One endpoint · current frontier models
EU/RO AI, invoiced properly.
Point your tools at codai. It routes each task to one of Anthropic's current Claude models with automatic failover and checks that it actually worked. The flat success fee is charged only for verified successes; anything else costs at most the raw tokens.
Drop-in for any OpenAI-compatible client
# any OpenAI-compatible clientnpm i codai-sdkbase_url: https://ai.codai.ro/v1model: codai
Same request shape, one model name. No SDK rewrite.
Works in
How it works
You keep the client you already use. codai does the routing, the verification and the accounting.
Set the base URL to ai.codai.ro/v1 and the model to codai. Copilot, Cursor, Claude Code, Codex CLI or any OpenAI SDK — nothing else changes.
Each request goes to one of Anthropic's current Claude models, with automatic failover between approved deployments. The model that answered comes back in the x-codai-routed-to header.
A task is only billed when it is verified as successful — by tests, by the harness, or by your one-tap confirmation. Failed attempts never pay the flat success fee — at most raw token cost.
What you get
Everything an agentic workflow needs, behind one name.
Anthropic's current Claude models behind one name, with automatic failover. Every response names the model that served it (x-codai-routed-to), and the live distribution below shows where traffic went. Other providers’ models are available by name.
HTTP, Python, shell and browser automation run server-side in sandboxes — so any client can use them, even without native tool support.
Facts, decisions and repo conventions survive across sessions and devices. Scoped per user, exportable, deletable.
Attach your own Model Context Protocol servers once and every codai client gets the same tools.
Outcome billing: a fixed charge per verified successful task instead of a token meter you cannot predict; only unusually large tasks pay for the excess usage at token rates.
Use your existing Anthropic, OpenAI or Google keys through the same router and tooling, from €5 a month.
Gateway, database and backups run on Google Cloud in europe-west1 (Belgium). Model inference runs on Vertex AI and Azure AI Foundry, including US regions; no training on your data unless you opt in.
Public benchmark runs with open methodology, each one dated. Check the numbers yourself before you trust us.
Model roster
codai routes each request across Anthropic's current Claude models. Models from OpenAI, Google, xAI, DeepSeek and Moonshot are available by name on plans that allow them.
4 models
Claude Fable 5.1
In the codai route · via Google Vertex AI, Microsoft Azure AI Foundry · Released 1 Sept 2026
Claude Opus 5.5
In the codai route · via Google Vertex AI, Microsoft Azure AI Foundry · Released 22 Sept 2026
Claude Sonnet 5.5
In the codai route · via Google Vertex AI, Microsoft Azure AI Foundry · Released 28 Sept 2026
Claude Haiku 5.5
In the codai route · via Google Vertex AI · Released 7 Oct 2026
2 models
GPT-6.1 Sol
By model name · via Microsoft Azure AI Foundry
GPT-6 Astra
By model name · via Microsoft Azure AI Foundry
2 models
Gemini 3.8 Flash
By model name · via Google Vertex AI · Released 2 Sept 2026
Gemini 3.1 Pro(preview)
By model name · via Google Vertex AI
1 model
Grok 4.6
By model name · via Microsoft Azure AI Foundry
2 models
DeepSeek V4.1 Flash(preview)
By model name · via Microsoft Azure AI Foundry
DeepSeek V4 Pro
By model name · via Google Vertex AI
1 model
Kimi K2.6
By model name · via Microsoft Azure AI Foundry
Open-weight models that run on your device, inside the codai apps.
codai-nano
smallRuns on your Android phone in the codai app. Fine-tuned by codai from Gemma 4 E2B. Steps it handles locally never leave the phone; harder steps go to codai in the cloud.
Open-weight base: google/gemma-4-E2B-it
codai-mini
mediumRuns locally in the codai desktop app (beta). Version 1 is the stock Qwen3.6-35B-A3B, not fine-tuned. Nothing is sent to a server while you use it.
Open-weight base: Qwen/Qwen3.6-35B-A3B
Use model="codai" in any OpenAI-compatible client. The full list, filtered by what the gateway routes right now, is at ai.codai.ro/v1/public/models.
Benchmarks
Public benchmark runs over real task suites — coding, agentic tool use, long context. Same harness for every model, and every result shows the date it ran.
Pricing
Every plan gives you a monthly usage pool and verified successful tasks. Beyond them, €1.20 per verified success; unusually large tasks are billed for the excess at standard token rates.
Chat with Lite models and try one real task.
Free
1 task included
€1.20 per successful task beyond your plan
Everyday chat for little money.
€4.99/month
VAT included
€3 usage pool + no included tasks / month
€1.20 per successful task beyond your plan
Cancel anytime
For one developer shipping every week.
€19.99/month
VAT included
€6 usage pool + 10 verified tasks / month
€1.20 per successful task beyond your plan
Cancel anytime
For builders running agents daily.
€49/month
VAT included
€14 usage pool + 28 verified tasks / month
€1.20 per successful task beyond your plan
Cancel anytime
Have your own provider keys? BYOK from €5.99/month with a 7-day trial.
See all plansTrust
Your prompts, keys and usage data are handled the way you would handle them yourself.
Gateway, database, memory and backups run on Google Cloud in europe-west1 (Belgium). Model inference runs on Google Vertex AI and Microsoft Azure AI Foundry, including US regions, under the providers' data-processing terms.
Prompts and completions are never used to train models unless you explicitly opt in from your account.
Export everything we hold about you, or delete your account and data, from your account page — no support ticket needed.
Sign in with passkeys; mint API keys with per-key budgets and model scopes, and revoke them instantly.
No card needed. Point your client at codai and see where the router sends your first request.
Live routing
Real production traffic — the actual distribution of upstream models the codai router selected. Updated every five minutes.
9,940 routed requests · last 24 hours · successful codai requests only