Orchard vs Deepgram.

Deepgram's headline rate is $0.0036 per minute on Nova-3 batch (the cheapest tier), climbing to $0.0058 per minute for streaming. Orchard sits at $0.00042 per minute flat — 8.5× cheaper than Deepgram batch, 14× cheaper than streaming. The compound effect at sustained volume is significant:

The Deepgram Growth tier ($4K prepaid for 25% off) brings the effective rate down to ~$0.0027/min. Orchard PAYG still beats that by ~6.5×, and the Pro plan ($25/mo with 30k included minutes) brings the effective rate to ~$0.00083/min — still 3× under Deepgram's deepest committed discount.

Deepgram Nova-3 is a proprietary model trained from scratch on a mix of public + licensed audio; it consistently ranks near the top on English WER benchmarks. Orchard runs a tuned Whisper-large derivative (whisper.cpp + Core ML) — same model family OpenAI used for the original Whisper API, with our optimizations for throughput and Spanish.

• English (clean speech): Nova-3 typically wins by 1-2% absolute WER on US-accent corpora. Whisper-class is competitive.
• English (noisy / accented): Roughly parity. Whisper's larger pre-training corpus helps with accents Nova-3 saw less of.
• Spanish (neutral LATAM): Parity within ±1.5% WER. Both handle it well.
• Spanish (rioplatense): Orchard wins by ~4% absolute — we ship a fine-tune for porteño speech that nobody else has.
• Multispeaker diar: Both ship diarization. Our pipeline (pyannote 3.1 GPU) is included on every plan, Deepgram bills it as an add-on.

This is the dimension where Deepgram still leads. Their streaming WebSocket lands transcripts in <300 ms — best-in-class for real-time voice agents that need partial transcripts as the user speaks. Orchard's sync endpoint targets the same use case but via short HTTP requests (~150 ms p50 on a 5 s clip), not a persistent WebSocket; full streaming is on the roadmap.

• Real-time partial transcripts: Deepgram wins today via WebSocket.
• Sync STT (utterance complete): Roughly parity (~150 ms vs ~200 ms p50 on a 5 s clip).
• Batch (long-form): Orchard's cluster shards a 60-min audio across nodes — landing in ~90 s wall-time, ~40× real-time end-to-end. Deepgram batch is single-stream; the comparable 60-min job is ~3-4 min wall.
• Webhooks for async: Both support callback URLs on job completion.

Deepgram is primarily an STT vendor that added Aura (TTS) as a separate product with separate metering. Orchard ships three products on the same balance:

Orchard mirrors the OpenAI Whisper request/response shape. If you already wrap Deepgram in your own service layer, swapping the backend is a single base URL change in the OpenAI SDK:

// Deepgram
const dg = createClient(process.env.DEEPGRAM_API_KEY);
const { result } = await dg.listen.prerecorded.transcribeFile(
  fs.readFileSync("podcast.mp3"),
  { model: "nova-3", smart_format: true, diarize: true }
);

// Orchard (via OpenAI SDK — drop-in compatible)
const client = new OpenAI({
  baseURL: "https://api.orchardrun.com/v1",
  apiKey:  process.env.ORCHARD_API_KEY,
});
const transcription = await client.audio.transcriptions.create({
  file:  fs.createReadStream("podcast.mp3"),
  model: "whisper-1",
});

For diarized + async jobs, see /docs#async. We accept diarize=true and webhook_url=... as form params on the upload endpoint.

Three scenarios where Deepgram is the better pick today:

• Real-time streaming voice agents. If you need partial transcripts as the user speaks (sub-300 ms WebSocket), Deepgram's streaming endpoint is the best-in-class today. Our streaming is on the roadmap; until it ships, the sync HTTP endpoint is the closest we have.
• On-prem / VPC deployment. Deepgram offers a self-hosted variant on the Enterprise tier for customers with strict data-residency or air-gap requirements. Orchard is managed-only — your audio leaves your environment, hits our cluster, and the result comes back. Audio is dropped from RAM on response, never trained on, but it does transit.
• English-only US-accent corpora. Nova-3 wins by 1-2% absolute WER on clean US-accent English. If that 1-2% matters more than the 10× price delta (regulated industries, medical transcription with strict accuracy SLAs), stay on Nova-3.

For everything else — high-volume batch, multilingual workloads, anything Spanish-heavy, anything that needs TTS or voice cloning on the same bill — Orchard is the economically obvious choice.