GPT Transcribe vs Whisper, Deepgram and Scribe

GPT Transcribe is OpenAI’s recommended starting model for new file-transcription integrations, and it shipped the way infrastructure usually does — as a model-card entry and a row on the pricing table, with no launch blog post and no named spokesperson. At the time of writing it lists at $0.0045 per minute of audio on OpenAI’s own pricing page, and an independent leaderboard scores it at 3.3% word error rate. Those two numbers, together, are the entire evaluation.

What makes this comparison worth running is that almost nobody publishes both halves. Vendor pages publish a price with no accuracy context. Independent leaderboards publish accuracy with a normalised price that frequently does not match what a developer sees on the vendor’s own checkout page. The gap between those two surfaces is not a rounding error: for one model in this comparison it is roughly 80%.

This guide sets GPT Transcribe against GPT-4o Transcribe, GPT-4o Mini Transcribe, Deepgram Nova-3, ElevenLabs Scribe v2 and open-weight Whisper across five hosting surfaces. Every price is labelled with the surface it came from and the date it was read. Where two sources disagree, both appear — never an average.

GPT Transcribe is a speech-to-text model built for three shapes of input: completed audio files, streamed transcripts of a file while it is still processing, and committed conversational turns inside a Realtime session over WebSocket. OpenAI’s model card describes it as supporting unstructured context, keyword hints, and multiple language hints — the mechanisms that let you steer it toward domain vocabulary and code-switched audio.

In practical terms, the API accepts three kinds of context alongside the audio. A prompt carries free-form context about the recording — the topic, the setting, who is speaking. A keywords list carries literal terms that may appear in the audio, such as product names, medications or acronyms; these are hints rather than required output. A languages list declares the languages you expect, which is what makes multilingual and code-switched audio tractable. When the model runs inside a Realtime session or a dedicated transcription session, it also carries earlier transcribed turns forward as context automatically, so a developer does not have to re-supply conversational memory on every turn.

The model was listed on OpenRouter on August 5, 2026. There is no accompanying launch post, no press coverage from the usual technology outlets, and no spokesperson statement — the closest thing to public discussion found during research was a low-signal submission of the model card itself to a developer news aggregator. That absence is worth noting up front, because it means every claim in this post comes from a vendor spec sheet or a benchmark table, not from marketing narrative.

One small finding is worth stating plainly, because it is rarer than it should be: OpenAI’s model card and OpenAI’s pricing page carry the same per-minute figure for GPT Transcribe, and the OpenRouter listing for the model shows the same figure again. Three surfaces, one number, no reconciliation needed. Most of the price comparisons in this post do not have that luxury.

OpenAI’s wider transcription lineup, read from the pricing page at the time of writing, prices as follows per minute of audio: gpt-transcribe at $0.0045; gpt-live-transcribe at $0.017; gpt-4o-transcribe at $0.006, which also carries token rates of $2.50 input and $10.00 output per million; gpt-4o-mini-transcribe at $0.003, with token rates of $1.25 and $5.00 per million; gpt-realtime-whisper at $0.017; and gpt-realtime-translate at $0.034. The transcription guide is explicit that the 4o transcription models and gpt-realtime-whisper remain available for existing integrations but are not the recommended starting models for a new one.

Legacy whisper-1 is a deliberate exception. It stays the recommended model for two capabilities GPT Transcribe does not cover: word-level timestamps with SRT and VTT subtitle export, and translating a completed recording into English through the audio-translations endpoint. Its standalone per-minute rate has aged off OpenAI’s current transcription pricing table, so this post does not print a figure for it. If subtitle export or translation is on your requirements list, price it directly with OpenAI rather than carrying an older number forward.

One more structural note before the table: speaker-labelled transcripts are not a GPT Transcribe feature. Diarisation is a separate specialised model, gpt-4o-transcribe-diarize. Teams that assume “transcription API” implies speaker labels tend to discover this late, usually after the pipeline is built.

The table below is the piece of work this post exists for. Each row carries the vendor’s own quoted price where one is published, a normalised cost for 100 hours of audio computed from that rate, and the independent word-error-rate and speed figures from the same leaderboard run. Prices and benchmark figures come from different organisations with different incentives, and the column headers say which is which.

Where a vendor’s public price and the benchmark-normalised price disagree, both appear as separate rows rather than being blended. The 100-hour column is arithmetic on the per-minute figure in the same row — 100 hours is 6,000 minutes — and it is there because per-minute rates in the fourth decimal place are almost impossible to compare by eye.

Three things fall out of the table immediately. First, GPT Transcribe is both cheaper and more accurate than the GPT-4o transcription model it supersedes — a 25% lower list price for a 0.7-point accuracy gain, which is why OpenAI’s own guidance now points new integrations at it. Second, nothing here is dominant: Scribe v2 posts the lowest error rate, Nova-3 posts by far the highest throughput, and the cheapest row is a hosted open-weight Whisper endpoint that is 2.4 points worse on accuracy than the best. Third, the two Scribe rows and the two Nova-3 rows are the same models priced by different observers, which is the subject of the next section.

For the self-hosted row, we deliberately do not restate a per-minute figure — the economics are hardware amortisation, not usage billing. Our earlier breakdown of the self-hosted Whisper breakeven math works that case in full, including the hardware assumptions and the volume at which a bought GPU beats a metered API.

ElevenLabs’ public speech-to-text FAQ states Scribe pricing as starting from $0.40 per hour of transcribed audio, falling below that at scale on Enterprise plans. The independent leaderboard normalises Scribe v2 at $3.67 per 1,000 minutes, which is approximately $0.22 per hour. Those are the same model. The retail figure is roughly 80% above the benchmarked one.

There is no scandal in that gap, and this post does not claim one. The most likely explanations are mundane: a different tier, a volume commitment, or a benchmarking-account rate not disclosed on a public FAQ. What matters operationally is that averaging the two produces a number that describes nothing. If you are budgeting against the public FAQ, budget $0.40 per hour and treat anything better as a negotiation outcome. If you are comparing accuracy-per-dollar across vendors on the leaderboard, use the leaderboard’s own price column for every row so the comparison stays internally consistent.

Deepgram’s divergence is smaller but more volatile. At the time of writing, Deepgram’s pricing page shows Nova-3 monolingual at $0.0048 per minute pay-as-you-go with the list price struck through above it at $0.0077, and Nova-3 multilingual at $0.0058 pay-as-you-go against a $0.0092 list. Pre-paid Growth-tier rates sit lower again at $0.0042 monolingual and $0.0050 multilingual. The leaderboard normalises the same model at $0.0043 per minute — about 10% below the current pay-as-you-go page rate. Our own snapshot from June 29, 2026 recorded $0.0043 per minute pre-recorded and $0.0077 streaming. The rate has moved inside a single quarter, and the vendor page explicitly labels current streaming rates as a limited-time promotion.

The pattern worth naming here is that speech-to-text pricing has started to behave like inference pricing generally: promotional entry rates, tier ladders, and a public retail figure that is a starting point rather than a price. That is a reasonable commercial move on the vendor side and a genuine planning problem on the buyer side, because a transcription workload is usually a fixed, predictable volume — exactly the kind of line item that gets budgeted once and then quietly re-prices underneath you.

Our expectation for the next few quarters is that the spread between benchmark-normalised and public retail rates widens rather than narrows, because the accuracy leaders have more room to segment by tier than the throughput leaders do. If that holds, the practical defence is to re-read every vendor page on a fixed cadence rather than at renewal, and to keep a second provider integrated and warm so that a promotional revert is a routing decision instead of a migration project.

The sharpest data point in this comparison is not about any closed vendor. It is about Whisper — the open-weight model most teams treat as a fixed, comparable baseline. On a single leaderboard run, the same Whisper Large v3 weights score 4.1% word error rate on one hosting provider and 10.1% on another. That is a 2.5× difference in reported error rate with identical model weights underneath.

The variance comes from the host, not the model: decoding parameters, chunking strategy, voice-activity handling, and how long files are segmented before they reach the model. Bars below are scaled to the widest value in the set, so the Replicate row fills the track and everything else is read relative to it. Lower is better throughout.

Two conclusions follow. The first is a buying conclusion: “we use Whisper” is not a specification. If accuracy matters, the hosting provider is as much of a decision as the model, and the cheapest host in this set is not the least accurate one — Groq’s Turbo endpoint is the cheapest row in the whole comparison at $0.67 per 1,000 minutes and still lands within half a point of the best Large v3 result. The most expensive Large v3 host here is also the least accurate and the slowest.

The second is an evaluation conclusion. If a set of identical weights can span six points of error rate depending on who runs them, then a vendor comparing its closed model against “Whisper” without naming the host and the decoding configuration has not told you enough to check the claim. That failure mode shows up across voice AI generally — we walked through a worked example of it in the Think Fast 2.0 speech-to-speech launch claims, where the headline accuracy comparison ran on the vendor’s own unaudited eval.

Word error rate is the industry’s default accuracy metric and it has a documented failure mode that changes how the whole table above should be read. AssemblyAI’s applied-AI engineering team published a finding in April 2026 that its own newer model was scoring worse than an older one on customer benchmarks — because the newer model was correctly transcribing words the human reference transcript had missed. Backchannels, quiet affirmations and overlapping speech that a human transcriber skipped were counted as insertions, so being more accurate than the ground truth registered as error.

This is not one vendor’s complaint about one dataset. Artificial Analysis, which produces the independent leaderboard used throughout this post, had to build proprietary manually-corrected ground-truth datasets for exactly this reason: public reference transcripts contain systematic human-transcription errors. When the benchmarking organisation and the model vendor independently arrive at the same conclusion about the reference data, that is about as close to confirmation as this field offers.

The metric has a second, quieter problem: it weights every error equally. AssemblyAI’s illustration is a good one — mishearing a car brand as a medical condition in a sentence about how something is doing counts as exactly one substitution, identical in weight to a trivial contraction mismatch. One of those breaks a downstream pipeline that feeds a language model; the other is invisible. A 0.3-point WER difference between two vendors tells you nothing about which of them makes the meaning-changing mistake more often on your audio.

The practical version of all this is simple and unpopular: build a small evaluation set from your own audio, transcribe it with two or three candidates, and read the errors rather than the aggregate. On a hundred representative clips you will learn more about which model fits your domain vocabulary than any leaderboard column can tell you. If you want the longer treatment of how to audit a vendor accuracy claim, our companion piece on how to read a WER claim without a published harness covers the methodology side in full; this post uses one section of it as the caveat on its own table.

For most production workloads the choice is settled by a capability requirement long before the third decimal place of the per-minute rate matters. Within OpenAI’s own lineup, four models cover four distinct jobs, and picking the wrong one is the most common early mistake.

Outside OpenAI, the feature sets diverge more sharply than the prices do. ElevenLabs launched Scribe v2 on January 9, 2026 as a batch transcription model, positioned around improved long-form audio stability, pause handling and tone-change handling relative to Scribe v1. It supports 90+ languages with per-language accuracy tiers published by the vendor: an “Excellent” tier at or below 5% error rate covering English, French, German, Spanish, Italian, Dutch, Polish, Portuguese and 26 others, then “High” at 5–10%, “Good” at 10–20% and “Moderate” at 20–50% covering the long tail. Publishing a per-language tier list is more transparency than most of this market offers.

Scribe v2’s feature set also carries keyterm prompting for up to 100 terms per its launch post, 56-category entity detection with timestamps for personal, health and payment data, speaker diarisation, word-level timestamps, automatic multi-language detection inside a single file, and a compliance list that includes SOC 2, ISO 27001, PCI DSS Level 1, HIPAA and GDPR, plus EU and India data residency and a zero-retention mode. A separate Scribe v2 Realtime variant targets live agents and meetings at under 150 milliseconds of latency.

Deepgram’s model is the opposite shape: a fast, cheap core with the extras itemised. At the time of writing its pricing page lists redaction at $0.0020 per minute, keyterm prompting at $0.0013, entity detection at $0.0017 and speaker diarisation at $0.0020, with smart formatting included at no charge. Stack diarisation and redaction onto the promotional monolingual base rate and the effective cost lands around $0.0088 per minute — above every OpenAI transcription row in the table. Deepgram also prices its conversational Flux model separately for real-time voice agents with built-in turn detection, at $0.0065 per minute for English pay-as-you-go against a $0.0077 list, and $0.0078 multilingual.

If your requirement is dictation rather than pipeline transcription — a person speaking into a text field, not a recording being processed — the trade-offs are different again, and the open-source options are stronger than most teams expect. We compared that stack separately in the open-source voice dictation roundup.

There is no overall winner in this comparison, which is the honest read of the table rather than a hedge. Each of the four common transcription workloads has a different constraint binding first.

The architectural recommendation that follows is routing rather than selection. Transcription is one of the few AI workloads where the output format is genuinely standardised, so swapping providers per workload class costs far less than it does for a language model. Keeping two providers integrated turns a promotional price revert into a configuration change. That is the same reasoning we apply when scoping AI transformation engagements — the evaluation is per-workload, and the deliverable is usually a routing policy rather than a single vendor choice.

Where transcription output feeds a downstream system — call summaries into a pipeline record, meeting notes into a deal timeline — the accuracy question stops being about the aggregate score and starts being about which specific errors survive into the record. That is a CRM automation design problem as much as a model-selection one, and it is where a single meaning-changing substitution costs more than a hundred harmless ones.