Phonic 1 vs Qwen Audio 3.1: 82.0% and 79.9% Are Not the Same Number

The comparison everyone will make between Phonic 1 and Qwen Audio 3.1 Realtime is a two-line table: Qwen scores 82.0% on τ-Voice, Phonic scores 79.9%, therefore the cheaper model wins. It is a clean story and it is wrong, because those two figures come from different instruments. Qwen's 82.0% is from the vendor's own paper, measured on an adapted, half-duplex speech-to-text variant of the benchmark. Phonic's 79.9% is the agentic component of Artificial Analysis's speech-to-speech index, measured on the real thing. Subtracting one from the other is the same error as comparing an exam score from a practice paper with one from the exam.

What makes the mistake worth writing about is that it is the only comparison available. Qwen Audio 3.1 Realtime does not appear on Artificial Analysis's speech-to-speech board at all — the board lists Qwen Audio 3.0 Realtime Plus at 69.5 and Qwen Audio 3.0 Realtime Flash at 67.7, and nothing newer. So the newest cheap challenger in this category has published its own numbers, in its own units, and no independent body has scored it. That gap is the story, and the honest article is mostly about what cannot yet be said.

What Qwen Audio 3.1 actually is

Alibaba's Qwen-Audio-3.1 landed in late September 2026 as a five-model stack rather than a single release: ASR, ASR-Next, TTS, TTS-Next and Realtime. The one that competes with Phonic 1 is Qwen Audio 3.1 Realtime, and specifically the Realtime-Plus tier, which is available only through QwenCloud over a WebSocket — no open weights, unlike the TTS and ASR members of the family, and no self-host path in the public material.

• Context — 262K total, 245K max input and 16K max output, against Phonic 1's unpublished context.

• Throughput limits — 60 requests and 100K tokens per minute by default on QwenCloud; Phonic publishes no equivalent session cap.

• Features — function calling, web search, structured outputs, context caching, fine-tuning and voice cloning are all listed for Qwen Audio 3.1 Realtime; Phonic lists inline synchronous tools, 50+ languages and mid-conversation language switching.

• Availability — Qwen Audio 3.1 Realtime is API-only; Phonic 1 offers cloud, single-tenant, VPC peering and self-host.

Those two rows are the ones that matter for procurement. Qwen's family usually ships weights; this member does not, which removes the main reason a team chooses Qwen over a hosted vendor. Phonic's self-host option is the mirror image — it is the only way either of these two models runs anywhere other than someone else's cloud.

The benchmark that is not the same benchmark

The 82.0% figure appears in Alibaba's arXiv paper on Qwen-Audio-3.1, presented as an improvement from 78.4% in the previous generation. The same paper is explicit that this is a half-duplex, speech-to-text adaptation of τ-Voice rather than the full-duplex agentic evaluation that produces the numbers on the public leaderboard. That is not dishonest — it is disclosed — but it means the figure answers a narrower question: given a transcript-style setup and a task, how often does the model complete it.

Phonic 1's 79.9% answers the wider question, and the gap between the two questions is not small. A half-duplex adaptation removes the parts of an agentic call that are hardest to get right: barge-in, overlapping speech, deciding when to stop listening, and recovering when the caller changes their mind mid-tool-call. Every one of those is a place a model loses tasks it would otherwise complete.

The direction of the error is therefore predictable, and it is the opposite of the obvious reading. Qwen's 82.0% is very likely an optimistic ceiling for its full-duplex behaviour, not a like-for-like win over Phonic's 79.9%. Whether it clears that bar is unknown. Alibaba has not put the model in front of the evaluator that produced the 79.9.

What Qwen does publish, and where its own numbers turn against it

Alibaba's release is unusually forthcoming about failures, which makes it more useful than the headline τ-Voice line. Read the whole table and the model looks less like a decisive challenger and more like a strong second-generation realtime system with two specific soft spots.

• FLEURS word error rate — 9.01 down to 3.98, a genuine improvement in recognition accuracy across languages.

• 14-language Big Bench Audio average — 81.7% to 88.1%.

• Audio MultiChallenge — 47.12 to 52.21.

• Background-speech response rate on Full-Duplex-Bench v1.5 — 73.0% down to 13.0%, meaning the model is far less likely to answer speech that was not directed at it.

• Interruption stop latency — 1.116s, against 0.383s for GPT-Realtime-2. This is the number to weigh against Phonic's 0.61s to first audio: one measures how fast a model starts, the other how fast it stops, and a slow stop is what callers describe as being talked over.

• Unwanted resume rate — 0.035 rising to 0.130, a regression the paper reports against its own predecessor; the model is more likely to start speaking again when it should not.

Those last two lines are the most underreported facts in the Qwen Audio 3.1 launch. A model that takes 1.116 seconds to yield and resumes when unwanted will feel, on a real phone call, worse than its accuracy scores suggest — and that is exactly the dimension where Phonic 1's marketing leans, with its "responds before the caller finishes" claim and its sub-600ms speech-to-speech latency.

The one comparable data point

Because Qwen Audio 3.1 Realtime has no index score, the only way to place it against Phonic 1 on the same instrument is through its predecessor. Qwen Audio 3.0 Realtime Plus scored 69.5 on the speech-to-speech index, with 54.6% agentic performance, 99% speech reasoning, 77.8% task success, 798 arena Elo and 1.54s to first audio.

• Agentic performance — Phonic 1 79.9 vs Qwen Audio 3.0 Realtime Plus 54.6. If 3.1 delivers the improvement its own paper implies on tasks, it closes some of that 25-point gap; nobody has measured how much.

• Arena preference — Phonic 1 1,012 vs 3.0's 798. This is the largest gap in the comparison and the one least likely to have closed, since arena preference is a human-vote measure that moves slowly between generations.

• Time to first audio — Phonic 1 0.61s vs 3.0's 1.54s. Qwen's realtime tier has historically been the slower family here, and 3.1 publishes no time-to-first-audio figure.

• Task success — Phonic 1 92.3% vs 3.0's 77.8%.

The three-generation pattern is consistent: Qwen's realtime models have been strong recognisers and weak agents, and 3.1 is the generation that claims to fix the agent half. Whether it did is precisely what the missing index score would tell you.

Two price units that do not convert

Qwen Audio 3.1 Realtime bills per token: $0.80 for a million text input tokens, $6.40 for audio input, $6.40 for text output and $24 for audio output in the Singapore region, with Beijing rates about 14% lower across the board. Phonic 1 bills $0.07 per conversation minute, flat, with no token charges.

There is no public conversion between those units. To turn a token rate into a per-minute cost you need the audio token density — how many tokens a minute of speech occupies — and Alibaba does not publish it for this model. Anyone who quotes you a per-minute equivalent for Qwen Audio 3.1 is either guessing at the density or borrowing it from a different model family with a different codec. The direction of the answer is not in doubt: Qwen's audio rates are low enough that a well-tokenised minute will very likely land below seven cents, and its ~85% price cut on the Realtime tier was the headline of the launch. The magnitude is unknown, and the unknown is large enough to matter for a high-volume deployment.

• Input audio — Qwen $6.40 / 1M tokens (Singapore) against Phonic $0.07 per conversation minute, with no published exchange rate between them.

• Output audio — Qwen $24 / 1M tokens; Phonic's flat minute covers output audio.

• Regional pricing — Qwen's Beijing region runs roughly 14% cheaper than Singapore; Phonic publishes one global rate card.

• Minimums — Phonic's Pro tier carries a $200 monthly minimum; QwenCloud publishes default rate limits rather than a platform minimum.

The shape difference is the part that survives the missing conversion. A per-token bill tracks how much the model actually does — long transcripts, verbose replies and tool-heavy turns all cost more. A per-minute bill tracks how long the phone line was open. For an agent that spends most of a call listening, the token model is the friendlier one; for an agent that answers at length, it is not.

Which question are you answering

If you are choosing a model to deploy on your own infrastructure, this comparison is not for you: Qwen Audio 3.1 Realtime has no weights, and Phonic 1 self-hosts, which makes Phonic the only option here that leaves the vendor's cloud at all. That is a one-line answer and it is easy to miss under the benchmark talk.

If you are choosing a hosted realtime model on cost, Qwen Audio 3.1 is the more interesting bet — dramatically cheaper per unit, a well-documented jump in recognition accuracy, and a vendor with a deep open-weight ecosystem behind it. The risk is that its agent half is unmeasured and its interruption behaviour regressed in the vendor's own table.

If you are choosing on measured agentic performance, Phonic 1 is the only one of the two you can actually rank, because it is the only one that has been ranked. That is a weaker claim than "Phonic is better" and it is the true one.

And if you are running Qwen's text models alongside a voice layer, that half of the stack is already routable — Qwen3.8-Max sits on OrcaRouter among 200+ models on a single key at provider list price with no markup, so a price cut on Alibaba's side is live the same day. The audio endpoint is not on our catalogue and neither is Phonic 1; each comes from its own vendor. What a router buys you in a market this young is the ability to keep a working voice path in place while a newer, cheaper one earns its way in — automatic failover means testing Qwen Audio 3.1 against your own traffic costs you a routing rule, not a migration.

The honest conclusion is that this matchup is currently undecidable on the axis both vendors want it decided on. Alibaba has published a number that flatters its model by measuring an easier task. Phonic has published a number that is measured properly but on one board, one configuration, one week of availability. The first thing to watch is not a price cut or a benchmark — it is the appearance of Qwen Audio 3.1 Realtime on a board that scores both models the same way. Until then, the only defensible claim is that Phonic 1 is the known quantity and Qwen Audio 3.1 is the cheap unknown, which is a perfectly good reason to try the second one behind a failover rule rather than in front of your customers.