codai
AI Command Center

One model.
Every model.

Codai is a universal AI model platform that routes every task to the best frontier model available — Claude, GPT-5, Gemini, and more. Top-tier quality by default. One API. One price. Zero model management.

< 80ms
Median latency
99.7%
Success rate
Cost savings vs. direct

Live Routing

Where codai routed your requests

Real production traffic — the actual distribution of upstream models the codai router selected. Updated every 5 minutes.

  • claude-fable-563.4%
  • claude-opus-4-834.6%
  • claude-sonnet-4-62.0%

5,247 routed requests · last 24 hours · successful `codai` requests only

Why Codai

Model intelligence, not just an API proxy

Most platforms forward your request to a hardcoded model. Codai runs a prior-driven decision engine that optimises every call for cost, latency, and quality.

Smart Model Selection

Classifies your prompt, checks historical success rates, and picks the optimal model tier. Haiku for quick fixes; Opus for deep reasoning — automatically.

Sub-80ms Dispatch

Codai runs a lightweight classifier locally before touching any cloud provider. Zero cold-start overhead on your inference path.

Benchmark-Driven Priors

Model selection is trained on real HumanEval, MMLU-Pro, and Aider Polyglot benchmarks. Prior success rates per model tier update weekly.

Predictable Pricing

Flat-rate task budgets. We don't sell tokens — we sell outcomes. Every task runs on the strongest model that clears the quality bar, with no per-token surprises.

Universal Model Roster

Claude 4 (Sonnet/Opus), GPT-5 (mini/nano/full), Gemini Pro, Gemini Flash — all behind one endpoint. Switch providers without touching your code.

Enterprise-Grade Security

Envelope encryption on all stored credentials. Zero plaintext secrets in DB. Audit-logged every request. GDPR-ready data residency in EU.

Live Benchmarks

Numbers that prove codai works

Every week, codai runs live benchmark suites against the full model roster — real runs on production infrastructure, not marketing slides. These numbers feed directly into model selection. No cherry-picking.

0.0%
HumanEval pass@1
Code generation accuracy on standard benchmark
0.0%
MMLU-Pro accuracy
Multi-task language understanding, professional
0.0%
Aider Polyglot
Agentic code editing across 10 languages
0.0×
Cost efficiency
Same frontier quality, fewer wasted tokens
0ms
P50 latency
First-token time including dispatch overhead
0.0%
Cascade escalation
Requests that needed a second, stronger model

Benchmarks run on production infrastructure. Updated weekly via automated worker. See full results →

Architecture

How it works

Three stages, sub-100ms overhead, fully observable.

⌨️1
Step 01

Your request arrives

Any client that speaks OpenAI or Anthropic API format. Point your base URL at api.codai.ro — no SDK changes.

🧠2
Step 02

Codai classifies & decides

A lightweight prompt classifier checks historical success rates per model tier. High-confidence? Skip LLM classifier entirely. Low-confidence? Run it in <15ms.

3
Step 03

Cascade guarantees quality

A verifier scores every answer against quality gates. If it falls short, a stronger frontier model takes over automatically — so you always get a top-tier result, transparently.

curl — api.codai.ro
$ curl api.codai.ro/v1/chat/completions \
-H "Authorization: Bearer $CODAI_KEY" \
-d '{"model":"codai","messages":[{"role":"user","content":"Fix this bug"}]}'
> x-codai-model-selected: claude-sonnet-4-6
> x-codai-decision-reason: prior_decision
> x-codai-confidence: 0.87
✓ 76ms — Bug fixed.

Model Roster

Every frontier model, one endpoint

Codai selects across Claude, GPT-5, and Gemini. As new models ship, codai benchmarks them and adds them to the pool. You never update a model name.

Anthropic
3 models
Claude Fable 5
Frontier model, autonomous coding
large
Claude Sonnet 4.6
Balanced performance & cost
medium
Claude Haiku 4.5
Ultra-fast, lightweight tasks
small
OpenAI
3 models
GPT-5
Frontier reasoning model
large
GPT-5 mini
Fast, cost-effective tasks
medium
GPT-5 nano
High-throughput micro-tasks
small
Google
3 models
Gemini 2.5 Pro
Long context, multimodal
large
Gemini Flash 2.5
Low latency, high throughput
medium
Gemini Flash 2.0
Sub-10ms, experimental
small

Use model="codai" in any OpenAI-compatible client. Codai handles the rest.

Live platform metrics

2.8M
Requests served
Total API calls routed
live
94.3B
Tokens processed
Input + output tokens
live
387.4KUSD
Cost saved
vs. always using Opus
live
9
Models live
Frontier models in pool

Pricing

Best model for everyone.

Every tier runs Claude Opus 4. Limits are on tasks, not on intelligence. No model lock-ins, no second-class inference.

Free
€0/mo
Try the real thing
  • Frontier models — same as paid
  • Full tool suite — same as paid
  • ~5 agent tasks / month to try
  • Codai smart router + sub-agents
Start for free
Student
€40/mo
Indie devs & learners
  • Frontier models — same as paid
  • Full tool suite — Python, shell, browser
  • ~33 agent tasks / month
  • Sub-agents
Get Student
Recommended
Builder
€100/mo
Daily agentic coding
  • Frontier models — same as paid
  • Full tool suite — Python, shell, browser
  • ~83 agent tasks / month
  • Persistent memory
  • BYOK support
Get Builder
Pro
€300/mo
Power users & teams
  • Every model — pin any provider
  • Full tool suite — Python, shell, browser
  • ~250 agent tasks / month
  • Persistent memory
  • Up to 100 seats
  • Weekly pay-as-you-go for overage
Get Pro

Full pricing details including team plans → /pricing

Security & Trust

Built for production from day one

We route sensitive AI workloads. Security is not a feature — it's the foundation.

GDPR
EU Compliant
TLS 1.3
Encrypted in transit
AES-256
Encrypted at rest
SOC 2
In progress

Envelope Encryption

Every API key is encrypted with a per-user data-encryption key, itself wrapped by a master key in GCP Secret Manager. Zero plaintext secrets in the database.

Zero-Trust Architecture

Gateway and management plane are separate Cloud Run services. Internal communication signed. No shared secrets between services.

EU Data Residency

All infrastructure runs in europe-west1 (Belgium). Your prompts and usage data never leave the EU. GDPR-compliant by design.

Immutable Audit Log

Every API call is persisted to an append-only usage_events table with model, latency, cost, and model decision. Full observability.

HMAC-Verified Keys

API keys are validated with constant-time HMAC comparison, immune to timing attacks. Keys can be scoped and revoked per session.

Open Telemetry

Prometheus metrics, structured logs, and distributed traces. See exactly what codai decided and why, in real time.