codai

Radical transparency

What we actually paid per million tokens.

Computed live from the last 7 days of real traffic on ai.codai.ro. No list prices, no “starting from” footnotes — the blended cost users actually got. Other gateways treat their routing as a black box. We treat it as the product.

Overall blended cost · all tiers · last 7 days

$1.76/M tokens

From 34,087,924,549 tokens across 113,631 requests.

Per tier

Small

Chat, factual lookup, single-function code.

$1.78/M

0%

485 requests · 323,693 tokens · $0.58 total

Routes to: Claude Haiku 4.5 · Gemini Flash · Gemini Flash-Lite

Medium

Multi-step reasoning, code generation, refactors.

—/M

No traffic this week.

Routes to: Claude Sonnet 4.6 · GPT-5-mini · Gemini Pro · GPT-5.4-mini

Large

Frontier-grade architecture, long context, agentic loops.

$1.70/M

93%

112,677 requests · 32,854,772,205 tokens · $55907.51 total

Routes to: Claude Opus 4.7 · GPT-5 · GPT-5.5 · Gemini 3.1 Pro

Vision

Multimodal requests with image content — auto-routed to Gemini 2.5 Pro.

$5.84/M

0%

469 requests · 2,153,581 tokens · $12.58 total

Routes to: Gemini 2.5 Pro

Note: 1,230,675,070 tokens (10,844 requests, $4111.44) came from upstream models not yet labelled by tier (claude-haiku-4-5, claude-opus-4-8, claude-sonnet-5, …). Counted in the overall total above, not in any tier card.

How we keep the number honest

  • →Every request hits a classifier (Claude Haiku 4.5) that picks the cheapest tier that can handle it. We bias toward small. Requests with max_tokens ≤ 200 skip the classifier entirely and go straight to small.
  • →Opt in to FrugalGPT cascading with x-codai-cascade: verify: small tier answers first, a medium-tier verifier vets the reply, and we only escalate when the verifier rejects. Saves 40–60% on ambiguous prompts.
  • →Every response carries x-codai-routed-to so you can see which upstream actually ran. No black-box routing.
  • →Sticky-cache session routing: pass x-codai-session-id and follow-up turns land on the same upstream so the provider's prompt cache hits. Free latency + cost wins.

Numbers refresh every 10 minutes from the live usage_events table. Last computed 5 Sept 2026, 14:54 UTC.