Fugu Ultra v2 so với MiniMax M3: Giá đầu ra gấp hai mươi lăm lần

Set the two output rates next to each other and the number looks like a typo. MiniMax M3 bills $1.20 per million output tokens. Fugu Ultra v2, announced by Sakana AI on 11 September 2026, is reported at $30.00. That is 25× — and on input the gap is 16.7×, $0.30 against a reported $5.00. MiniMax M3 is also faster, accepts video natively, and ships as downloadable open weights. Faced with that table, the reasonable first reaction is that the orchestration tier cannot possibly justify its price. The more interesting question is what it would have to be true for it to.

The gap, stated plainly

• Input — MiniMax M3 $0.30 per 1M vs Fugu Ultra v2 a reported $5.00 per 1M. The M3 rate is stated by Min​iMax as a permanent 50% discount, not a promotion with an expiry.

• Output — MiniMax M3 $1.20 per 1M vs Fugu Ultra v2 a reported $30.00 per 1M.

• Cached input — MiniMax M3 $0.06 per 1M vs Fugu Ultra v2 a reported $0.50 per 1M. M3's cache read is a fifth of its own input price, which makes repeated agent loops over a stable prompt unusually cheap.

• Above 512K input — MiniMax M3 doubles to $0.60 / $2.40 / $0.12. Fugu Ultra v2's reported tier above roughly 272K moves to about $10.00 / $45.00 / $1.00. Both reprice the whole request; M3 starts later and lands at a fraction of the rate.

• Context — MiniMax M3 1,048,576 tokens with a vendor-guaranteed minimum of 512K. Fugu Ultra v2's context is reported at 1M by third-party listings; Sakana's announcement page states no figure.

One caveat on the Fugu column, and it is not a small one: Sakana does not publish Fugu Ultra v2's token rates on its own announcement page. The $5.00 / $0.50 / $30.00 figures come from third-party model listings and should be treated as unconfirmed until you check the live rate card. The M3 column, by contrast, is documented by the vendor. That difference in provenance is itself part of the comparison.

What M3 does natively that an orchestrator has to route

MiniMax M3 accepts text, image and video input and returns text. Video is the notable one. The model was trained on mixed modalities from step zero, and it is built on Min​iMax Sparse Attention, an architecture the vendor claims delivers up to 9× faster prefill and 15× faster decode at 1M context compared with its predecessor, at roughly a twentieth of the per-token compute.

That matters in this matchup because it is a capability Fugu Ultra v2 does not claim for itself. An orchestrator's answer to "handle this video" is to route the video to a model that can see it, which means a sub-call, which means more tokens billed at the orchestrator's rate. M3's answer is to just read it. At $0.30 input, an hour of video analysis is a rounding error; at $5.00 input plus whatever the routing costs, it is a budget line.

The architecture comparison is lopsided in a different way too. MiniMax M3 is roughly 428B total parameters with about 23B active per token — a Mixture-of-Experts design where the sparse routing keeps serving cost low. Fugu Ultra v2 has no published parameter count, because its capability sits in a scheduling policy over other companies' models rather than in weights Sakana trained. One of these is a model. The other is a decision procedure.

The uncomfortable implication

Here is the thing worth thinking through before paying 25× for output.

MiniMax M3 is open-weight, cheap, fast, multimodal and long-context. It is, on paper, an excellent sub-agent. If you were building an orchestration layer from scratch — the thing Fugu Ultra v2 is — a model like M3 is close to the ideal component: cheap enough to call speculatively, capable enough to handle the bulk of decomposed subtasks, and able to see images and video without a second hop.

Sakana does not publish Fugu Ultra v2's model pool, so we cannot say whether M3 is in it. What we can say is that the economics only make sense one of two ways. Either Fugu Ultra v2 is not calling models like M3, in which case it is leaving the cheapest capable capacity on the table. Or it is calling them, in which case a meaningful share of the tokens you are billed at $5.00 and $30.00 are M3 tokens that cost $0.30 and $1.20 upstream, and the premium is the orchestration policy plus a margin.

Neither reading is fatal to the product — a good scheduler can be worth a large multiple, and paying for one avoids you having to build and maintain it. But it does mean the question to ask about Fugu Ultra v2 is not "is it better than M3." It is "is the scheduling worth more than the tokens." And that is a question only your own task measurements can answer.

There is no shared benchmark

Unlike some other matchups in this series, there is not a single benchmark on which both vendors publish a figure.

Min​iMax's published set for M3 is broad and mostly reproducible in shape: SWE-bench Verified at 80.5%, SWE-bench Pro at 59.0%, Terminal-Bench 2.0/2.1 at 66, BrowseComp at 83.5%, OSWorld-Verified at 70.1%, and PostTrainBench at 37.1 — third place, behind Claude Opus 4.7 at 42.4 and GPT-5.5 at 39.3. All vendor-reported. Independently, Artificial Analysis scores MiniMax M3 at 30 on its v4.3 Intelligence Index, ranked #16 of 113, with a throughput of 95.7 tokens per second and a time-to-first-token of 1.95 seconds — fast, and among the better-served models in its class. If you have seen M3 quoted at 44 or 45, that predates the index restatement.

Sakana's eight benchmarks for Fugu Ultra v2 — GDP.pdf, Chartography, SWEFish, DeepSWE and Toolathon among them, with best-or-joint-best on five and top-two on seven — do not intersect with that list at all. Two of them, SWEFish and Toolathon, are internal Sakana tests. The headline claims, Chartography at 48.3 against Opus 5's 27.3 and Fable 5's 29.5, and DeepSWE at 74.3, are vendor-reported and were published today. Nothing about Fugu Ultra v2 has been independently verified.

So the comparison every aggregator will produce in the next two weeks — a table of shared scores — is not available. The only honest thing to say about capability is that MiniMax M3 has a measured, independent position and Fugu Ultra v2 has a claim.

Speed versus configuration

MiniMax M3's numbers are known: 95.7 tokens per second, 1.95 seconds to first token. That is roughly two and a half times the throughput of Kimi K3 and among the fastest frontier-adjacent models measured, which fits a design explicitly optimised for cheap long-context serving.

Fugu Ultra v2 publishes no latency or throughput figure. As with the pricing, part of that is structural — an orchestrator's response time depends on which models it picks and how many passes it takes, so a single number would be misleading. But it means the wall-clock cost of adopting it is unmeasured and unmeasurable from the outside. For the previous Fugu Ultra, testers publicly reported runs stretching toward 30 minutes; that is the June-2026 model rather than v2, and it is not evidence about the new one — but if you are spending 25× on output, knowing how long you wait is not a nice-to-have.

The licence, and what the Arabic run says

MiniMax M3's weights are downloadable under the Min​iMax Community License, which is open weights with commercial conditions attached — notably a revenue threshold above which written authorisation is required. It is not MIT, and any page telling you it is has not read the licence tag. Fugu Ultra v2 gives you nothing to license because there is nothing to download.

The more interesting recent development for M3 is what other people are doing with those weights. On 3 September 2026, HUMAIN — the Saudi PIF-backed AI company — released an Arabic flagship built by post-training MiniMax M3 on more than a trillion Arabic tokens, at 428B total and 23B active parameters, reporting an equal-weighted 89.37% across seven Arabic benchmarks and placing it ahead of GPT-5.6 Sol at 87.30 and Claude Opus 5 at 87.34. It is a research preview, and those are HUMAIN's own numbers.

The contrast with Fugu Ultra v2 could not be sharper as a statement about what each product is. MiniMax M3 is a base other companies build sovereign models on top of. Fugu Ultra v2 is a gate you call through, whose composition you cannot see and whose weights you cannot hold. Both are legitimate positions. They are not remotely the same kind of thing.

Calling either of them

MiniMax M3 is on OrcaRouter at Min​iMax's own rate — $0.30 per million input, $0.06 on a cache hit, $1.20 per million output — passed through with zero markup, so if Min​iMax changes the discount structure it lands here the same day rather than on a migration cycle. It is OpenAI-compatible on the same base URL as the rest of the catalogue, which means you can route to it by rule, put it behind a failover chain, or include it in a model-fusion panel without a second contract or a code change. For a model whose entire argument is price, that pass-through is the part that matters: the cheapest route stays cheapest.

Fugu Ultra v2 is not routed by us. It is available from Sakana AI's own OpenAI-compatible API, and the company says an existing Fugu integration moves onto it with a one-line parameter change. Both speak the same request format, so putting them behind a single flag for evaluation is a base-URL swap.

Bản án

For the overwhelming majority of workloads, MiniMax M3 wins this comparison on the arithmetic alone. It is 16.7× cheaper on input and 25× cheaper on output, faster by every measured metric, natively multimodal including video, open-weight under a licence most companies will never hit the ceiling of, and independently scored at 30 on the current Artificial Analysis index. Its weaknesses are real but narrow: it is verbose, its agentic category score is middling relative to its coding scores, and its vendor benchmark set is strong on software engineering and thinner on long-horizon autonomy.

Choose Fugu Ultra v2 only if you believe the thing Sakana is actually selling — a peak-capability tier that decomposes hard multi-step tasks across a pool no frontier vendor controls, reaching frontier-level output without depending on frontier models — and you have task-level measurements showing that decomposition beats a single cheap model on your work. That is a genuinely different proposition from anything Min​iMax offers, and it is the reason the 25× exists.

But do the measurement before the purchase, not after. If a $0.30 model finishes your task in two attempts and Fugu Ultra v2 finishes it in one, Fugu is still more expensive. The only workload where 25× is a bargain is one where the cheap model cannot finish at all — and that is a much smaller set of tasks than the launch copy implies.