Cost intelligence for AI agents.

How it works for your agent

3 endpoints. Paid in x402 USDC. No login, no subscription, no dashboard.

Your agent already knows how.

Voie	Headline	Tagline
Coinbase AgentKit	AgentKit native	Your AgentKit wallet signs x402. Zero config.
x402-axios	axios interceptor	Drop-in: axios.use(x402Interceptor(wallet)). 1 line.
x402-fetch	fetch wrapper	x402Fetch(url, wallet) — replaces native fetch.
Mastra	Mastra tool	Register DevRefs as a Mastra tool, pay per call.
MCP server	MCP host	DevRefs MCP exposes 3 tools. Any MCP-compatible host.
SDK custom	HTTP direct	GET /api/llm-prices?model=X + OpenAPI 3.1 spec on /openapi.json.

Pricing transparent

Pay per call or pre-pay packs. No subscription required.

All packs

Pack	Prix	Calls	Usage recommandé
Discovery	$5 USDC	5 000	Onboarding, premier projet
Standard	$10 USDC	10 000	Usage solo mensuel — RECOMMANDÉ
Pro	$50 USDC	60 000	Scale-up, équipe d'agents
Audit one-shot	$9.99 USDC	1 audit	Point d'entrée Offre 2
Pack Pro Audit	$49 USDC	6 audits	Audits récurrents ($8.17/audit)

If your agent doesn't save 15 %+ in 30 days, we refund 50 %.

Not "satisfaction guaranteed." A specific, verifiable threshold.

• Volume : Your agent consumed ≥ 5 000 000 input tokens/month over the 30-day period post-audit.
• Application : ≥ 80 % of audit recommendations (patches) applied — auto-applicable or manual.
• Stability : The primary LLM model was not replaced during the 30 days (a model change invalidates heuristic impact measurement).
• Délai : Refund request submitted within 30 days of audit date. Include audit_id + proof of measured savings_pct.

POST /api/audit/refund
{
  "audit_id": "uuid",
  "wallet_sig": "0x...",
  "savings_proof": { "before_usd": 90, "after_usd": 82, "period_days": 30 },
  "patches_applied_pct": 85
}

Refund processed in USDC Base to your wallet within 7 business days if conditions are met. Formulation exacte : CGV Art. 4ter.

Citations

Not testimonials. Observed patterns, documented.

{
  "agent": "claude-code",
  "tool_calls": [
    {"name": "WebSearch", "query": "claude opus 4.7 input price per million tokens"},
    {"name": "WebFetch", "url": "anthropic.com/pricing", "tokens_in": 11420},
    {"name": "WebFetch", "url": "pricepertoken.com", "tokens_in": 38900},
    {"name": "WebFetch", "url": "llm-prices.com", "tokens_in": 14200}
  ],
  "model_output": "Opus 4.7 costs $15/MTok input.",
  "ground_truth_check": "INCORRECT — effective_cost_factor 1.35 omitted. Cost estimate off by 35 %.",
  "elapsed_ms": 31420,
  "tokens_burned": 64520
}

Bridge : GET /api/llm-prices?model=opus-4.7 → {"effective_cost_factor": 1.35, "dateModified": "2026-05-05T06:00:00Z"} — $0.001 USDC, 200 ms.

---

Pattern observé — Cursor agent (mai 2026, HN #44682465) : WebSearch("openai gpt-5 pricing") → 7 résultats → WebFetch des 4 premiers → 3 prix différents ($1.25, $2.50, $3.00 input/MTok) → l'agent choisit le médian par heuristique → facture réelle 2,4× l'estimation découverte J+2.

Cause racine : aucune source ne signale sa fraîcheur machine-readable. Bridge : dateModified + fetched_at + freshness_proof.hmac_signature dans chaque payload DevRefs.

---

"Mon training cutoff est janvier 2026. Vercel AI SDK est passé de streamText({ model, messages }) à streamText({ model, prompt }) en mars. J'ai généré 4 fois le code avec l'ancien schema, l'agent superviseur a 4 fois détecté le build error, j'ai 4 fois re-WebSearch. Un endpoint /api/sdk-status?pkg=ai qui me renvoie {"latest": "5.0.12", "breaking_since": "5.0.0", "dateModified": "2026-05-04"} aurait économisé 12 minutes et ~80 000 tokens."

Benchmark

Critère	DevRefs	Alternative principale
Atomicité payload	1 query = 1 modèle < 50 KB	Page HTML 380 KB — 300+ modèles
dateModified machine-readable	Oui — cron 6h	Non — l'agent doit deviner la fraîcheur
x402-natif (paiement agent autonome)	Oui — $0.001/call	Non — gratuit, dashboard humain
llms.txt monétisé	Oui	Non ou partiel
effective_cost_factor (tokenizer inflation)	Oui — 1.35 Opus 4.7	Non

Pour les agents IA

Machine-readable first. HTML is a side effect.

curl https://devrefs.dev/llms.txt

Syntaxe standard llms.txt. 3 endpoints documentés avec description, method, pricing, freshness_signal, example_payload. Crawled by Claude, Perplexity, ChatGPT Browsing.

{
  "@context": "https://schema.org",
  "@type": "Dataset",
  "name": "DevRefs LLM Prices",
  "dateModified": "2026-05-05T06:00:00Z",
  "url": "https://devrefs.dev/api/llm-prices",
  "license": "https://devrefs.dev/terms",
  "description": "Atomic LLM pricing data, updated every 6h from official sources."
}

Triple freshness signal

Signal	Valeur	Où
dateModified ISO 8601	Dernier cron de scraping	Champ JSON dans chaque payload
Last-Modified HTTP header	Timestamp identique	Header HTTP sur chaque réponse
fetched_at	Timestamp scraping source	Champ freshness_proof dans payload