled choice. But it should be priced as what it is: an ongoing service relationship, not an asset.

Step 5 Preview is the other shape. The weights exist. They are not open in the OSI sense — StepFun Model License 1.0 keeps an attribution condition on materially used outputs, requires hardware makers to negotiate a separate licence, and requires a commercial licence for model-as-a-service operators above $20 million in annual revenue — but the behaviour you test today is the behaviour you can run in a year, on hardware you control, without asking permission. The FP8 checkpoint needs about 773 GB of VRAM and a minimum of four GPUs with B300 as the default target; the Hopper path needs eight H200s. The NVFP4 variant cuts that to roughly 473 GB and caps the usable context at 320k tokens. Neither path is cheap, and neither is as cheap as the API, which is the usual argument for calling a model rather than hosting it.

The disclosure gap runs both ways

Each side is missing something the other publishes.

Step 5 Preview publishes a rate card, a context limit, a parameter count, a licence, serving recipes for vLLM and SGLang on the day the weights landed, and one independent score. What it does not publish is a harness-matched head-to-head against any competitor, and its vendor benchmark table — Terminal-Bench 4.0 at 33.3, DeepSWE v1.1 at 67.7, FrontierFinance at 66.4, CyberGym at 84.7 — is exactly that, a vendor table.

Fugu Ultra v2 publishes an architecture story, a pool policy, a benchmark table of its own, and subscription pricing that is genuinely unusual in this market. What it does not publish is a single number that an independent evaluator has verified, and the evaluators have not built a page for it, which is the cleanest possible way of saying the score does not exist rather than that it is low.

When both sides are vendor-reported, the tiebreak is not "which vendor is more trustworthy". It is which vendor has staked something that can be checked. StepFun staking the weights on 15 October is a checkable claim: anyone can download the checkpoint and reproduce the behaviour. Sakana staking a score without a downloadable artefact is not.

Where either one is callable today

Neither is on OrcaRouter. The live catalogue carries 214 model IDs across 16 providers behind one key and one bill, and as of 29 October 2026 there is no StepFun entry and no Sakana entry in it — no string "step" and no string "fugu". If you want Fugu Ultra v2 you go to Sakana, and if you want Step 5 Preview you go to StepFun's API or to the checkpoints. Nothing in this article should be read as a claim that we serve either.

What the platform is for, in this comparison, is the step you take after: once you have decided which of these two shapes fits your workload, the models you actually route per request — the cheap one for classification, the expensive one for synthesis — are the part that scales badly by hand. OrcaRouter's routing configuration lets you declare that as a primary and a fallback chain rather than a hard-coded model ID, with automatic failover when a provider degrades, and it bills at provider list price with no markup, so a vendor price cut is live on the platform the same day it is announced. That is a property of the models we do carry, not of these two.

The question that decides it

Not "which scores higher", because only one of them has been scored, and the score is 29 under a rubric that landed on 27 October and rewrote what the composite measures. Ask instead whether your workload is one where a single 1M-context model with vision input, at $1.00 and $2.70 per million tokens, is good enough — and if it is, Step 5 Preview is the cheaper and more portable buy by a wide margin. If your workload is one where routing per request is the hard part, Fugu Ultra v2 is doing labour you would otherwise do yourself, and the five-to-one ratio is the price of that labour rather than a markup on the same good.

What you should not do is compare the Chartography 48.3 against anything StepFun has published. They are two vendors' numbers, on two vendors' harnesses, for two different kinds of product. Nobody has run them against each other, and until somebody does, the only honest comparison between these two is the one about what they are.