Dessa vs Qwen3.8-Max: A $6 Output Token Against a Model With No Price

Qwen3.8-Max charges $2.00 per million input tokens and $6.00 per million output tokens. Dessa charges nothing, because Block is not selling it, hosting it, or releasing it. That is not a knock on Dessa — it is the entire comparison. One model is a priced, callable, independently scored flagship; the other is a speech research preview announced on 23 October 2026 with a waitlist and a demo page. Putting them side by side is useful precisely because they sit at opposite ends of the same product question: what does it cost to give a user a voice?

Qwen3.8-Max is Alibaba's hosted flagship, announced on 3 August 2026, a 2.4-trillion-parameter sparse mixture-of-experts model with roughly 95 billion parameters active per token. Dessa is Block's full-duplex speech-to-speech model, trained from scratch on human conversation with no text and no pretrained models. One reads and reasons over a million tokens of text. The other listens and talks at the same time. Neither can do the other's job, and only one of them can be called today.

What each model is actually for

Qwen3.8-Max is a general-purpose reasoning and generation engine. It takes text, images, and video and returns text, with a context window of about one million tokens, a maximum output of 131,072 tokens, and a thinking mode that can spend up to 262,144 tokens on chain-of-thought. Alibaba's own benchmark table is aimed at frontier reasoning and agentic work: GPQA Diamond 92.6, PaperBench 93.0, OSWorld 86.1, BabyVision 82.0, all vendor-reported. Artificial Analysis, which is independent, puts it at 45 on its Intelligence Index and measures 42.4 output tokens per second.

Dessa is a conversational interface. It takes an audio stream and returns an audio stream, with no text in between, and it is full-duplex, meaning it can listen while it speaks. Block's demos show it backchanneling, laughing mid-sentence, matching volume, and finishing a speaker's forgotten film title. There is no benchmark table, no parameter count, and no context window because there is no token interface to measure.

These are not competing answers to one question. They are answers to two questions that a voice product has to ask in sequence: how does the system understand, and how does it speak?

The numbers, and the hole where Dessa's should be

Every Qwen figure below is from Alibaba's published pricing and model documentation or from Artificial Analysis, and is labelled accordingly. Every Dessa figure is from Block's announcement page, which publishes very little.

• Price — Qwen3.8-Max $2.00 input and $6.00 output per million tokens vs Dessa no price.

• Cached input — Qwen3.8-Max $0.25 per million tokens vs Dessa not applicable.

• Parameters — Qwen3.8-Max 2.4T total, roughly 95B active vs Dessa undisclosed.

• Context — Qwen3.8-Max about 1,000,000 tokens; Artificial Analysis lists 984k vs Dessa not stated.

• Max output — Qwen3.8-Max 131,072 tokens vs Dessa not applicable, because output is audio.

• Modality — Qwen3.8-Max text, image, and video in, text out vs Dessa audio in, audio out.

• Independent score — Qwen3.8-Max 45 on the Artificial Analysis Intelligence Index vs Dessa none.

• Vendor benchmarks — Qwen3.8-Max GPQA Diamond 92.6, OSWorld 86.1 vs Dessa no quantitative claims at all.

• Availability — Qwen3.8-Max a priced API since August 2026 vs Dessa a waitlist and a small safety-researcher preview.

• Licence — Qwen3.8-Max a hosted commercial API, with a separate open-weight release under its own terms vs Dessa no licence published.

One hears, one reads

The modality gap is the part that gets lost in a spec comparison. Qwen3.8-Max accepts text, images, and video, and emits text. Audio is not among its input modalities and it does not emit audio at all; nothing in the Qwen3.8 line is a speech-to-speech model.

That means a voice product built on Qwen3.8-Max is necessarily a pipeline: speech recognition turns the caller's audio into text, Qwen3.8-Max reasons over that text, and speech synthesis turns the reply back into audio. Each stage is a separate vendor, a separate failure mode, and a separate line on the invoice. The pipeline works — it is how most voice agents in production today are built — but it has a structural cost that Dessa's architecture is designed to remove.

Dessa's structural advantage is not intelligence. It is that there is no transcript to commit to. A pipeline has to decide when you have finished speaking before it can respond; a duplex model does not. That is why Block's demos can show overlap and interruption, and why a pipeline has to fake those behaviours with endpoint detection and barge-in logic that cancels the synthesiser mid-word.

What a live voice agent actually costs on Qwen3.8-Max

The arithmetic is more reassuring than the rate card suggests. A ten-minute spoken conversation runs to roughly 1,500 words, which is about 2,000 tokens of input once the system prompt and conversation history are included. A reply of similar length is about 1,000 output tokens if the agent is concise. At $2.00 and $6.00 per million, that is $0.004 of input and $0.006 of output — roughly one cent of language-model cost per ten-minute call.

At 10,000 calls a day, the language model bill is about $100 a day, or roughly $3,000 a month. That is not nothing, but it is almost certainly smaller than the speech recognition and synthesis bill for the same volume, and smaller than the engineering cost of the orchestration layer. The intuition that voice is expensive because tokens are expensive is wrong at current prices. Voice is expensive because audio in and audio out are expensive, and because the pipeline's latency is the sum of its stages.

The one-million-token context changes what the agent can hold. A ten-minute call is a rounding error against that window, which means the system can carry the full transcript, the customer's file, and a policy document at once without retrieval. That is a capability Dessa does not have an interface for at all — a speech-only model cannot read the policy document.

The latency budget is where the architectures diverge

In a pipeline, the user stops speaking, the recogniser waits for an endpoint, the language model produces a first token, and the synthesiser produces first audio. Streaming each stage hides some of that, but the chain still has a floor, and barge-in means tearing down the synthesiser and restarting the recogniser. The result is the familiar voice-assistant rhythm: a pause, then a reply, then a pause.

Dessa's claim is that the rhythm disappears because there is no chain. Block says inference runs faster than real time on a MacBook Pro M4 Max, which — if true — means the model is not the bottleneck. But the company publishes no latency figure in milliseconds, no interruption-recovery rate, and no measurement under network or background-noise conditions. "Faster than real time" is a throughput statement, not a responsiveness one. A model can generate audio faster than playback and still take 800 milliseconds to react to an interruption.

For a product team, that missing number is the whole ballgame. The reason to adopt a duplex model is timing. A vendor that will not publish timing has not yet made the case.

Availability, routing, and what you can do on Monday

Qwen3.8-Max is in the OrcaRouter catalogue, which means it is reachable through the same key as the rest of the 200+ models we route, at 0% markup — provider list price passed through — with automatic failover across providers. If Alibaba changes the rate, the change is live on our side the same day, because there is no markup layer to reprice. The routing DSL can compose Qwen3.8-Max with a cheaper model for the bulk of a transcript and reserve the flagship for the reasoning step; model fusion can run a panel of models over the same question when the answer matters more than the bill.

There are cheaper points in the same family if the flagship is more than the job needs. Qwen3.8-27B is on our catalogue at $0.33 input and $2.40 output per million tokens, and Qwen3.8-Flash at $0.15 and $0.47. For a voice agent whose language-model cost is already about a cent per call, the flagship is affordable; the cheaper tiers matter at higher volume or for simpler turns.

Dessa is not in our catalogue and we do not host it. Block's page offers a waitlist and early access for a small group of safety researchers. There is no API to route, no price to pass through, and no endpoint to fail over to.

When Dessa would change the answer

If Block ships an API at a plausible price, the comparison stops being about tokens and starts being about minutes. The right metric would be cost per minute of natural conversation, including the speech recognition and synthesis that Dessa removes. On that metric, a duplex model at even a few cents a minute could beat a pipeline whose audio stages cost more than its language model.

The second condition is quality. A duplex model that is pleasant for thirty seconds and confused after five minutes is worse than a pipeline that is slower but reliable. Block's demos are too short and too curated to answer that, and the company's own open questions — factual retention, language learning, in-context learning — are exactly the ones that determine whether a long conversation holds together.

What to do now

Build the voice product on the pipeline, because it is the only one with a price and a contract. Use a strong recogniser and synthesiser, put Qwen3.8-Max in the middle where the reasoning happens, and keep the language-model layer swappable so the expensive part can move when prices move. Then watch for the two numbers that would justify rebuilding around a duplex model: a published latency figure and a price per minute.

The comparison in one line: Qwen3.8-Max is what you can buy today, and Dessa is what you might want to buy tomorrow. The gap between them is not intelligence. It is a rate card, an endpoint, and a benchmark — three things Block has not yet published, and three things Alibaba publishes for a living.