A generated title card headed 'Tell 1 vs Qwen 3.8'. Left card labelled 'Tell 1' carries a stopwatch icon and the lines 'first on latency' and 'fifth on tool calls'; right card labelled 'Qwen 3.8' carries a wrench icon and the lines '1M token context' and 'native tool calling'. A divider reads 'speed versus the transaction'. The OrcaRouter logo is composited in the bottom-right corner.
Guides & Insights

Tell 1 vs Qwen 3.8: The Tool-Calling Gap Nobody Puts in the Headline

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Retell publishes a leaderboard for voice models and puts Tell 1 first on it. The same leaderboard, read a little further down each column, says Tell 1 is fifth at calling tools. That single rank is the most useful thing on the page, because it marks the exact boundary between what Tell 1 was built to do and what a general-purpose model like Q​wen 3.8 — the generation whose flagship is Qwen3.8-Max and whose open dense member is Qwen3.8-27B — is now genuinely good at. A phone agent that cannot book the appointment is worse than a slow one. A 2.4-trillion-parameter model that needs 8 seconds to say hello cannot answer the phone at all. The interesting question is not which model is better; it is where in your stack each one stops being the right tool.

Worth naming the evidence levels before anything else, because these two models are documented in incompatible ways. Qwen3.8-Max launched as a priced API on 3 August 2026 with a published capability matrix, a 1M-token context and vendor benchmarks; Qwen3.8-27B followed as downloadable Apache-2.0 weights; both are independently scored by Artificial Analysis. Tell 1 has one row in a benchmark its owner wrote and no model card at all. So one side of this page is measured and the other is reported, and every number below is labelled accordingly.

What Tell 1's fifth place actually costs you

The vCX table's own numbers are worth reading in the order they appear rather than in the order ranks suggest. On the row Retell calls colloquial, Tell 1 scores 96.6% and no configuration in the field of 62 beats it. On the row it calls tool calling, the same model scores 95.4% and four others do better — which is not a flaw in the model so much as a statement about what it was tuned to be good at. The 94.45% average and the 72.5 ms median both follow from that priority; so does a 3.5 cent per minute model cost that ranks 19th, comfortably mid-table.

The fifth-place figure — 95.4%, 954 of 1,000 evaluations passing — is not bad in absolute terms. It is the pattern that matters. In a contact-centre flow, tool calls are where the agent does something irreversible: books a slot, takes a payment, transfers a caller, cancels an order. A 4.6% failure rate on those steps is a system that mostly works and occasionally does the wrong thing, which is precisely the failure mode that gets an AI agent pulled out of production after a month. The colloquial and latency wins, by contrast, are the ones that get it approved in the first place. That asymmetry — better at the demo than at the transaction — is the thing to test, not the thing to assume.

Two caveats on the benchmark itself, since it is the only public evidence for Tell 1. Its row carries version 17 of the harness while 50 of the 62 rows carry version 1, so "#1 overall" aggregates across two different test sets. And it is Retell's own benchmark, run on Retell's traffic patterns, which is a legitimate methodology and not an independent one.

What Q​wen 3.8 brings to the same problem

Q​wen 3.8 is a generation, not a model, and the split matters here. Qwen3.8-Max is the hosted flagship: 1M-token context, text, image and video input, thinking mode, function calling, built-in tools and structured outputs, served in O​penAI-, A​nthropic- and DashScope-compatible wire formats. Artificial Analysis scores it 33.7 on its Intelligence Index — better than 73% of the models it compares — with 76.2 on its Coding Index, 92.8 GPQA Diamond, 88.8 on Terminal-Bench 2.1 and 47.8 on tau-banking, the vendor-and-independent mix our model page cites. Qwen3.8-27B is the open-weight dense member: 27B parameters, 256K context, multimodal, Apache 2.0, and on Q​wen's own card 90.3 LiveCodeBench v6, 89.2 GPQA Diamond, 84.3 OSWorld-Verified, 79.0 QwenSWEBench.

Read the tau-banking number twice if tool use is your concern. It is a benchmark of tool-augmented customer-service tasks — looking up an account, changing an order, following a policy — and 47.8 is a hard score on a hard test, not a marketing figure. General-purpose models got good at this before they got good at sounding human, and the ordering is the whole story of this comparison: Q​wen 3.8 will run your tool loop more reliably than a 72-millisecond voice model will, and it will take seconds to start talking.

• Latency — 72.5 ms median for Tell 1 inside Retell's harness; our own served measurements put Qwen3.8-Max at 8,235 ms p50 and Qwen3.8-27B at 4,197 ms p50 over a seven-day window, which is a different quantity measured a different way, and neither is a phone-call turn.

• Tool calling — 95.4% (rank 5 of 62) for Tell 1 on Retell's harness; 47.8 tau-banking and full native function calling with structured outputs for Qwen3.8-Max, on benchmarks scored by people who do not sell either model.

• Context — 1M tokens for Qwen3.8-Max, 256K for Qwen3.8-27B; unpublished for Tell 1, and largely irrelevant on a call, where the knowledge base sits outside the model.

• Price — Tell 1 is inside Retell's $0.07 to $0.31 per minute, with 3.5 cents of that attributed to the model; Qwen3.8-Max is $2 input and $6 output per million tokens, Qwen3.8-27B $0.33 and $2.40, both at provider list through OrcaRouter with no markup added.

A generated two-column scoreboard headed 'Tell 1 vs Qwen 3.8'. The Tell 1 column reads 'Latency: 72 ms', 'Tool calling: 95.4% rank 5', 'Context: unpublished', 'Weights: none', 'Evidence: vendor benchmark', 'Price: 3.5c per minute'. The Qwen3.8-Max column reads 'Latency: 8235 ms p50', 'Tool calling: tau-banking 47.8', 'Context: 1M tokens', 'Weights: open 27B sibling', 'Evidence: AA Index 33.7', 'Price: $2/$6'. A footer reads 'Tell 1 per Retell; Qwen figures per Artificial Analysis and OrcaRouter.' The OrcaRouter logo is composited in the bottom-right corner.

The arithmetic that decides most deployments

Converting between the two price units is where comparisons usually go soft, so here is the conversion done out loud. A minute of speech is roughly 150 spoken words, which is roughly 200 tokens of output and perhaps 1,500 to 3,000 tokens of input once the prompt, the retrieved knowledge and the tool schemas are counted. At Qwen3.8-Max's $2/$6, that is under two cents a minute of model cost; at Qwen3.8-27B's $0.33/$2.40 it is a fraction of a cent.

So the per-token model is not the expensive part, and that is the trap. What you cannot get from Q​wen 3.8 at any price is a first syllable in 72 ms, because a 2.4-trillion-parameter flagship answering in eight seconds has already lost the caller. The cost of the hosted flagship is not the tokens; it is the round trip. Building the same thing yourself means owning the latency budget end to end — colocating the model with the speech stack, capping output length, streaming the first clause before the tool call resolves — and that engineering is the actual line item, not the $0.02.

Which is why the honest verdict is architectural rather than competitive. Tell 1's value is that someone has already spent the engineering on the first 800 milliseconds of a phone call. Q​wen 3.8's value is that it is measurably better at the part of the call where something gets done, and it is available both as a callable API and as weights you can host.

The split-stack pattern, and where a router earns its place

The configuration that resolves this is not a choice between the two models; it is a division of labour, and voice platforms already practise it internally. Retell's own homepage describes its stack exactly that way: a model for reasoning and orchestration, an agent harness that assembles knowledge and tools, and a separate speech layer. Tell 1 occupies one seat in that arrangement, not all of them.

Outside a voice platform, the same split looks like this. A small, fast model handles the turn-taking layer — the acknowledgement, the clarification, the "let me check that for you" — while the tool-heavy, reasoning-heavy steps go to a model with a better record on function calling and long context. Q​wen 3.8 fits the second role well and both of its main members are routable through OrcaRouter on one key: switch between Qwen3.8-Max and Qwen3.8-27B by changing the model string, keep a fallback chain so an upstream 429 does not become a dropped call, and use the routing DSL to send the hard turns to the flagship and the routine ones to the cheaper member. Provider list price passes through with nothing added, so a Q​wen price change reaches you the same day rather than at the next contract renewal.

Tell 1 is not on our catalogue and will not be; it is not sold as a model. What we can do is the escalation path around it — the model your voice agent calls when the caller asks something the script does not cover.

How to settle it in a week

Take fifty real calls from your own queue, not synthetic ones — the benchmark's weakness is that its calls are Retell's, and your failure modes live in your own transcripts. Run each through two configurations: Tell 1 inside Retell, and the same agent logic pointed at Qwen3.8-Max. Score three things and ignore everything else. Did the tool call fire with the right arguments? Did the answer match your knowledge base? And at what timestamp did the agent start speaking after the caller stopped?

My expectation, from the published numbers rather than from testing: the Q​wen configuration will win the first two and lose the third badly, and the size of that third gap is the only number that decides the architecture. If your calls are mostly lookups with a right answer, a voice-tuned model's latency advantage is worth less than a few points of tool-calling accuracy, and the general model wins. If your calls are mostly conversation — triage, reassurance, qualification — the tool calls are shallow, the latency is everything, and a model built for the phone is the right part for the job.

Both answers are defensible, which is why the "which is better" framing fails. The one thing I would not do is pick Tell 1 on the strength of a first-place rank without reading the column it is fifth in, or pick Q​wen 3.8 on the strength of an Intelligence Index without putting a stopwatch on its first token inside a live call.

A screenshot of the OrcaRouter model page for qwen3.8-max captured on 13 November 2026, showing a median time-to-first-token of 8.13 seconds and a listing dated 2026-08-03.A screenshot of the OrcaRouter model page for qwen3.8-27b captured on 13 November 2026, showing a 262K context badge, a median time-to-first-token of about 4.21 seconds and a listing dated 2026-08-13.