GPT-6.1 Sol vs Fugu Max: The Same Input Price, and a Model That Does Not Exist in the Usual Sense

Sakana AI's Fugu Max and Open​AI's GPT-6.1 Sol both charge $2.00 per million input tokens. That is where the resemblance in the rate cards starts and stops — Fugu Max is $6.00 per million output against GPT-6.1 Sol's $10.00, and Fugu Max is 40 percent cheaper on output while publishing no benchmark score that a third party has reproduced. The reason those two facts belong in the same sentence is that Fugu Max, released 2026-09-11, is not a model in the sense GPT-6.1 Sol is. It is an orchestrator: one endpoint that picks a model from a pool, calls it, and returns the answer. Comparing its token price to a single checkpoint's token price is comparing two different things that happen to be denominated the same way.

What follows is drawn from Sakana's own launch material for Fugu Max and from Open​AI's model documentation for GPT-6.1 Sol. Where a number is the vendor's own and nobody outside has reproduced it, the sentence carrying it says so. Where there is no number at all, that gets said too.

An architecture, not a checkpoint

Sakana's own framing is unusual enough to quote rather than paraphrase: Fugu Max is an architecture, not a single model. The system holds a pool of models, routes an incoming request to what it judges to be the cheapest capable option, and returns a single answer. There is no Fugu Max checkpoint to download, no fixed parameter count, and no way to know from the outside which model answered a given call — Sakana does not publish a routing trace.

That design has a property worth naming before anything else in this comparison: the capability of the endpoint is not fixed. It changes when the pool changes, when the router's policy changes, and when the models underneath it are upgraded or withdrawn. A benchmark score for a checkpoint describes that checkpoint. A benchmark score for a router describes the router as it was configured on the day it was measured, which is a narrower claim.

GPT-6.1 Sol is the opposite arrangement. It is a single snapshot identifier, gpt-6.1-sol, with no dated variant, a 1,050,000-token context window, 128,000 maximum output, text, image and file input, and five reasoning effort levels with medium as the default. What you measure on Monday is what you get on Friday.

The two rate cards, side by side, with the rows that differ:

• Input — $2.00 per million for both. Identical.

• Output — $6.00 for Fugu Max against $10.00 for GPT-6.1 Sol. Fugu Max is 40 percent cheaper.

• Cached input — Sakana's launch page does not state a cached-input rate for Fugu Max; a $0.25 figure circulates but is not on the page we read, so we are not treating it as confirmed. GPT-6.1 Sol's is $0.10 per million, halved from the $0.20 that GPT-6 Sol charged.

• Long-prompt repricing — GPT-6.1 Sol reprices the whole request to $4.00/$15.00 above 272,000 prompt tokens. Fugu Max publishes no length tier at all.

• The tier above — Sakana also sells Fugu Ultra v2 at $5.00/$30.00 with $0.50 cached input, for callers who want the router to spend more per call.

The claim that does not survive its own arithmetic

Sakana's headline claim for Fugu Max is that it undercuts Claude Sonnet 5, GPT-5.6 Terra and Kimi K3 by 40 to 60 percent. Checked against the list prices of those three, the arithmetic holds — that one is fine.

The second claim is the one to read carefully. Sakana describes Fugu Max as landing "within striking distance of elite models at two to six times lower cost." At the list prices in play, the raw gap between Fugu Max and the three models it names is between 1.7 and 2.5 times, not two to six. A range that wide, quoted alongside a set of models that only spans the narrow end of it, is a marketing range rather than a measured one. It may well be true of some other comparison Sakana has in mind; it is not true of the comparison in the same sentence.

What is not the case is that Sakana has hidden its reasoning. It published ten scatter plots in the launch material. The transparency problem is different: there are no published benchmark scores attached to them, so a reader cannot check the position of any point.

Six wins and one number you can check

Sakana's launch page highlights six benchmarks. Five of them are presented as wins without published figures. The sixth, SWEFish, is Sakana's own benchmark — the company defined the task, and the company reports the result. An in-house benchmark is not worthless; it is a legitimate way to publish a directional signal. It just is not independent evidence, and it is the only number in the set that can be quoted at all.

Against that, GPT-6.1 Sol's independent record is thin but real. Artificial Analysis scores it at 51.8 on its Intelligence Index — rank 10 of 147 models, 93.2nd percentile — with 52.9% on Humanity's Last Exam, 83% on Long-Context Recall, 54.2% on SciCode and 56.06% on Terminal-Bench 4.0. All five are third-party measurements on a model whose weights never move.

One recent development is worth noting because it is the first independent score on the model and it post-dates the earlier coverage on this blog: when GPT-6.1 Sol launched, no independent evaluator had published anything. Artificial Analysis has since published the set above. The direction is unremarkable — the model lands where Open​AI positioned it, mid-frontier, rather than above its own flagship — but it is now measured rather than merely claimed, and that is a different thing from Fugu Max's position today.

What an orchestrator's token price does not include

This is the part of the comparison that a rate card cannot express, and it is the reason a 40 percent cheaper output token is not automatically a 40 percent cheaper job.

A router that decides per request will sometimes decide wrong, and the standard remedy for a wrong decision is a retry or an escalation. Both cost tokens. A single checkpoint that answers in one pass bills once. Sakana does not publish the number of underlying calls per answered request, the share of requests that escalate, or the token overhead of the routing decision itself — so on any workload where Fugu Max's router picks a weaker model first and climbs, the effective cost per completed task is not $2.00/$6.00 times the tokens you see. It is $2.00/$6.00 times tokens you cannot see.

The same is true of quality, in the direction that matters more. GPT-6.1 Sol's failures are the failures of one known model on your prompts, reproducible and measurable in advance. Fugu Max's failures are partly its router's failures, and the router is a component you cannot inspect, version, or pin. For a workload where a wrong answer costs more than the tokens to produce it — which is most professional work — that is a real exposure, not a theoretical one.

The place this cuts the other way is worth stating too. If Sakana's router is good, it is buying capability that would otherwise cost list price from a frontier model, and paying a blended rate for it. That is a genuinely useful product. The problem is that "if the router is good" is the entire claim, and there is no published evaluation of the router to check it against. A pool that includes NVIDIA Nemotron — a partnership Sakana announced in August 2026 — with the rest of the pool unnamed, and no per-call trace, is a system whose quality you can only discover by running your own traffic through it.

Where GPT-6.1 Sol is genuinely exposed

Two rows, and neither is in GPT-6.1 Sol's favour.

The first is price at the top of the output range. $10.00 per million output is the Sonnet-tier price Open​AI chose, and it is not cheap. On a workload that produces long answers — long-form drafting, code generation at volume, document synthesis — the $6.00 per million that Fugu Max charges is $4.00 per million you are choosing to spend. At a billion output tokens a quarter, that is $4,000, and it is a real line item.

The second is the long-prompt step. Above 272,000 tokens the whole GPT-6.1 Sol request reprices to $4.00/$15.00, which is double on input and 50 percent on output. Fugu Max publishes no equivalent tier, so on a genuinely long prompt the Fugu Max rate card is the simpler one — with the standing caveat that "no published tier" and "no tier" are not the same statement.

There is a third exposure that is neither price nor performance: availability. Fugu Max was reported unavailable in the EU and EEA pending GDPR work, which removes it as an option for a European deployment regardless of what it scores. GPT-6.1 Sol offers US and EU data residency, with the caveat that Open​AI's fast mode is unavailable under EU residency — a narrow but real restriction, and a reminder that neither model's availability story is a single line.

Calling one, piloting the other

GPT-6.1 Sol is on OrcaRouter as openai/gpt-6.1-sol, at Open​AI's list price with 0% markup — $2.00 and $10.00 per million, cached input at $0.10, and the above-272K tier applied exactly as Open​AI sets it. Because the routing layer passes the provider price through rather than marking it up, a change in Open​AI's rates is live here the same day it takes effect rather than at the next billing cycle. Automatic failover sits underneath, which is the cheap way to find out how a $2.00 input rate behaves on your traffic without making it your only path.

Fugu Max is not on OrcaRouter and we do not host it. It is reachable through Sakana's own API, and upgrading an existing Fugu integration to Fugu Max is a one-line parameter change by Sakana's own account. If you are already a Fugu customer the cost of trying it is near zero, which makes it the obvious first move for that audience.

The decision this pairing earns: Fugu Max is worth a pilot and not yet worth a production commitment, because the only performance evidence for it is the vendor's, and the vendor's own benchmark is one of six highlights. GPT-6.1 Sol is the safer default, because its independent scores exist, its failures are attributable to one model, and its ceiling is a published number rather than a pool. Paying $4.00 more per million output tokens for a system whose behaviour you can predict is not obviously the worse trade — and if Sakana publishes a routing trace and a third-party evaluation, this verdict should be revisited immediately.