Simba 3.5 Takes #1 in the Voice Arena by Two Elo Points, Which Is Not the Same Thing as Winning

SpeechifyAI Simba 3.5 is now first on Artificial Analysis' Controlled Voice arena at an Elo of 1,192, and the margin over HeyGen Voice at 1,190 is two points on a board whose confidence interval at the top is plus or minus 17. Third is Alibaba Qwen-Audio-3.1-TTS-Plus at 1,176. Read the leaderboard as a picture and Simba 3.5 has the crown; read the interval and four models are tied for it. Both readings are correct, and which one matters depends on whether you are writing a headline or choosing a vendor.

The more useful news about this model is not the rank. Simba 3.5 launched on 21 October 2026, and in the three and a half weeks since, Speechify has cut its list price, published the language list, and set a date — 21 November — on which two legacy model ids start being served by it. The arena position is the part everyone will quote. The dates are the part that changes a bill.

What the board actually says, in numbers you can check

Artificial Analysis runs two speech arenas and they measure different things. Controlled Voice clones the same eight reference voices — four American, four British — for every entrant, so the speaker is held constant and only the synthesis differs. Provider Voice lets each vendor use its own voice cast. A model with excellent default voices can win the second and lose the first. Simba 3.5 currently does the opposite of that.

• Controlled Voice — SpeechifyAI Simba 3.5, rank 1, Elo 1,192, plus or minus 17, interval 1,175 to 1,209, 1,783 samples, released November 2026

• Controlled Voice — HeyGen Voice, rank 2, Elo 1,190, plus or minus 13, interval 1,177 to 1,203, 2,626 samples, listed at $30.00 per million characters

• Controlled Voice — Alibaba Qwen-Audio-3.1-TTS-Plus, rank 3, Elo 1,176, plus or minus 12, interval 1,164 to 1,188, 2,968 samples

• Controlled Voice — HeyGen Voice Turbo, rank 4, Elo 1,174, plus or minus 15, listed at $4.00 per million characters

• Controlled Voice — ElevenLabs Eleven v4, rank 7, Elo 1,155, plus or minus 12, listed at $80.00 per million characters

• Controlled Voice — SpeechifyAI Simba 3.2, rank 14, Elo 1,082, plus or minus 12, released July 2026

• Provider Voice — SpeechifyAI Simba 3.5, rank 5, Elo 1,294, plus or minus 17, interval 1,277 to 1,311, 1,732 samples

• Provider Voice — ElevenLabs Eleven v4, rank 1, Elo 1,340, plus or minus 15

Two things follow immediately. The first is that the Controlled Voice ordering at the top is unresolved: Simba 3.5's interval spans 1,175 to 1,209, HeyGen Voice's spans 1,177 to 1,203, and Qwen-Audio-3.1-TTS-Plus sits at 1,164 to 1,188. Those three ranges overlap almost completely. The rank range column says the same thing in fewer characters — Simba 3.5 is listed as occupying ranks 1 to 3, not rank 1 alone.

The second is that this is not a fluke of a weak field. Simba 3.5 is 110 Elo points above its own predecessor, SpeechifyAI Simba 3.2, at 1,082, and that gap is wider than either model's interval. The generational step is real even though the lead over the field is not yet settled.

Why the top of this particular board is compressed

Elo is a relative scale, and this board is anchored at 1,000 by OpenAudio S1, a fixed reference entry that never moves. Everything else is placed against it by blind pairwise votes. That construction has a consequence people miss: as the field improves, the top does not spread out, it crowds. The distance from rank 1 to rank 14 on this board is 110 points. The distance from rank 27 (SpeechifyAI Simba 3.0 at 1,035) to rank 51 (OpenVoice v2 at 823) is 212.

When a hundred models are all being judged on the same eight cloned voices by listeners who cannot see labels, the differences that survive are small and the noise around them is large. A two-point lead on a board like this is a real data point about a real model — and it is not a product decision. Anyone who needs a decision has to add the things Elo does not measure: what the model costs, which languages it speaks, how fast the first audio arrives, and whether the API shape fits the code already written.

The price moved twice in three weeks

This is the part of the Simba 3.5 story that is genuinely recent, and it is entirely absent from launch coverage written in October. Speechify published the model at $20, $16 and $12 per million characters across Starter, Pro and Scale. On 9 November it cut the list to $15, $12 and $10. The launch offer has not changed: until 21 November 2026, Simba 3.5 bills at the rates of Simba 3.0 and Simba 3.2 — $10, $8 and $6 — and from 22 November it bills at the new list price rather than the old one.

Worked through against the plan balances Speechify publishes, that means a Pro month of 9 million characters costs $72 at the offer rate and $108 at list. A Scale month of 46.8 million characters costs $280.80 at the offer rate and $468 at list. The 9 November cut is what keeps that second number from being $561.60, which is what the original launch list would have charged.

One number on the leaderboard needs correcting before anyone budgets from it. Artificial Analysis lists Simba 3.5 at $6.58 per million characters — the same figure it shows for Simba 3.2 and Simba 3.0. That is a plan-embedded rate, not Speechify's published list, and it does not match the $10 to $15 range in the vendor's own changelog. If you are modelling spend, use the vendor's tier table; the board's price column is useful for ranking, not for a forecast.

What changes on 21 November

Speechify's two Simba 1.6 model ids, simba-english and simba-multilingual, stopped being selectable at API version 2026-09-21 — a new workspace that sends either gets a 400 model_retired. On 21 November the second half of that retirement lands: on every API version that can still name the old ids, both are served by Simba 3.5, in every language, and billed as simba-3.5.

The practical reading is that a set of integrations which have not been touched since September will change voice and change price on the same day, without a code change and without a warning beyond this changelog. Nothing breaks. The audio is different, and the invoice is denominated in a different model. Anyone running simba-multilingual in production should test simba-3.5 against their own content before that date, because the alternative is discovering the change from a support ticket.

Speechify's own numbers, labelled as such

Everything in this section is vendor-reported and has not been reproduced by anyone outside Speechify. It is worth reading, but it is not the same class of evidence as the arena figures above.

• Preference — 63% of blind side-by-side comparisons against Simba 3.2, with a stated 95% confidence interval of 60% to 67%, across all content types

• Latency — 72 milliseconds time to first audio at the 95th percentile of streaming requests, under production load

• Throughput — 1.8 times Simba 3.2 per serving instance, and 1.4 times Simba 3.0 Multilingual on average

• Coverage — 48 languages under one model id, up from six for Simba 3.0, with Estonian and Cantonese explicitly excluded

• Carried over unchanged — voice ids, output formats, speech marks, and the SSML subset that is actually applied: break, prosody rate, prosody pitch, prosody volume and sub alias

The 63% figure is the one to hold loosely. Blind preference against your own previous generation is the easiest comparison a speech vendor can run and the one least likely to predict a win against a different company's model — which is exactly what the arena above measures, and where the result is a two-point lead rather than a 63% one.

What this means if you are actually buying synthesis

Arena rank is an input, not a purchase order. Simba 3.5's case at the moment rests on three things that are unusually good together: a top-of-board quality position, 48 languages under a single id, and a price in the $10 to $15 range that most of the models above it on Provider Voice do not approach. ElevenLabs Eleven v4 is at $80. HeyGen Voice is at $30. The models that beat Simba 3.5 on price — HeyGen Voice Turbo at $4.00 — sit 18 Elo points lower on the same controlled board.

OrcaRouter does not carry the Simba line, and nothing here should be read as a claim that it does. What our catalogue does carry is a small set of synthesis models that make a useful control group when you are evaluating a switch: openai/tts-1 at $15.00 per million characters and openai/tts-1-hd at $30.00, both listed at provider list price with no markup added. Running your own script through those and through Simba 3.5 on the vendor's own API is a two-hour experiment that will tell you more about your content than any Elo column.

If the evaluation turns into a production decision, the argument for putting it behind a router is not the quality — it is the schedule. Simba 3.5 has changed its price once in three weeks, has a model-id migration landing on 21 November, and sits in an arena where the top four are statistically tied. Automatic failover means a provider-side change becomes a routing event rather than an outage, and a single key across vendors means the comparison you ran in November does not have to be rebuilt from scratch when one of them re-prices in December.

What to watch from here

Three things will settle this faster than another round of launch posts. Sample count is the first: Simba 3.5 has 1,783 Controlled Voice samples against HeyGen Voice's 2,626, and its interval will narrow as votes accumulate. If the two-point lead is real, it should survive that; if it is not, the rank range will drift back toward 1 to 3 and stay there.

The second is what Speechify does after 22 November, when the launch pricing ends and the $15, $12, $10 list takes over. A quality tier that costs 50% more than its predecessor is a different product decision than one that costs the same for a promotional window.

The third is the Provider Voice gap. Simba 3.5 is fifth there at 1,294 while ElevenLabs Eleven v4 leads at 1,340 — a 46-point deficit that does not appear on the controlled board, where Simba 3.5 is 37 points ahead. That split is the most interesting unresolved fact about this model, and it is the subject of its own comparison rather than a footnote here.