Serving inference for
Managed endpoints for language, documents and voice — plus the open-source libraries our own engineers use to run them. One API, one SDK, one invoice.
Chat, completion, embeddings and function calling on 40+ open-weight models — with streaming, batching, structured output and prompt versioning built in.
POST /v1/chat/completionsInvoices, IDs, contracts, handwriting, tables. Layout-aware models return typed JSON with coordinates — not a wall of text you have to re-parse.
POST /v1/vision/ocrNatural synthesis in 42 voices and 30 languages, streamed in under 300 ms, plus realtime transcription with word-level timestamps.
POST /v1/audio/speechOur engineers design the system with yours: model selection, guardrails, data flow, rollout plan and the cost model — then hand over the documentation.
SCOPED 4–12 week engagementsInference is the easiest part of an AI feature to get wrong in production. We handle the parts nobody wants to build twice.
Warm pools keep cold starts under 400 ms, and idle capacity scales to zero so staging costs nothing.
Server-sent events and WebSockets everywhere. Tokens, partial transcripts and audio chunks arrive as they are produced.
PII redaction, prompt-injection filters and regression suites that run before a prompt reaches production.
Per-key budgets, spend alerts and hard caps. Cache hits and batch jobs are billed at a fraction of realtime rates.
Token accounting, latency percentiles and model versions on every call — exported to your own OpenTelemetry collector.
First-class TypeScript, Python and Go clients, a CLI, and a self-hostable gateway that speaks the same API.
Six regions, burst to thousands of concurrent requests, and a rate limit you will only ever meet if you try to.
“We replaced three vendor SDKs and eleven million pages of OCR with one integration. The p95 latency chart stopped being a topic in our standups.”
Swap models without touching your code, keep the same auth, the same SDKs and the same observability — whether the request is a chat turn, a scanned invoice or a voice line.
Create an API key and make your first call in the playground. No card required for the free tier.
Use the typed SDKs or plain HTTP. Point at api.sapphire.studio and stream the response.
Add budgets, guardrails and evals as traffic grows — or lift the whole stack into your own VPC.
Start with a million free tokens a month. Metered usage after that, with hard caps you control.
For prototypes, side projects and evaluation.
$0 / month
Create an API keyFor products with real traffic and real SLAs.
$0.60 / 1M tokens
Start with 10M tokensFor regulated workloads and your own hardware.
Custom
Talk to an engineerEvery plan includes streaming, batching, TLS in transit, audit logging and unlimited API keys. Batch and cached requests are billed at 50% off.
Never. Prompts, documents and audio are processed for your request only. Standard plans retain inputs for 30 days for abuse monitoring, and Enterprise plans can switch to zero retention — we also sign a DPA and, where needed, a HIPAA BAA.
Yes. Enterprise customers deploy the same serving stack into their own VPC or on-prem cluster — same API surface, same SDKs, your GPUs or your reserved capacity. We ship Helm charts, container images and a reference architecture, and our engineers run the first rollout with your team.
On our public benchmark set of invoices, receipts and IDs, sapphire-ocr-v3 returns 99.2% character accuracy and 96.8% field-level accuracy. Every field comes back with a confidence score so you can route the uncertain 3% to a human review queue instead of guessing.
Voice cloning requires a verified consent recording from the speaker, is limited to your organization, and every synthesis is watermarked in the audio signal. We take down reported misuse, and cloning is disabled by default on the Developer plan.
You pay per million tokens, per page and per character of audio — no seat fees. Bursts are fine: we autoscale and bill the extra usage. To make invoices predictable you can reserve throughput, set per-key budgets and hard caps, and see spend alerts at 80% of any limit.
The SDKs, CLI and our self-hostable gateway are MIT licensed, and we publish the serving benchmarks we use internally. If you would rather run everything yourself, start with the gateway and bring your own models — the API stays the same.
No card, no sales call, no expiry. Create a key, copy three lines of code and see your first response stream back in under a minute.