OpenRouter API: полный гайд по GPT, Claude и Gemini — routing, код и production (2026)

OpenRouter — aggregation layer над inference endpoints: один API key + OpenAI-compatible /chat/completions вместо отдельных интеграций с OpenAI, Anthropic, Google, Meta, DeepSeek.

• Endpoint: https://openrouter.ai/api/v1/chat/completions
• Auth: Authorization: Bearer $OPENROUTER_API_KEY
• Протокол: OpenAI Chat Completions — миграция = смена base_url и api_key
• Model slug: provider/model — openai/gpt-4o, anthropic/claude-3.5-sonnet, google/gemini-2.5-pro, deepseek/deepseek-chat

Ключ к пониманию latency и failover — два независимых routing hop внутри gateway:

Provider Routing в деталях: один slug (anthropic/claude-3.5-sonnet) может резолвиться в несколько upstream endpoints (официальный Anthropic, reseller, regional mirror). Gateway выбирает candidate по price curve, health bit и rate-limit state — без изменения client payload.

Model Routing + Fallback chain: поле models: [...] с route: "fallback" задаёт ordered list. При 429/5xx/timeout L1 переключается на следующий slug; L2 внутри каждого slug повторяет provider failover. Client-side circuit breaker не нужен.

• Free tier: 25+ моделей; ~50 req/day без balance, ~1000/day после ≥$10 (cap 20/min).
• Pricing: token rate = provider list price; top-up 5.5% (min $0.80); crypto +5%; BYOK: first 1M req/month free, then 5% on excess.
• Extra hop latency: типично +10–80 ms TTFT — измерять под ваш prompt size и region.

• Account fragmentation: N провайдеров → N keys, billing consoles, audit trails.
• Schema drift: message format, streaming SSE deltas, error codes — каждый SDK свой.
• Single-provider SPOF: 429 на одном upstream без fallback = hard fail для user-facing agent.
• Observability gap: TTFT, throughput, spend размазаны по пяти dashboards.

Routing overhead оправдан для multi-model A/B и agent prototypes; для single-model hyperscale и sub-10 ms TTFT budget — стройте direct path или hybrid (OpenRouter только для fallback tier).

• One key, all slugs: model swap = string replace в model field.
• Cross-provider failover: declarative models array, sequential L1 routing.
• Unified billing/Observability: один dashboard для token, cost, TTFT across slugs.
• No token markup: только top-up fee; BYOK снижает effective rate на high QPS.
• Clear scope: prototyping, model shootouts, monthly spend до low thousands USD.

Direct API предпочтительнее:

• Single slug, monthly spend $10k+ — 5.5% fee окупает свой routing layer
• Нужны Anthropic Prompt Caching, OpenAI Batch/Assistants, Google Vertex hooks
• TTFT SLA не допускает +10–80 ms gateway hop
• Data residency: payload не должен проходить US middleware

1. Register: openrouter.ai — GitHub или email.
2. API key: Keys page; per-project key + spend limit в production.
3. Env var: export OPENROUTER_API_KEY="sk-or-..." — не коммитить в git.
4. Smoke test: curl или OpenAI SDK; verify HTTP 200 и resolved model id в response.
5. Model catalog: curl https://openrouter.ai/api/v1/models -H "Authorization: Bearer $OPENROUTER_API_KEY"
6. Deploy: fallback chain, streaming, HTTP-Referer/X-Title (rankings attribution), key на 7×24 agent host.

Runnable snippets после подстановки key.

cURL

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-3.5-sonnet",
    "messages": [{"role": "user", "content": "Объясни квантовые вычисления одним предложением"}]
  }'

Python (requests)

import requests, os
r = requests.post(
  "https://openrouter.ai/api/v1/chat/completions",
  headers={"Authorization": f"Bearer {os.environ['OPENROUTER_API_KEY']}"},
  json={"model": "google/gemini-2.5-pro",
        "messages": [{"role": "user", "content": "Quicksort на Python"}]}
)
print(r.json()["choices"][0]["message"]["content"])

Python (OpenAI SDK migration)

from openai import OpenAI
client = OpenAI(
  base_url="https://openrouter.ai/api/v1",
  api_key=os.environ["OPENROUTER_API_KEY"],
)
completion = client.chat.completions.create(
  model="openai/gpt-4o",
  messages=[{"role": "user", "content": "Hello!"}],
  extra_headers={"HTTP-Referer": "https://jexcloud.com", "X-Title": "JEXCLOUD Demo"},
)

Node.js (OpenAI SDK)

import OpenAI from "openai";
const openai = new OpenAI({
  baseURL: "https://openrouter.ai/api/v1",
  apiKey: process.env.OPENROUTER_API_KEY,
});
const completion = await openai.chat.completions.create({
  model: "deepseek/deepseek-chat",
  messages: [{ role: "user", content: "Explain OpenRouter in one sentence" }],
});

Streaming (SSE token deltas)

const stream = await openai.chat.completions.create({
  model: "anthropic/claude-3.5-sonnet", stream: true,
  messages: [{ role: "user", content: "Короткое осеннее стихотворение" }],
});
for await (const chunk of stream) {
  const c = chunk.choices[0]?.delta?.content;
  if (c) process.stdout.write(c);
}

L1 fallback routing JSON

{
  "model": "anthropic/claude-3.5-sonnet",
  "models": ["anthropic/claude-3.5-sonnet", "openai/gpt-4o", "google/gemini-2.5-pro"],
  "route": "fallback",
  "messages": [{"role": "user", "content": "Hello"}]
}

• Catalog scale: 70+ providers, 400+ models — verify via GET /models before routing rules.
• Free tier: 25+ models; 50 req/day no balance, 1000/day after $10, rate cap 20/min.
• Fees: top-up 5.5% (min $0.80); crypto +5%; BYOK 1M req/month free tier.
• Gateway latency tax: +10–80 ms — benchmark under your payload size.
• Cost control: default route to Flash/cheap slugs; Opus only on critical steps; per-project spend limits.

OpenRouter не накручивает token list price — total cost = top-up fee + gateway hop + compliance boundary, не «model price × mystery multiplier».

OpenRouter платный? Free + paid slugs; paid at provider token rate, top-up 5.5%.

Доступен из РФ/СНГ? API endpoint глобальный; оцените network path и data compliance самостоятельно.

Какие модели? 400+ в формате vendor/model; full list /api/v1/models.

OpenRouter vs Claude direct? Multi-model + failover: OpenRouter; exclusive APIs + hyperscale: direct.

Monthly cost? Зависит от slug и QPS — Flash agents часто $10–100, full Opus load может $1000+.

Security? Payload через gateway; sensitive data: BYOK или direct API.

Python call? Секция 05 — requests или OpenAI SDK.

P0 — Crawl/index:

• CDN/WAF blocks Googlebot — Search Console URL Inspection
• Missing hreflang — one language wins canonical
• robots.txt / noindex on /en/
• sitemap without per-lang URLs or alternates
• CSR empty shell — crawler sees no body

Content layer:

• Literal translation vs native queries (OpenRouter vs OpenAI API)
• No EN keyword research; weak E-E-A-T

Authority:

• CN distribution with backlinks; EN pages zero Reddit/HN/dev.to exposure

1. GSC: EN URL indexed?
2. CDN/WAF vs Googlebot
3. hreflang, canonical, lang sitemap
4. Rewrite 3–5 EN posts (not translate)
5. First EN distribution: dev.to, Reddit, Hacker News

RU: OpenRouter API, tutorial, OpenRouter vs OpenAI; long-tail: pricing, free models, Python example, fallback routing.

EN: OpenRouter API tutorial, is OpenRouter worth it, fallback routing, streaming response.

Title signals RU: полный гайд, пошагово, 2026 — keyword в title, lead, H2.

Principle: shared code blocks; per-lang title, lead, H2, FAQ; meta 120–160 chars; balanced «when NOT to use» section.

URL pattern (subdir, shared domain authority):

https://jexcloud.com/ru/blog/2026-0724-openrouter-api-tutorial-gpt-claude-gemini-guide.html
https://jexcloud.com/en/blog/2026-0724-openrouter-api-tutorial-gpt-claude-gemini-guide.html

hreflang (cross-ref + x-default):

<link rel="alternate" hreflang="ru" href="https://jexcloud.com/ru/blog/..." />
<link rel="alternate" hreflang="en" href="https://jexcloud.com/en/blog/..." />
<link rel="alternate" hreflang="x-default" href="https://jexcloud.com/en/blog/..." />

Per-lang canonical self-reference; sitemap xhtml:link alternates.

Schema: BlogPosting + FAQPage JSON-LD; FAQ questions in target language.

Distribution RU: Habr, Telegram tech channels; EN: dev.to, HN, Reddit r/LocalLLaMA.

P0: GSC index check → CDN/WAF → hreflang/canonical/sitemap.

P1: localized posts → keywords → Article + FAQPage schema.

P2: distribution → GSC sitemap → impressions/CTR per lang path.

OpenRouter agent на личном Mac: lid closed = connection drop. Oversubscribed VPS без native macOS: SSH jitter рвёт multi-step tool loops. Shared team machine: key rotation и CLI version drift — production blockers вне scope API integration.

Для команд с 7×24 Cursor Agent, OpenClaw Gateway и OpenRouter routing стабильнее JEXCLOUD multi-region bare-metal Mac: dedicated Apple Silicon, authentic macOS, 120s provisioning, monthly flex. Model billing через OpenRouter, compute через bare metal — clean separation. Specs: pricing, onboarding: help center.