GPT, Claude, Gemini, DeepSeek, GLM — and Seedance 2.5 video — routed to the real, official endpoints with public per-model multipliers you can audit. Change one line, keep your SDK.
46B tokens/month · 4K+ paying users · 100+ enterprises in production · 0 model swaps — public audit log ↗
Together and Segmind killed their free tiers; Novita is down to $0.50. flatkey still lets you run real models before you even sign in.
One model name; the gateway picks the best official model per request — task-aware, health-aware, always the lowest multiplier that meets quality.
No fp8 re-serves, no silent swaps. Every model probed against its official fingerprint, on a public log — 160+ models probed, every hour, on the record.
Prices in BRL, INR or CNY at checkout, 3DS-secured. Plus Anthropic-shaped API: point Claude Code or Cline at one base_url and go.
Claude Opus 4.8, $ per 1M input tokens after top-up bonus, measured across gateways on 2026-07-14.
Together, Novita and Segmind don’t carry official Claude at all. flatkey self-serve: 60–90% of list, bonus stacks to an effective 50% — and enterprise interfaces price below self-serve.
One request, five upstream channel classes. When a channel degrades, the router reroutes automatically — and the SLA is a public document, not a sales promise.
When one upstream errors, the router recovers on the next-best official channel — you see one flat line.
Models at 60–90% of list; the top-up bonus stacks to an effective 50%. One prepaid balance across every model.
claude-opus-4-8 · $/1M input tokens, after bonus
Every request hits the real lab API — no fp8 re-serves, no silent swaps. 160+ models probed against official fingerprints, on a public log.
Text, video, audio and image behind the same OpenAI-compatible gateway — and an Anthropic-shaped API for Claude Code.
client = OpenAI( base_url="https://router.flatkey.ai/v1") client.chat.completions.create( model="seedance-2.5", messages=[…]) # video, same SDK
Tree-structured sub-keys with budgets and model allowlists, a per-request ledger API, invoices in 48h — on audited infrastructure.
Every hour, every model is probed against its official API fingerprint — logits, tokenizer, refusal shape. The log is public. Cheap gateways ask for trust; flatkey shows receipts.
Per-request ledger with input, output and cached tokens — exportable, API-accessible. Invoices in 48h (VAT fapiao supported), 3DS payments, sub-key caps and model allowlists for teams.
OpenAI, Anthropic, Google, DeepSeek, Alibaba, Zhipu, Moonshot, ByteDance… new models live within 24h of official release — always the real endpoints, never quantized re-serves.
Text-to-video and image-to-video from every major lab — metered per second on the same prepaid balance as your text models.
Realtime voice, TTS and transcription from the official labs — same key, same ledger, priced per minute or per million characters.
“We route support and product-copy workloads across GPT and Claude through one flatkey balance. Same SDK we already had — roughly half the bill.”
“The hourly authenticity probes ended the ‘is this the real model?’ debate. Procurement gets one invoice; engineers get official endpoints.”
“Compliance was the blocker until flatkey: zero retention of request content, SOC 2 controls, and a per-request ledger. Our health-data reviews finally pass — and the bill dropped 47% in the first month.”
Create a flatkey account, open the console, and generate an API key for your app.
Point your existing SDK — OpenAI or Anthropic protocol, GPT, Claude, Gemini alike — at router.flatkey.ai. Nothing else changes.
Review usage, cost, routing and errors from the same dashboard — per model, per token, per request.
Integration, model choice, or billing — online support and usage consultation for everything on flatkey.
Whether you ship an agent today or route billions of tokens a month — one key, a signed SLA, and prices that drop as you grow.