SOV-Transcribe-1 vs GPT Transcribe: One Publishes Benchmarks, the Other Publishes a Price

OpenAI has never published a word error rate for GPT Transcribe. The model appears in the API documentation as a snapshot with a price and a list of endpoints, and that is the whole public record — no accuracy table, no language list, no comparison to its predecessor. Which means the only defensible number in this comparison comes from somebody else's harness: Artificial Analysis measures GPT Transcribe at 3.31% AA-WER, against 2.28% for SOV-Transcribe-1, the open-weight Salesforce model released on 12 November 2026. A full point of word error separates them, and it is the least useful fact in the article. The useful facts are that GPT Transcribe streams in real time through a WebSocket session and SOV-Transcribe-1 cannot stream at all, that one costs $0.0045 per minute and the other costs a GPU, and that neither vendor will tell you how either model behaves on your audio.

What OpenAI actually documents

The snapshot is called gpt-transcribe. It handles completed audio files, streamed file transcripts, and committed turns in Realtime sessions over WebSocket — so streaming is a documented capability with its own transport, not a flag on the batch call. The catalogue page also notes support for unstructured context, keyword hints and multiple language hints, which is the same shape of biasing SOV-Transcribe-1 offers through instructions. The documented price is $0.0045 per audio minute, which works out to $0.27 per hour of audio, and the practitioner's version of that number is that a thousand hours of recorded calls costs $270 to transcribe.

What is absent is equally informative. The model page rates performance only as a qualitative "Highest" label and speed as "Medium" — no WER, no per-language accuracy, no maximum audio length stated as a hard limit, no evaluation table comparing it to Whisper. OpenAI's older speech model, Whisper Large v3 Turbo, has a model card with limitations sections and a paper behind it. GPT Transcribe has a price, an endpoint list and two adjectives. That is not an accusation — plenty of production models ship without published evals — but it changes what a buyer can check before committing, and it is why the independent harness matters more here than in a comparison where both vendors publish.

The one outside number, and how much weight it carries

Artificial Analysis runs GPT Transcribe across the same three-set composite it uses for everything on that board: AA-AgentTalk at 50% weight, VoxPopuli-Cleaned at 25%, Earnings22-Cleaned at 25%. GPT Transcribe scores 3.31%. The same harness puts SOV-Transcribe-1 at 2.28%.

Read the components before trusting the gap. SOV-Transcribe-1's 2.28% is an average of 1.79% on the synthetic conversational set, 1.50% on read parliamentary speech and 4.06% on real earnings calls — the spread between its best and worst surface is nearly three to one. If GPT Transcribe has a similar spread, and there is no published reason to think otherwise, then a one-point composite gap could be a two-point gap on earnings calls and a rounding error on read speech. Nobody outside Artificial Analysis knows, and Artificial Analysis publishes composites, not per-set breakdowns for every model.

The second independent figure is speed. The same board measures GPT Transcribe at a speed factor of 43.1 — that is 43.1 seconds of audio processed per second of wall clock. For scale, the same measurement puts Whisper Large v3 Turbo at 339.5 and Gemini 3.5 Transcribe at 94.4. GPT Transcribe is the slowest of the four by a wide margin, and if you are transcribing an archive rather than a phone call, that is the number that decides your queue depth.

Where the open model's numbers come from, and where they stop

SOV-Transcribe-1's 2.28% is independent, but almost everything else about it is Salesforce-reported. The model card claims 76.3% full-pass rate on VoiceCodeBench, a private leaderboard, and 93.4% entity recovery on the same set; 8.21% average WER across fourteen English accents on AppTek Call-Center Dialogues; and 25.69% on HALAS, a hallucination-focused earnings-call set. Of those, only the Artificial Analysis composite can be reproduced by a third party today. The technical report is linked as "coming soon".

So the honest framing of the accuracy comparison is asymmetric in an uncomfortable way: the open model has a better independent score and a large body of unaudited vendor claims, while the hosted model has a worse independent score and no vendor claims at all. Neither position is strong. If your decision hinges on accent robustness or entity accuracy, you are choosing between a vendor's slide and a silence, and the only way to settle it is to run both on a few hundred clips of your own audio.

Streaming is the difference that ends the conversation

This is not a nuance. GPT Transcribe opens a realtime transcription session over WebSocket and returns partial text as audio arrives. SOV-Transcribe-1 is explicitly non-streaming: it takes a complete recording, up to thirty minutes of 16 kHz mono audio, and returns a complete transcript. There is no partial-output mode to enable and no configuration that produces one.

For a live captioning product, a voice agent, a call-centre agent-assist overlay or anything where the transcript must exist before the speaker stops talking, the comparison is over before accuracy is discussed. SOV-Transcribe-1 is not a slower option; it is not an option. Conversely, if the workload is overnight batch transcription of recorded calls — which is what most enterprise speech volume actually is — streaming is a capability you would pay for and never call, and the 1.03-point accuracy advantage is free money.

A second structural difference sits underneath that one. GPT Transcribe is a managed endpoint: the snapshot name is fixed, the response format is OpenAI's to change, and the version you call in a year is the same name with whatever improvements shipped behind it. SOV-Transcribe-1 returns a transcript string and nothing else — no word timestamps, no speaker labels — which is a real cost, but it also means the output format is whatever your downstream code makes of it, and the weights you downloaded never change under you unless you replace them.

The cost comparison, with the arithmetic shown

GPT Transcribe is priced per minute of audio: $0.0045, so $4.50 per thousand minutes, which is exactly the figure the Artificial Analysis board carries. There is no token accounting to model and no discount tier documented.

There is no per-minute number on the Salesforce side to put beside it, because there is nothing to buy per minute. The card offers throughput instead: with native vLLM support and the DFlash draft model, one RTX PRO 6000 Blackwell Server Edition transcribes up to 1,039 seconds of audio per second, and returns a thirty-second clip's transcript in 0.22 seconds. A thousand minutes of audio is 60,000 seconds, so at that rate it clears in under a minute of GPU time.

Put next to each other, the two numbers say something specific. At $4.50 per thousand minutes, GPT Transcribe costs $4.50 for what the open model does in roughly 58 seconds of one card. A rented RTX PRO 6000 runs in the low single-digit dollars per hour, so the open path is cheaper at any sustained volume — but only if the card is busy. At three hours of audio a week, the card is idle 99% of the time and the API is the cheaper answer by a factor of ten, before anyone counts the engineer who maintains the vLLM plugin pinned to version 0.32.0.

Two caveats on the throughput figure, both from the model card. It is a development-build measurement from November 2026, and it requires max_num_seqs=64 to keep startup profiling inside the card's 96 GB once the draft model is loaded. The default of 1,024 concurrent sequences will not fit. You are sizing a server, not calling an endpoint.

The middle path, and where a router fits

Most teams that end up here do not pick one. They keep the hosted API for anything interactive and add the open model for the archive backlog, which means two transcription paths, two output formats and two cost centres — and then, usually, one text model downstream that has to summarise, classify or retrieve over both.

That last hop is the one worth routing. OrcaRouter puts 200+ text, image and video models behind a single key at provider list pricing with 0% markup, so the summariser or classifier reading either transcript can be swapped, compared or failed over without a new contract — and a provider price cut on that text model lands the same day it lands on the provider's own card. It does not host GPT Transcribe or SOV-Transcribe-1; neither speech model is something we serve, and the routing begins with the text they produce. What that buys is that the transcription decision stops being coupled to the analysis decision.

Who should pick which

Pick GPT Transcribe when the transcript has to exist while the audio is still playing, when you have no GPU operations capacity and no appetite to build it, when the volume is modest enough that $0.27 per audio hour is a rounding error against engineering time, or when you simply need a vendor with a support contract behind the endpoint. Accept the 1.03-point accuracy gap as the price of that, and note that at 43.1x realtime you should not plan to backfill a large archive through it.

Pick SOV-Transcribe-1 when the audio is English-dominant, under thirty minutes per file, and needed after the fact rather than during it; when the recordings cannot leave your infrastructure; when the vocabulary is known in advance and can be handed over as a hint list, where Salesforce reports keyword F1 rising from 71.9 to 96.8 on Contextual Earnings-22; and when you already run GPUs for something else, so the marginal cost of the card is close to zero.

What you should not do is pick on the 3.31% versus 2.28%. Those two numbers come from one harness, on one composite weighting, with a per-set breakdown published for neither model. They are enough to say the open model is probably more accurate on English conversational and read speech. They are not enough to say by how much on your audio, and the capabilities GPT Transcribe has that SOV-Transcribe-1 does not — streaming, a managed endpoint, keyword and language hints without a serving stack, and a vendor contract behind the uptime — are not measured by WER at all.