GPT-6 Astra vs Fugu Max: The Flat Rate Meets the Tiered One

Fugu Max charges $2.00 per million input tokens and $6.00 per million output tokens, and it charges that at 8,000 tokens or at 800,000. GPT-6 Astra charges $10.00 and $50.00 — until a request passes 272,000 input tokens, at which point it charges $20.00 and $75.00 for the whole request. That single structural difference is the most consequential thing about this matchup, and it is the one that gets buried under a benchmark argument neither side can win. Sakana AI released Fugu Max on September 11, 2026; O​penAI has had GPT-6 Astra generally available since September 3, 2026, with its meters unchanged through the repricing that hit the tiers beneath it on October 7. So this is not a new-model comparison. It is a comparison between two pricing philosophies, one of which is a pricing philosophy and the other of which is an orchestration system wearing one.

The two rate cards, read rather than summarised

Sakana's own pricing page is unusually blunt about what Fugu Max is selling. Fugu Ultra v2 is priced like a frontier model should be — $5.00 in, $30.00 out, $0.50 cached, with a higher band at $10.00 / $45.00 / $1.00 above 272K tokens. Fugu Max's rates are then described on the same page as fixed regardless of context length. There is no upper band to plan around. $2.00, $6.00, $0.25 cached, at any prompt size.

GPT-6 Astra is the opposite construction. Its first tier runs to 272,000 prompt tokens at $10.00 / $50.00 with cached input at $1.00 and cache writes at $12.50. Past that line the request is repriced at $20.00 / $75.00, with cached input at $2.00 and cache writes at $25.00 — applied to the whole request, not to the excess. That is O​penAI's documentation as our own catalogue reads it today, and it has not moved since September.

Put the two side by side on the dimensions that decide a bill:

• Input, short prompt — GPT-6 Astra $10.00 per 1M vs Fugu Max $2.00; Fugu Max is 5x cheaper

• Output — GPT-6 Astra $50.00 per 1M vs Fugu Max $6.00; Fugu Max is 8.3x cheaper

• Cached input — GPT-6 Astra $1.00 per 1M vs Fugu Max $0.25; Fugu Max is 4x cheaper

• Long prompt, above 272K — GPT-6 Astra $20.00 / $75.00 per 1M vs Fugu Max $2.00 / $6.00, unchanged

• Context tracked — GPT-6 Astra 1,050,000 tokens, 128,000 max output vs Fugu Max not published

• Input modality — GPT-6 Astra text, image and file in, text out vs Fugu Max not published

• Independent score — GPT-6 Astra 53 on the Artificial Analysis Intelligence Index vs Fugu Max none, no entry on that board

• Weights — both closed; Fugu Max is not a checkpoint at all

Where the 272,000-token line actually bites

A 300,000-token request on GPT-6 Astra is not 10% more expensive than a 270,000-token request. It is priced at $20.00 input and $75.00 output on every token in the call, which makes the step change roughly a doubling of the input bill and a 50% increase on output the moment you cross the line. On Fugu Max the same request costs $2.00 and $6.00 per million like any other.

That matters more here than on most comparison pages, because of what GPT-6 Astra is sold for. O​penAI's positioning is long-horizon, end-to-end work — the enterprise page published on September 9 frames it as a model that takes a task from instruction to finished artefact over many steps and a long stretch of time. An agent loop that re-sends a growing context is exactly the shape that sails past 272,000 tokens without anyone deciding to. Astra's 1,050,000-token window invites the workload that triggers its own surcharge.

Fugu Max is constructed for the opposite instinct. Sakana describes it as routing each task to the leanest model capable of solving it, and the flat rate is what makes that routing legible from the outside: whoever the coordinator picks and however long the prompt, the meter reads the same. There is one further difference that cuts the other way and is worth stating, because the pricing page does not state it as a warning. Sakana's page notes that web search and web fetch are billed as stack fees — a single rate based on the top-tier model involved in the call. So the flat per-token rate is genuinely flat, and the composite call is not necessarily $2 and $6.

Whose scores are these, exactly

Sakana's launch material says Fugu Max achieves the best overall score on six benchmarks: Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and SWEFish, described as Sakana's own internal coding benchmark. No figures accompany any of those six claims. The company also states that Fugu Max expands the cost-performance Pareto frontier on seven of ten benchmarks and that its output pricing is 40–60% below Sonnet 5, GPT-5.6 Terra and Kimi K3. Every one of those statements is Sakana's own, and none of it has been reproduced — there is no Artificial Analysis entry for Fugu Max or for any Fugu model.

That absence is more than a labelling issue on this particular matchup, because of how an orchestrator's scores work. A Fugu Max result is not a property of one model at one version. It is a property of a routing policy, a pool whose membership Sakana has not published, and a coordinator trained to choose among them. Sakana's pool grew in this release to include an unprecedented number of open-weights and specialised models, naming the NVIDIA Nemotron family through an NVIDIA collaboration — and stops there. Change one pool member and the benchmark number moves without the product name changing.

There is a narrow exception worth knowing, because it is the kind of detail that gets lost. Sakana's note about the September release says explicitly that Fable 5, Fable 5.1 and GPT-6 Astra are not in Fugu Ultra v2's pool. That sentence is about Ultra v2. It says nothing about Fugu Max, whose membership remains undisclosed. So the question a reader actually wants answered — is the $10/$50 model I am being told to replace already inside the $2/$6 coordinator — is not answered by either statement. Anyone who tells you which way it falls has inferred it.

On the GPT-6 Astra side the evidence is a different shape but not a simpler one. Artificial Analysis measures it at 53 on its current Intelligence Index, with a cost of $3.26 per completed index task, 60M output tokens across that run against a board median near 85M, and 44.6 output tokens per second — which places it seventh of 216 models on intelligence and 168th of 185 on speed. Those are third-party measurements on a fixed suite, and they are the most defensible numbers on this page. O​penAI's own headline claims for Astra — 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, 100% on ExploitBench, 57.9% on Terminal-Bench 4.0 — are vendor-reported and unreproduced, and the ARC-AGI-3 figure in particular needs its harness attached: ARC Prize's independent run under its standard harness measured 62.7%, and published the 99.9% as the result under a provider adapter that carries reasoning state between requests.

What routing does to a comparison like this

The practical difficulty is that one side of this matchup is on our catalogue and the other is not. GPT-6 Astra is served at openai/gpt-6-astra through one O​penAI-compatible endpoint, at OpenAI's own list rates passed through with 0% markup of our own — which means that when a meter on it moves, the meter here moves the same day rather than whenever a margin table is refreshed. Fugu Max is not routed by us and this page is not claiming it is. It reaches you through Sakana's own O​penAI-compatible API and several third-party platforms, and it is not available in the EU or EEA while Sakana works toward GDPR compliance.

So a genuine A/B between these two is two integrations, two bills and two sets of credentials, and the second one has a regional restriction attached. What the shared key does buy is the cheap half of the experiment: pinning a slice of traffic to Astra behind automatic failover, so that a rollback is a config change rather than an incident. That is the transport-layer version of the resilience Sakana sells as the premise of orchestration — weaker in kind, but available today and on a model whose behaviour can be pinned to a version.

The decision, stated as a decision

Choose Fugu Max if your work is text-heavy, verifiable, and lives at the long end of the context range — or if the median cost of a completed task is the number that governs your budget rather than the cost of any single call. The flat rate is a real structural advantage over a meter that doubles above 272K, and $0.25 cached input is the most aggressive cache rate in either rate card. Do not choose it for a multimodal pipeline on the strength of the release page, because the release page does not document whether it accepts an image. Test it first, and set an output-token budget before you start, because an orchestrator's token consumption is not something you can estimate from the model name.

Choose GPT-6 Astra if the work is genuinely agentic, computer-use heavy, and leans on the top of a million-token window — the workload where the independent evidence is strongest and where per-task cost, not per-token price, is what you are managing. Its $3.26 cost per index task is high, and it is high for a reason a careful buyer can see: the model is concise on that suite, emitting 60M tokens where the median model emits 85M. On a deep, multi-step task that produces a finished artefact, a higher rate on fewer tokens can land in the same place as a low rate on many. Do not choose it for short-prompt, high-volume traffic. At five times Fugu Max's input price and eight times its output price, a workload that never approaches the 272K line gets none of what Astra's rate card is paying for.

The honest summary is that these two products are not priced against each other so much as priced against different fears. Astra's rate card is built for the customer who fears a cheap model failing a long task. Fugu Max's is built for the customer who fears a good model being wasted on an easy one. Which fear is yours is a question about your traffic, and neither vendor's benchmark table answers it.

A release is also the moment to check what has gone wrong before: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.