Fugu Ultra v2 vs GLM 5.2: You Pay for Coordination, Not Parameters

Put the two rate cards side by side and the matchup looks like a mistake. Fugu Ultra v2, which Sakana AI announced on 11 September 2026, bills $5 per million input tokens and $30 per million output tokens. GLM 5.2, Z.ai's flagship from 16 June 2026, bills $1.40 and $4.40. Both claim a context window in the million-token range. Both are pitched at exactly the same buyer — the team running long, multi-step, repository-scale agentic work. One is roughly 3.6× the input price and 6.8× the output price of the other, and the cheaper one is the one you can download and keep. So the whole comparison reduces to a single question: what is the extra money actually buying, and is it buying something a single model cannot be?

They are not the same kind of object

The reason the price gap exists is that these two things solve long-horizon tasks with completely different machinery, and the difference shows up before you ever run a benchmark.

• Fugu Ultra v2 — an orchestration system, not a foundation model. It is the capability-maximising tier of Sakana AI's Fugu line: a single OpenAI-compatible endpoint that decomposes an incoming task and dispatches the pieces across a pool of other models, some open-weights and some specialised. Sakana's own framing is that it "pushes peak performance higher than ever before, without the indispensable reliance on the frontier models it orchestrates." You are buying a scheduler, and you pay for every token that scheduler spends thinking, including the internal orchestration tokens.

• GLM 5.2 — a single Mixture-of-Experts model. Z.ai publishes it as a text-in / text-out model with a "truly usable" 1M-token context window and up to 128K output tokens, hybrid reasoning controlled by reasoning_effort, and native tool calling. Reported architecture is roughly 753B total parameters with about 40B active per token, though Z.ai has not confirmed the active count for 5.2 specifically — treat that figure as carried over from GLM-5.1.

That is the fork in the road. One of these is a thing you call; the other is a thing that calls things.

What Fugu Ultra v2 would not tell us

Sakana's announcement page is unusually candid about one thing and unusually quiet about another, and both matter for a buying decision.

The candid part: the company states outright that Claude Fable 5, Claude Fable 5.1 and GPT-6 Astra are not in Fugu Ultra v2's model pool, even while claiming to beat them on benchmarks. The stated rationale is resilience — a swappable pool of open and specialised models that cannot be revoked, repriced, or cut off by a vendor or a geopolitical shift. The training cutoff is listed as 20260828. Whatever you think of the argument, it is a real architectural stance, and it is the clearest reason to pick Fugu Ultra v2 over a stack you assembled yourself.

The quiet part: the announcement does not state a context window, and it does not state Fugu Ultra v2's token pricing on the page itself. Third-party model listings put the window at 1M tokens and the rates at $5 / $0.50 cached / $30, with a repricing tier above roughly 272K input tokens that moves to about $10 / $1.00 / $45. Those are the numbers the rest of this article uses, but they are listings, not vendor documentation — confirm them against Sakana's live rate card before you budget on them.

The cost question nobody answers per token

Per-token pricing is the wrong unit for this comparison, and every page currently ranking for it makes the same error. An orchestrator's bill is not the size of your prompt multiplied by a rate; it is the size of your prompt, plus every sub-call it decides to make, plus the reasoning tokens of the models it rents, all marked up at the orchestrator's own rate.

Sakana acknowledges this directly: orchestration tokens consumed internally are billed as ordinary input and output tokens. Which means the honest way to compare Fugu Ultra v2 with GLM 5.2 is cost per completed task, and neither vendor publishes that figure.

Some structural points that do hold regardless:

• Input price — Fugu Ultra v2 $5.00 / 1M vs GLM 5.2 $1.40 / 1M on OrcaRouter at Z.ai's provider rate. Fugu is 3.6× the cost before it makes a single sub-call.

• Output price — Fugu Ultra v2 $30.00 / 1M vs GLM 5.2 $4.40 / 1M. This is the number that hurts, because agents emit a lot of output, and an orchestrator emits output that you did not ask for.

• Cached input — Fugu Ultra v2 lists $0.50 / 1M; GLM 5.2 caches at $0.26 / 1M, an 81% discount off its own input rate. Repeated agent loops over a stable system prompt are where GLM 5.2's advantage compounds hardest.

• Long-context repricing — Fugu Ultra v2's reported tier above ~272K input moves the whole request to roughly double the input rate and 1.5× the output rate. GLM 5.2 has no context-tiering at all: $1.40 / $4.40 flat, all the way out to 1M.

That last row is the one to circle. A long-horizon agent run spends most of its life beyond 272K tokens. GLM 5.2's flat rate is worth real money exactly where Fugu Ultra v2's doubles.

The benchmarks barely touch

Here is the genuinely strange thing about this matchup, and the thing no current ranking page says plainly: these two models are almost never measured on the same test.

Sakana reports Fugu Ultra v2 as best or joint-best on five of eight benchmarks — GDP.pdf, Chartography, SWEFish, DeepSWE and Toolathon — and top-two on seven of eight. The two headline figures it gives are Chartography at 48.3, against Opus 5 at 27.3 and Fable 5 at 29.5, and DeepSWE at 74.3. SWEFish and Toolathon are Sakana's own internal benchmarks. All of it is vendor-reported and none of it has been independently reproduced as of publication.

Z.ai's numbers for GLM 5.2 come from a different shelf entirely: Terminal-Bench 2.1 at 81.0, SWE-bench Pro at 62.1, GPQA-Diamond at 91.2, AIME 2026 at 99.2, HLE at 40.5 — also vendor-reported. On the independent side, GLM 5.2 carries an AA Coding Index of 68.8 and an Intelligence Index of 39 on Artificial Analysis's re-scored page, where it is flagged as a provisional estimate, ranked #7 of 113. Older citations of 51 or 53 for GLM 5.2 come from a superseded index scale and should not be quoted.

So if you want a clean head-to-head number, there isn't one. The closest thing to a shared axis is that both are strong on agentic coding, and the honest reading is that each company has published on the tests it does well. That is a reason to run your own evaluation, not to pick the higher number.

One thing worth flagging for anyone shopping GLM 5.2 today: it is no longer Z.ai's newest model. GLM-5.3 arrived around 14 August 2026 at the same list price, but under a custom licence that restricts large cloud providers — which makes GLM 5.2 the last fully MIT-licensed Z.ai flagship, and the reason it still has a distinct buying case.

The licence is the sharpest divider

Set the benchmarks aside and the structural difference is starker than any score.

GLM 5.2 ships as open weights under MIT on Hugging Face, with no regional restrictions. You can download it, run it on your own hardware, fine-tune it, and ship it inside a product without asking anyone's permission or paying a per-token rate at all. Z.ai trained it — notably, by the company's own claim, entirely on Huawei Ascend accelerators rather than Nvidia.

Fugu Ultra v2 gives you none of that. It is a hosted service: an orchestration policy over a pool of models whose membership Sakana has described only at the edges. You cannot download it, inspect it, pin it, or run it air-gapped. What you get instead is the abstraction — a single endpoint whose behaviour is, in principle, resilient to any one upstream model disappearing. That is a genuine benefit for a production system that cannot afford a dependency to vanish. It is also a genuine cost, because you have traded control for it.

If your constraint is "the model must not change under me," only one of these answers. If your constraint is "the model must not be removable by someone else," only the other one does.

Speed, and what testers said about the last one

GLM 5.2 is measurably not fast: Artificial Analysis recorded roughly 66.3 tokens per second with a time-to-first-token around 4.03 seconds, which is modest for a frontier-class model, and Z.ai's own users have complained about serving congestion during peak hours.

Sakana publishes no latency or throughput figures for Fugu Ultra v2, which is itself informative — a system that dispatches to multiple models has a latency profile that depends entirely on what it decides to call, so a single throughput number would be close to meaningless. For the previous Fugu Ultra, testers publicly reported runs stretching to around 30 minutes and costs far above a single model's rate on the same job. That is the June-2026 model and not evidence about v2 — but it is the failure mode to test for, and the reason a per-token comparison flatters orchestrators.

How you would actually call each of these

GLM 5.2 is on OrcaRouter today at Z.ai's own provider rates — $1.40 per million input and $4.40 per million output, passed through with zero markup, so any cut Z.ai makes lands here the same day. It is OpenAI-compatible, so switching to it is a base-URL and model-ID change in the SDK you already have, and you can put it behind a failover chain or a routing rule rather than betting a production path on one provider.

Fugu Ultra v2 is not something we route. It is available from Sakana AI's own OpenAI-compatible API, and a single parameter change moves an existing Fugu integration onto it. If you are evaluating it against GLM 5.2, the practical arrangement is to keep GLM 5.2 on your existing gateway and call Fugu Ultra v2 directly from Sakana while you decide — the two are both OpenAI-compatible, so they can sit in the same code path behind a flag.

Which one to pick

Choose GLM 5.2 if you want a known quantity at a known price: MIT-licensed weights you could self-host tomorrow, a flat $1.40 / $4.40 with no long-context penalty, a genuinely usable 1M window, and a discount that compounds on cached agent loops. Accept that it is text-only, that it is a generation behind GLM-5.3, and that its independent index score is provisional.

Choose Fugu Ultra v2 if the thing you are buying is resilience plus peak capability on hard multi-step work, and you are willing to pay for coordination by the token and to give up the ability to inspect or self-host what you are calling. The vendor's own claim is that it reaches frontier-level output without depending on frontier models — a real and unusual proposition, and one that, as of today, exactly nobody outside Sakana has been able to verify.

What we would not do is pick between them on the benchmark table. The tests barely overlap, both vendors published on their own strongest ground, and the deciding number — cost per completed long-horizon task — is the one neither company has put on a page.