Simba 3.5 vs Simba 3.2: 110 Elo Points, 48 Languages and a Bill That Doubles in Nine Days

The upgrade path inside one vendor is usually the least interesting comparison on this blog. This one is not, because SpeechifyAI Simba 3.5 and SpeechifyAI Simba 3.2 differ by more than a version number and the differences point in opposite directions. Simba 3.5 is 110 Elo points better on the Controlled Voice arena, speaks 48 languages instead of one, and costs 50% more at list. Simba 3.2 is faster on the one latency number anyone has actually measured, has a voice allow-list that no longer restricts anything, and keeps its old price. A team running Simba 3.2 in production has a real decision to make before 22 November, and "upgrade to the newer one" is not automatically the right answer.

The dates first, because they are what makes this urgent rather than academic. Simba 3.2 shipped on 8 July 2026 as an English-only streaming model and was recommended until 18 October. Simba 3.5 was marked recommended on 18 October, launched publicly on 21 October, had its list price cut on 9 November, and on 21 November becomes the model behind two retiring legacy ids. Speechify is not deprecating 3.2 — it stays listed, stays priced, and stays callable — and that is precisely why the choice has to be made on merits rather than on a deadline.

The arena gap is the largest in either model's history

Both models are scored on both Artificial Analysis boards. The controlled board holds the speaker constant across eight cloned reference voices; the provider board lets each vendor use its own cast. The gap between these two generations is larger on the controlled board, which is the more informative of the two for anyone cloning a voice.

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

• Controlled Voice — Simba 3.2 rank 14, Elo 1,082, plus or minus 12, interval 1,070 to 1,094, 3,630 samples

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

• Provider Voice — Simba 3.2 rank 12, Elo 1,250, plus or minus 11, interval 1,239 to 1,261, 3,966 samples

• Two generations back — Simba 3.0 rank 27 at 1,035 controlled, rank 41 at 1,114 provider

The controlled-board gap is 110 points with intervals that clear each other by more than 80 points of empty space. This is not a marginal re-run. For context, the entire top ten of that board spans 62 points; one generation of Simba covers nearly twice that range. The provider-board gap is smaller at 44 points, which says something useful: part of what 3.5 improved is synthesis quality on unfamiliar voices, and part of it is simply a better default sound.

The generational pattern is consistent across three releases, too. Simba 3.0 to 3.2 was 47 points on the controlled board; 3.2 to 3.5 was 110. Speechify is getting better at this faster than it was, and the 3.2 model that was recommended in September is now mid-table.

Languages: six, one, forty-eight

This is the difference that changes architectures rather than quality scores. Simba 3.2 is English only — a non-English voice returns a 400 with the rejected locale named in the message. Simba 3.0 covers six languages across seven locales. Simba 3.5 covers 48 under one id, with Estonian and Cantonese explicitly excluded and regional variants in scope.

What makes the 48 usable is that the voice travels. Every catalogue voice speaks all 48 on Simba 3.5 with the same voice id, and so does a voice you cloned on a paid plan. That is the feature a localisation pipeline has been waiting for: one cloned brand voice, forty-eight languages, no per-locale voice substitution. The vendor's caveat is that a voice keeps the accent of its source recording, and its guidance is to generate with a staff-pick voice native to each language — which is why staff_pick_rank appeared in the voices endpoint on 18 October.

On Simba 3.2, that whole workflow does not exist. One voice, one language. If your product ships in more than one market, this is not a scoring difference; it is the difference between one integration and two vendors.

Latency, where the older model still wins

Speechify publishes measured latency for Simba 3.2 and only aggregate claims for 3.5, and the published measurement is good. On 9 October 2026, timed from a client in US East over warm connections, Simba 3.2 reached first audible audio in a median of 47 milliseconds over the WebSocket input stream and 64 milliseconds over the HTTP stream — the lowest of the thirteen endpoints Speechify timed that day from that location. Server-side first byte was 56 ms at p50 and 102 ms at p90 on production traffic measured 15 September. The company is explicit that every value is a measured percentile rather than a guarantee, and that from Zurich the picture changes: 142 ms for another vendor's flash model against 145 and 153 ms for its own two stream paths.

For Simba 3.5 the documented figures are vendor claims rather than published percentiles: 72 milliseconds time to first audio for 95% of streaming requests under production load, and 1.8 times Simba 3.2's throughput per serving instance. Those cannot be compared directly with a p50 measured over warm connections. A 72 ms p95 and a 47 ms p50 may be entirely consistent, and nobody outside Speechify has run the round that would establish it.

The practical consequence is narrow but real. If your product's user experience is a voice assistant where a hundred milliseconds of silence is perceptible, you have a measured number for 3.2 and a claim for 3.5, and the honest thing to do is measure 3.5 yourself before migrating. If you are rendering audio into a file, latency is irrelevant and this section does not apply to you.

Price: the offer ends on 21 November

Speechify's pricing for these two models is about to diverge, and the changelog lays it out precisely.

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

• Simba 3.5, through 21 November — $10.00, $8.00 and $6.00, the rates of the older models

• Simba 3.5, original 21 October launch list — $20, $16 and $12, superseded by a 9 November cut

• Simba 3.2 — unchanged at $10.00, $8.00 and $6.00 per million characters on the same plans

So through 21 November the two models cost exactly the same and 3.5 is strictly better. From 22 November, 3.5 costs 50% more at every tier. On the Pro plan that is $72 versus $108 for nine million characters; on Scale, $280.80 versus $468 for 46.8 million. The balances are unchanged, so the same plan buys a third fewer characters of the newer model at list.

That is a straightforward value judgment rather than a trick: 110 Elo points and 47 extra languages for a 50% premium is a good trade for almost any workload that uses either feature, and a bad one for an English-only pipeline that was already satisfied with 3.2's quality.

Migrating is one line, and that is the trap

Speechify's documentation is unusually clear that moving from 3.2 to 3.5 requires no API version change: set model to simba-3.5, keep the same voice ids, the same output formats, the same speech marks, and the same applied SSML subset. Requests that omit model still resolve to simba-3.0, so the default does not silently move you.

Two details complicate the one-line story. The first is that language now matters: if the text differs from the voice's locale you must set it explicitly, or the API uses the voice's locale and you get an accent you did not intend. The second is 21 November itself. On that date the two Simba 1.6 ids, simba-english and simba-multilingual, stop being served by their old training and start being served by Simba 3.5, billed at Simba 3.5's rate, on every API version that can still name them. An integration pinned to the old ids will change voice and change price on the same day without a code change. Testing 3.5 deliberately before then is the only way to choose.

There is also a quiet improvement worth knowing about: on 20 October, prosody pitch and prosody volume started being applied on Simba 3.5, 3.2 and 3.0, where they had previously been accepted and ignored. If you send those attributes today and did not want the change, Speechify's own advice is to remove them. Emphasis and emotion tags remain accepted and unapplied on all three models.

Where OrcaRouter fits into an in-family upgrade

Neither Simba model is on OrcaRouter — the line is reached through Speechify's own API, and nothing here should be read as a claim otherwise. What we do route is the synthesis catalogue you would use as a control while deciding: openai/tts-1 at $15.00 per million characters and openai/tts-1-hd at $30.00, at provider list price with nothing added.

The useful thing a router does in a migration like this is make the old path survivable while the new one is proven. Run 3.5 on the vendor's API for a week of real traffic, keep 3.2 as the fallback, and route around an incident rather than discovering it in production — automatic failover costs nothing when nothing fails and buys the whole launch window when something does. The same single key also means the OpenAI voice you were already using stays reachable while you decide whether either Simba generation replaces it.

Which one to run, by workload

Stay on Simba 3.2 if your workload is English, your voices are English, your quality bar was already met in July, and the measured 47 ms first-audible-audio figure is load-bearing in your product. You are paying the lower rate through 21 November and forever after, and nothing in 3.5 improves your output. There is no deprecation date on 3.2; Speechify has said so explicitly.

Move to Simba 3.5 if any of the following is true: you need a second language, you clone voices and want one voice id across locales, your English quality bar has moved, or you are building something new. The 110-point controlled-board gap is the largest generational jump Speechify has produced, and the 48-language coverage is not available any other way inside this family. Do it before 21 November if you want to evaluate it at the older models' rates, and set the model id explicitly so the retirement does not choose for you.

Whichever way you go, the decision has a review date rather than a permanent answer. Simba 3.5 has 1,783 controlled-voice samples against 3.2's 3,630, so its interval will narrow. Its list price has already moved once in three weeks. And Alibaba's Qwen-Audio-3.1-TTS-Plus, at 1,176 on the same board, sits 16 points behind the current Simba flagship at a comparable board price — close enough that the next release on either side reopens the question.