xAI diarized transcription cost model

This note estimates the provider cost of using xAI Speech to Text with diarization for the first paid Dictate subscription, Dictate Pro.

For current model/provider direction, use ../TRANSCRIPTION_PLAN.md. This file is cost-focused and its older "xAI as default" language should not be read as overriding the current meeting rule: Meeting mode requires speaker-attributed output, and local Parakeet + dedicated diarization lanes must be benchmarked against hosted providers before a production default is chosen.

Research date: 2026-06-25. Prices are USD and should be rechecked before launch.

Currency assumption for AUD planning: 1 USD ~= 1.45 AUD, based on current USD/AUD rates around 2026-06-25. Recheck before pricing is finalized.

Sources #

• xAI pricing docs: https://docs.x.ai/developers/pricing
• xAI Speech to Text docs: https://docs.x.ai/developers/model-capabilities/audio/speech-to-text
• xAI STT/TTS launch post: https://x.ai/news/grok-stt-and-tts-apis
• Baruch College Tools for Clear Speech speaking-rate reference: https://tfcs.baruch.cuny.edu/speaking-rate/

Current xAI pricing #

xAI lists Speech to Text pricing as:

• REST/batch Speech to Text: $0.10 / audio hour
• Streaming Speech to Text: $0.20 / audio hour

xAI's Speech to Text docs expose diarization as a diarize=true request parameter. The public pricing page does not list a separate diarization add-on or surcharge. For planning, treat diarization as included in the STT audio-hour rate, but verify this in the xAI console before paid rollout.

The same endpoint also supports word-level timestamps, multichannel transcription, and text formatting. These are useful for meeting processing and speaker-attributed transcripts.

How batch works for meetings #

For xAI Speech to Text, "batch" means file-based REST transcription:

1. Dictate records or receives an audio file.
2. The backend sends the file to POST https://api.x.ai/v1/stt.
3. The request can include options such as diarize=true, format=true, and language=en.
4. xAI returns a transcript response with text, duration, and optional word-level metadata. With diarization enabled, each word includes a speaker identifier.

This is different from xAI's separate Batch API for text/chat models. The text Batch API can discount token-priced LLM calls, but the public STT pricing already has separate REST and streaming audio-hour prices. Do not assume the text Batch API discount applies to Speech to Text.

Batch/file STT is the right default for long meetings and recordings because the user does not need sub-second transcript updates after the meeting is already recorded. xAI's public STT docs list a 500 MB maximum file size. A normal one-hour compressed meeting recording should fit well under that limit; very long or high-bitrate recordings should be compressed or split into chunks before upload.

For long meetings, use this processing path:

• record locally,
• normalize/compress to an accepted format such as MP3, M4A, WAV, FLAC, or Opus,
• upload through REST STT with diarize=true,
• run summary/action-item post-processing after the transcript is returned,
• keep chunk metadata if the file had to be split.

Use streaming only when the user wants live dictation, live captions, or live meeting notes while the meeting is still happening.

Usage assumptions #

Dictate Pro should cover a normal user's regular dictation plus meeting and recording processing.

Baseline meeting/recording assumption:

• 3-4 one-hour conversations per week
• Monthly equivalent: 13.0-17.3 audio hours
• Best processing mode: REST/batch STT at $0.10 / hour, because recordings do not need live streaming latency

Everyday dictation assumption:

• Light: 15 min/day live dictation
• Typical: 30 min/day live dictation
• Heavy but still ordinary: 60 min/day live dictation
• Best processing mode: streaming STT at $0.20 / hour, if the user expects live transcript feedback

Speaking-rate assumption:

• Average English speech is roughly 150 words/minute.
• One hour of speech is roughly 9,000 words.
• That is enough text that meeting summaries/action items should be modeled as token-billed LLM post-processing, separate from audio transcription.

Direct STT cost #

Meeting and recording processing only:

Everyday live dictation only:

Combined target-user scenarios:

If all audio is sent through streaming instead of REST where possible, the expected target scenario rises to about $6.51/month before post-processing and overhead.

Meeting post-processing cost #

Meeting processing usually means more than transcription:

• clean transcript formatting,
• speaker labels,
• summary,
• decisions,
• action items,
• follow-up email or notes.

At 150 words/minute, a one-hour transcript is about 9,000 words. Depending on tokenizer and meeting style, plan around 10k-14k input tokens per meeting plus 1k-3k output tokens for summaries and action items.

Using xAI's current general chat pricing shown for Grok 4.3 ($1.25 / 1M input tokens, $2.50 / 1M output tokens), a one-hour meeting summary is roughly:

• Input: 12k tokens * $1.25 / 1M = $0.015
• Output: 2k tokens * $2.50 / 1M = $0.005
• Total: about $0.02 per one-hour meeting

For 13-17 one-hour meetings/month, this is only about $0.26-$0.35/month. Even after allowing for retries, richer prompts, and multiple output formats, budgeting $0.50-$1.00/month for meeting post-processing is conservative.

Recommended allowance #

For a paid subscription, a practical allowance is:

• include 35 audio hours/month for target users, or
• include 40 audio hours/month if the product promise should clearly cover 3-4 one-hour meetings per week plus 30 min/day of live dictation.

Estimated provider COGS for a 40 audio hours/month allowance:

• all REST/batch: $4.00
• all streaming: $8.00
• realistic mixed mode, meetings REST and dictation streaming: about $4.50-$6.00
• with meeting post-processing and retry overhead: about $5.50-$8.50

The paid plan can cover the target workload if meeting/recording uploads are processed with REST/batch STT and live streaming is reserved for actual live dictation. A soft cap at 40 hours/month, with warnings near 30 and 38 hours, should protect margin without surprising normal users.

If the market price is around $7/month, a 40 audio hours/month plan is possible but margin is thin. The expected target-user STT cost is about $4.78/month before payment fees, Store fees if applicable, backend costs, support, taxes, failed jobs, and heavy-user skew. At $7/month, prefer one of these safer shapes:

• 25-30 audio hours/month on the paid plan, with the 3-4 meetings/week target covered only when everyday live dictation is modest.
• 40 audio hours/month on annual billing or a higher monthly tier.
• A fair-use paid plan with warnings and throttling after 30 hours, plus paid hour packs for heavy users.

Dictate Pro at AUD 7 and 30% API markup #

Pricing target:

• Customer price: $7 AUD/month
• Markup rule: customer price = API cost * 1.3
• Maximum API cost budget: $7 / 1.3 = $5.38 AUD/month
• At 1 USD ~= 1.45 AUD: $5.38 AUD ~= $3.71 USD

At current xAI STT prices, that API budget buys:

If the plan is primarily a diarized meeting-processing plan, then A can be:

• 30 one-hour diarized meeting recordings per month,
• REST/batch processing only for those meetings,
• transcript with speaker labels, word timestamps where available, and formatted text,
• meeting summary/action-item processing included,
• live xAI dictation not included in the same allowance, or limited separately.

Cost of that A package:

• STT: 30 hours * $0.10 = $3.00 USD
• Meeting summaries/action items: roughly $0.60-$1.00 USD at this usage level
• Total estimated API cost: $3.60-$4.00 USD
• AUD equivalent at 1.45: $5.22-$5.80 AUD

The upper end is slightly above the strict $5.38 AUD budget, so the cleaner strict-margin package is:

• 25 one-hour diarized meeting recordings per month,
• meeting summaries/action items included,
• fair-use cap or paid packs above that.

Cost of the stricter 25 meeting package:

• STT: 25 hours * $0.10 = $2.50 USD
• Meeting summaries/action items: roughly $0.50-$0.85 USD
• Total estimated API cost: $3.00-$3.35 USD
• AUD equivalent at 1.45: $4.35-$4.86 AUD
• Charged at 30% markup: $5.66-$6.32 AUD

Recommended A for a $7 AUD/month Dictate Pro plan:

• 25 diarized one-hour meeting recordings per month included.
• This comfortably covers a target buyer with 1-2 important team meetings per week, which equals about 4-9 one-hour meetings/month.
• Keep an internal soft budget of 30 batch audio hours/month, but market the plan as 25 included meeting hours to leave margin for summaries, retries, exchange-rate movement, and unusually long files.
• Offer extra meeting packs at a clean marginal price, for example 10 extra meeting hours.

Qualification and positioning #

A useful pre-qualification question is:

How many important meetings do you have per week with your team?

The best-fit Dictate Pro buyer answers 1-2. That user has enough recurring meeting pain to value diarized transcripts and summaries, but does not create heavy transcription cost. At 1-2 one-hour meetings/week, expected monthly meeting usage is only 4-9 hours:

This means the included allowance can be generous without being the headline. Market the outcome instead:

• "Turn your important meetings into speaker-labelled notes."
• "Covers your weekly team meetings."
• "Includes 25 meeting hours/month."

Avoid leading with "25 hours" alone; high-volume users will anchor on the raw allowance and consume the margin. The qualification question should identify whether the buyer values the outcome, not whether they are shopping for bulk transcription minutes.

Do not include 30 min/day of live xAI streaming dictation in the same $7 AUD Dictate Pro package under the strict 30% API-markup rule. That alone costs about 15.2 hours * $0.20 = $3.04 USD/month, leaving only about $0.67 USD for batch meetings, or roughly 6 one-hour meetings before summaries. For this price point, everyday dictation should use local/on-device transcription by default, with xAI live dictation sold as a higher tier or add-on.

Future higher tier shape #

The first subscription is Dictate Pro, so this is not part of the initial launch. If a later higher tier is added, reserve it for users who answer 3-5 important meetings/week, or who want hosted live dictation in addition to meeting processing.

A future higher tier could include:

• 50 diarized meeting/recording hours per month through REST/batch STT,
• 10 live xAI dictation hours per month through streaming STT,
• meeting summaries, decisions, action items, and follow-up drafts,
• priority processing queue for uploaded meetings,
• professional exports,
• paid extra hour packs above the included allowance.

Diarization alternatives #

xAI is already close to the lower end of hosted diarized transcription pricing for one-hour meetings. Keep it as the default until a benchmark proves another provider is cheaper at equivalent accuracy, diarization quality, latency, and operational complexity.

Hosted alternatives to evaluate:

Local/open-source alternatives:

Local diarization has no per-minute API cost, but it is not free operationally. Expect higher support burden, model downloads, GPU/CPU variability, packaging weight, and lower reliability on noisy/overlapping meeting audio. It should be an experimental/private-mode feature until benchmarked.

Recommended benchmark order:

1. Keep xAI as hosted baseline.
2. Test Rev AI Reverb Turbo because it is price-competitive at $0.10/hr.
3. Test Google dynamic batch only if latency can be low priority.
4. Test Soniox if official pricing confirms $0.10-$0.12/hr all-in.
5. Prototype local faster-whisper + pyannote.audio Community-1.
6. Compare WhisperX as the heavier integrated local path.

Benchmark criteria:

• word error rate on real Dictate meeting samples,
• speaker diarization error and speaker-confusion rate,
• handling of overlap/crosstalk,
• processing time for a one-hour meeting on common laptops,
• install size and packaging risk,
• memory/VRAM usage,
• whether the output maps cleanly into Dictate transcript segments,
• total provider cost per successful one-hour meeting.

Product guidance #

• Use REST/batch STT for uploaded meetings and recordings by default.
• Use streaming STT only for live dictation or live meeting capture where the user needs immediate text.
• Track billable audio seconds server-side by subscription account.
• Store per-job provider mode (rest or streaming) so cost reporting is auditable.
• Expose subscription language as an allowance, not as "unlimited", unless there is a fair-use policy.
• Treat failed uploads/retries as part of gross margin planning. A 15-25% overhead buffer is reasonable before there is real telemetry.

Open checks before launch #

• Confirm in the xAI console or account terms that diarize=true has no separate surcharge.
• Confirm rate limits for expected upload sizes and concurrency.
• Confirm whether streaming billing rounds by second, minute, or session.
• Decide whether Dictate absorbs overage, hard-limits after the included hours, or sells additional hour packs.
• Reprice the model if xAI changes STT pricing before Store submission.