SubQ Snap 1.1 vs Qwen 3.8: Forty Times the Price per Token, a Different Job

Qwen3.8-Max costs roughly forty times what SubQ Snap 1.1 does per input token — about $1.65 per million against $0.04 — and that ratio is the least interesting fact about the matchup. The two models are billed for different things. Qwen3.8-Max, Alibaba's 2.4-trillion-parameter flagship that launched on August 3, 2026, charges for what it thinks and what it writes. Snap 1.1, Subquadratic's decision model released October 22, 2026, charges only for what it reads, because the answer is a label rather than a sentence. Run the decision on a model you control, and the generation on the one that earns its bill: Qwen3.8-Max is served at provider list price with 0% markup and automatic failover.

Put them in the same pipeline and the price ratio stops looking like a bargain or a ripoff and starts looking like a design decision. One of these models should be doing every routing call in your system. The other should be doing almost none of them — and the cases where Qwen3.8-Max is the right answer are precisely the cases a decision model cannot handle at all.

What Qwen3.8-Max actually is

Alibaba previewed it at WAIC in Shanghai on July 19, 2026 and launched it on August 3. The architecture is a sparse mixture of experts at genuinely frontier scale: 2.4 trillion total parameters with roughly 95 billion active per token, a 1M-token context window, and a maximum output of 131K tokens. It is natively multimodal, accepts text, images and video, and supports thinking mode, function calling, built-in tools and structured outputs across OpenAI-compatible, Anthropic-compatible and OpenAI Responses wire formats.

The benchmark row is a frontier-model row, not a small-model one: GPQA Diamond 92.6, HLE 43.6, OSWorld 86.1, DeepSWE 1.1 at 56.6, MRCR v2 at 0.93, and an Artificial Analysis Intelligence Index of 45. Those are the kind of numbers that justify a $1.65/$4.95 list price — roughly $4,950 to push a billion input tokens through it, or about what Snap 1.1 charges for 124 billion.

A note on naming, because it matters for anyone shopping: Qwen3.8-Max is the hosted flagship. It has open weights released as Qwen3.8-2.4T-A95B under a custom licence, and there is a separate, genuinely open sibling — Qwen3.8-27B, Apache 2.0, 262K native context, an Intelligence Index of 34, and roughly $0.42 to $0.50 per million input tokens. The 27B is the one most teams should be comparing against small models. The Max is the one you call when the task needs the biggest model available.

What Snap 1.1 actually is

Snap 1.1 is a decision model, which is a narrower claim than it sounds. It accepts up to 384K tokens of text and images and an answer set of two to 255 alternatives per question, and returns a selection with a calibrated probability. No prose. No JSON to repair. No thinking tokens. Subquadratic's case is that a fixed-answer decision is a different computational job from generation, and its sparse-attention architecture processes the input at constant compute per token rather than paying quadratic attention across a long document.

The headline result is a score of 65 on the Artificial Analysis Decision Index v1.0, tied for first with d1-ultra and Elephant 9B, and ahead of Jev 1.14 at 57. The number underneath it is the one to hold onto: accuracy of 20%, against 32% for d1-ultra and 12% for Elephant 9B, with the index score carried by a reported 86% abstention rate on questions it cannot answer. Snap 1.1 is a model that has been trained to decline, and its ranking is a measurement of how well it declines.

The specs, side by side

• Scale — Qwen3.8-Max is a 2.4T-parameter MoE with ~95B active; SubQ Snap 1.1 is a sparse-attention decision model of undisclosed size.

• Context — 1M tokens for Qwen3.8-Max, 131K max output; 384K tokens for Snap 1.1, no generated output.

• Input — text, images and video for Qwen3.8-Max; text and images for Snap 1.1.

• Output — free-form text, reasoning traces, tool calls and structured output for Qwen3.8-Max; a choice or score with a probability for Snap 1.1.

• Price — roughly $1.65 to $2.00 per 1M input and $4.95 to $6.00 per 1M output for Qwen3.8-Max depending on the provider, with OrcaRouter listing $2.00 and $6.00; $0.04 per 1M input and nothing for output for Snap 1.1.

• General intelligence — Artificial Analysis Intelligence Index of 45 for Qwen3.8-Max; not measured on that index for Snap 1.1.

• Decision quality — 65 on the Artificial Analysis Decision Index v1.0 for Snap 1.1; not measured on that index for Qwen3.8-Max.

• Deployment — hosted API and open weights under a custom licence for Qwen3.8-Max; hosted API only, no weights, for Snap 1.1.

Those last two rows are why this comparison keeps getting made badly. Each model is measured on an index the other has never been run on, so any claim that one "beats" the other is an inference, not a result. Nobody has published a head-to-head.

The anti-pattern: using a flagship as a router

It is common, and it is expensive in a way that does not show up until the bill arrives. A team builds an agent, needs it to decide between five tools, and reaches for the strongest model available because the decision matters. The decision does matter. The model is simply being asked to do something it was not optimised for, at a price built for something else.

Run the arithmetic on a million routing decisions with 6,000 tokens of context each — six billion input tokens. Qwen3.8-Max at the cheapest listed rate of $1.65: $9,900 for input, plus perhaps twenty output tokens per decision, or 20 million output tokens at $4.95 per million, for another $99. Call it $10,000. At OrcaRouter's $2.00 and $6.00 listing the same volume comes to $12,120. Snap 1.1 over the same volume: $240, and no output line at all. The gap is not a discount. It is a factor of forty to fifty.

Latency tells the same story from the other side. A 2.4-trillion-parameter model generating even a short answer incurs a first-token delay and a decode cost that a decision model simply does not have — Snap's pitch is a sub-second decision layer, and Subquadratic publishes no latency figure for 1.1 beyond that. In an agent loop that makes twenty decisions per task, that difference compounds into seconds per task, which is the difference between an agent that feels instant and one that feels broken.

What Qwen3.8-Max does that Snap cannot

Everything that involves producing something. It writes the escalation note, drafts the customer reply, summarises the 900-page contract, generates the code, calls a tool with a valid argument schema, and reasons through a problem whose answer is not in a list you supplied in advance. Snap 1.1 cannot do any of those things, not because it is worse but because it has no mechanism for them.

The 1M-token window is also a real capability, not a spec-sheet flex. Alibaba reports MRCR v2 at 0.93, which is a retrieval score rather than a nominal context length, and it is the number that determines whether a million-token window is usable. A 384K window with strong effective retrieval can beat a 1M window with weak retrieval, but Qwen's retrieval score suggests it is not relying on the nominal figure.

And the open weights matter for a different reason: Qwen3.8-2.4T-A95B can be self-hosted under a custom licence, which means the largest model in the family is not purely an API dependency. Snap 1.1 has no such option.

Where they fit together

The defensible architecture is not a choice between them, it is a division of labour. Snap 1.1 handles the high-volume, fixed-answer decisions where a calibrated probability and a refusal option are worth more than fluency: which queue, which tool, whether to escalate, whether this document is relevant at all. Qwen3.8-Max handles the small fraction of cases where the decision requires reasoning over material that is not in the option list, and handles everything downstream of the decision that requires writing.

The proportion is the whole economic argument. If Snap abstains on 14% of cases and those route to the flagship, you pay flagship prices on 14% of traffic instead of 100%, and you get a calibrated signal telling you which 14% — which is a better allocation mechanism than a fixed threshold on a general model's softmax.

Qwen3.8-Max is one of more than 200 models behind a single OrcaRouter key, served at provider list price with 0% markup and automatic failover across providers. Snap 1.1 is not ours; it runs through Subquadratic's own API. That split is worth knowing before you design the pipeline, because it means the generation tier can be consolidated onto one contract and one billing line while the decision tier stays a separate, single-vendor dependency. If that second dependency is unacceptable, the alternative is to run the decision on a Qwen3.8-27B instance you host yourself — worse at abstaining, much better at explaining itself, and entirely under your control.

The honest default

If you are choosing one model to make decisions for an agent, and the decisions have a fixed answer set, choose the decision model. The forty-to-one price ratio is real, but it is the smaller half of the argument; the larger half is that a model trained to say "I don't know" gives you something a flagship cannot, which is a place to put the human in the loop.

If you are choosing one model to do everything — decide, reason, write, call tools, read a million tokens — choose Qwen3.8-Max, and accept that you are paying flagship rates for routing calls that a $0.04 model would have handled better. Most teams discover that they are doing exactly this about three months after launch, when the invoice arrives and nobody can explain why the router costs more than the product.