AI Model Rankings

We measured every AI model on cleaning up real dictation — quality, speed, and which ones to actually pick.

VivaDicta lets you point AI processing at a lot of different models, and a model name on its own tells you very little about whether it is any good at cleaning up dictation. So we measured them: every model below was sent the same real transcripts through the app's own prompt, and graded on what came back.

The short version: almost any current model is good enough. The top twenty score between 9.2 and 9.7, which is close enough to be a tie. What actually separates them is speed, and whether they quietly do something wrong.

Our picks

gemini-3.8-flash - the best results of anything measured, 9.7 quality at 0.9s.

gpt-5.4-mini - the same 9.7 at 0.9s, and one of the cheaper models OpenAI sells. Pick whichever account you already have.

grok-4.20-non-reasoning - 9.6 at 0.8s. With reasoning off there is no thinking pause before the text comes back, so it feels immediate even on long dictation.

qwen-3.8-27b on Cerebras - the quickest thing here, 9.2 at 0.3s, and what VivaDicta now selects by default on Cerebras. It replaces gemma-4-31b, which scored higher (9.6 at 0.29s) but which Cerebras retired from its public endpoints in September 2026, so it has been dropped from the table below. Cerebras asks for an account with a payment method before it hands out an API key, but you can delete the card again straight after signing up, so nothing can ever be charged - and the $5 signup credit it gives you lasts a very long time at dictation-sized requests.

mistral-medium-latest - free with nothing but an account: no card, no credit to spend down. 9.4 at 0.4s, which makes it the easiest of the free picks to get running.

gemma4:31b on Ollama Cloud - another free option. 9.4 at 0.6s.

claude-haiku-4-5 or gemini-3.1-flash-lite - if you already have credit with Anthropic or Google, both are sub-second and score 9.4-9.5.

How this was measured

Each model was given five raw dictation transcripts through the exact prompt VivaDicta sends, and graded on four things: whether it returned only cleaned text (rather than answering a question it was asked to clean up), whether every name and number survived, how well it removed fillers and stutters, and formatting. Obedience is weighted hardest - a model that replies to your dictation instead of tidying it is useless here, however good it sounds.

The cases deliberately target the hard parts: a speaker correcting themselves mid-sentence, a transcript that is entirely a question, technical terms and numbers, a spoken list, and non-English text.

Speed is the median of three runs. Value combines quality and speed as (2 × quality + speed) / 3, with speed on a curve - under a second is indistinguishable while you wait, and past about eight seconds you have already given up. Cost is not part of the score.

Full ranking

60 models, sorted by value - tap Quality, Speed or Value to re-sort. Graded by claude-opus-5 over 5 cases. Latest batch added October 3, 2026; older rows may have been measured earlier.

One row per model. Where a gateway resells weights already listed - the same claude-haiku-4.5 through Vercel AI Gateway and OpenRouter as well as Anthropic - only the best-scoring row is kept, so a long table is a list of things to choose between rather than the same model billed three ways. Two providers running the same open weights on their own hardware, like Cerebras and Groq with Qwen, both keep a row: those are separate measurements.

What the numbers do not tell you

These are 60 models measured for VivaDicta, not every model the apps offer. Newly measured models may appear here before an app update makes them selectable. A model missing from this table was either not measured or is the same weights as a row above, reached through a different gateway - neither is a verdict on it.

Scores within about 0.3 of each other are ties. Five cases graded by a single model is enough to separate good from broken, not enough to rank 9.6 above 9.5 with a straight face.

Speed depends on where you are. These were measured from one machine on one network. The ordering held up across runs; the exact seconds will not match what you see.

Two things models get wrong regardless of score. Several ignore the instruction to avoid the Russian letter "ё", and several use em-dashes despite being told not to. Both show up in models that otherwise score well.

Where this runs in the app

Pick a model under AI Providers - on macOS a shortlist is listed there directly: the best model from each provider, so the list shows what you could set up rather than repeating one you have already seen. Once a provider is configured you can set a different model per Viva Mode, so a fast cheap model can handle everyday dictation while something stronger handles the work that needs it.