Simba 3.5 vs Eleven v4: The Two Leaderboards Disagree, and Only One of Them Is About Your Product

Put SpeechifyAI Simba 3.5 and ElevenLabs Eleven v4 on the same table and the result depends entirely on which Artificial Analysis board you read. On Controlled Voice — the same eight cloned reference voices for every entrant — Simba 3.5 is first at 1,192 and Eleven v4 is seventh at 1,155, a 37-point gap whose intervals do not touch. On Provider Voice — each vendor's own voice cast — Eleven v4 is first at 1,340 and Simba 3.5 is fifth at 1,294, a 46-point gap in the other direction, also with clean separation. Neither number is wrong. They are measuring different products, and the one you should trust is determined by a question about your stack rather than about either vendor.

The two boards, and why they disagree so completely

Controlled Voice holds the speaker constant. Artificial Analysis clones four American and four British voices and has every model say the same lines in those same voices, so a listener comparing two samples is hearing synthesis quality alone. Provider Voice lets each vendor put forward its own voices, which means the listener is judging the model and the voice cast together — the timbre, the accent, the acting choices that the vendor curated.

ElevenLabs has spent years building that cast. It is the part of the company that is not a model at all, and it is why the provider board has Eleven v4 and Eleven v4 Turbo in the top two slots with 32 samples of clear air between them and the rest. Take the cast away and Eleven v4 lands seventh. Speechify's position is the mirror image: fifth with its own voices, first without them.

• Controlled Voice — SpeechifyAI Simba 3.5 rank 1, Elo 1,192, interval 1,175 to 1,209, 1,783 samples

• Controlled Voice — ElevenLabs Eleven v4 rank 7, Elo 1,155, interval 1,143 to 1,167, 3,221 samples

• Controlled Voice — ElevenLabs Eleven v4 Turbo rank 6, Elo 1,156, interval 1,144 to 1,168, 3,327 samples

• Provider Voice — ElevenLabs Eleven v4 rank 1, Elo 1,340, interval 1,325 to 1,355, 2,742 samples

• Provider Voice — ElevenLabs Eleven v4 Turbo rank 2, Elo 1,339, interval 1,323 to 1,355

• Provider Voice — SpeechifyAI Simba 3.5 rank 5, Elo 1,294, interval 1,277 to 1,311, 1,732 samples

• Provider Voice — Alibaba Qwen-Audio-3.1-TTS-Plus rank 3 at 1,300, between the two on that board

Both gaps are larger than the intervals around them, so neither ordering is a coin flip. The 37-point Controlled Voice win for Simba 3.5 and the 46-point Provider Voice win for Eleven v4 are both real measurements of different things. That is unusual — most model pairs land the same way on both boards — and it is the whole content of this comparison.

The question that decides it

If your product ships whichever voice the vendor selected — a default assistant voice, a stock narration track, a single house voice across an app — you are buying what Provider Voice measured, and Eleven v4 wins that by 46 points. If your product lets users pick a voice, clone their own, or brings a fixed brand voice to every language, you are buying what Controlled Voice measured, and Simba 3.5 wins that by 37.

There is a practical test that resolves it in an afternoon: take three minutes of your own reference audio, clone it on both platforms, and listen. A model that leads the controlled board is a model that handles an unfamiliar voice well. A model that leads the provider board is a model with good taste in voices. Those are different skills and your workload only needs one of them.

Price is not close, and it is not the whole story either

Eleven v4 lists at $0.08 per thousand characters, which is $80 per million. That is the rate on ElevenLabs' own API pricing page and the figure Artificial Analysis carries for the model; the promotional window that ran through mid-October has ended, so $80 is the standing number rather than a launch artifact. Eleven v4 Turbo is half that at $40 per million.

Simba 3.5 lists at $15, $12 and $10 per million characters across Starter, Pro and Scale — a cut from the $20, $16 and $12 it launched at on 21 October. Through 21 November 2026 it bills at the older models' rates instead: $10, $8 and $6. After 22 November the new list applies and the offer does not.

• Eleven v4 — $80.00 per million characters, list, no active promotion

• Eleven v4 Turbo — $40.00 per million characters

• Simba 3.5 — $15.00 Starter, $12.00 Pro, $10.00 Scale per million characters, from 22 November

• Simba 3.5 — $10.00, $8.00 and $6.00 per million characters through 21 November

• Simba 3.2 and Simba 3.0 — unchanged at the $10, $8, $6 rates on the same plans

At five million characters a month — about a hundred hours of finished speech — Eleven v4 costs $400 and Simba 3.5 costs between $50 and $75 depending on plan and date. That is a factor of five to eight. Anyone quoting the leaderboard's $6.58 for Simba 3.5 is quoting a plan-embedded rate that does not match Speechify's published list; budget from the tier table above.

Languages: 48 under one id against 90-plus

This is where the comparison stops being a tie and starts being a routing decision. Simba 3.5 speaks 48 languages under a single model id, up from the six that Simba 3.0 covered, and every catalogue voice — including a cloned one — speaks all 48 with the same voice id. Estonian and Cantonese are explicitly not in the list. Eleven v4 covers 90-plus languages with a 10,000-character cap per request, which is roughly ten minutes of speech before you have to chunk.

Forty-eight languages is enough for most European, Latin American and East Asian product lines and not enough for a genuinely global catalogue. Ninety-plus is enough for almost anything, and if your requirement is "every locale we sell in," the quality argument stops mattering — one of these two can do the job and the other may not.

There is a caveat on the 48 that the vendor says out loud and most coverage skips: a voice keeps the accent of its source recording. A cloned American voice speaking Japanese will sound like an American speaking Japanese. Speechify's own guidance is to start from the staff-pick voice native to each language, which is why staff_pick_rank now appears in the voices endpoint. Cross-lingual voice transfer is available; it is not accent-free.

Latency, cloning and the API shape

Speechify reports 72 milliseconds time to first audio at the 95th percentile of streaming requests under production load, and 1.8 times Simba 3.2's throughput per serving instance. Those are vendor figures, unreproduced. ElevenLabs does not publish a comparable number for Eleven v4 itself. The one measured latency figure in the Simba documentation is older and narrower: Simba 3.2 at 56 milliseconds median and 102 milliseconds at the 90th percentile for server-side first byte, measured in US East on 15 September 2026 — a different metric from time to first audio and not transferable.

Voice cloning is zero-shot and self-serve on both, with a difference in friction that matters at team scale. Simba 3.5 cloning is available on paid plans without approval. ElevenLabs offers instant clones from ten seconds of audio alongside its professional clone tier, and its consent verification is the more developed of the two — a point in its favour for anyone shipping user-supplied voices.

The request shapes are not interchangeable, which is worth knowing before you write an abstraction layer. Speechify authenticates with a bearer token, sends the text as input, the model as model, and has no voice_settings object at all. ElevenLabs uses its own header and its own field names. Both support streaming; Speechify's SSML support is a subset, with break, prosody rate, prosody pitch, prosody volume and sub alias applied and emphasis and emotion tags accepted but ignored.

Where a router actually helps here

Neither model is on OrcaRouter, and this piece does not claim otherwise: the Simba line is reached through Speechify's own API and the ElevenLabs line through ElevenLabs'. What we do route is the smaller synthesis catalogue that makes a fair control group — openai/tts-1 at $15.00 per million characters and openai/tts-1-hd at $30.00, both at provider list price with nothing added on top.

The routing argument in this specific matchup is not about quality, because the quality answer is already split. It is that the two boards disagree, the prices differ by a factor of five to eight, and both facts are stable only until the next model ships. A stack that can send the same script to two vendors and compare on its own audio has an answer that survives the next leaderboard re-run; a stack hard-wired to one of them has to re-litigate the decision every quarter. Automatic failover also covers the case nobody benchmarks — a provider-side incident during a launch week.

An honest verdict, with the tie left in

Eleven v4 is the better model if you need broad language coverage, a curated voice cast you do not have to build, the most mature consent and cloning tooling, or a provider board lead that has held at the top of a hundred-model field. It costs five to eight times as much and the price is not promotional.

Simba 3.5 is the better model if your voices are yours — cloned, brand-owned, or chosen per locale — and if 48 languages covers the markets you actually sell in. It leads the board that holds the speaker constant, which is the board that predicts performance on a voice the vendor has never heard. It is also three weeks into a pricing schedule that has already changed once, with a model-id migration landing on 21 November, so the number you sign up at is not necessarily the number you will pay in December.

What neither board tells you is which one your listeners prefer on your content. Thirty-seven points of Elo is a meaningful gap and it is not a substitute for a listening test on three minutes of your own script in your own voice. Run that test before you commit volume to either.