OpenRouter API: полное руководство по GPT, Claude и Gemini (2026)

OpenRouter — aggregation proxy между вашим HTTP client и inference backends. Один account → one API key → 70+ providers, 400+ models, ~100T tokens/month platform throughput (июнь 2026). Wire protocol = OpenAI Chat Completions: messages[], optional stream: true (SSE), max_tokens, tool calls где backend поддерживает.

Model ID всегда vendor/model-slug. Discovery: GET /api/v1/models → JSON array с pricing per 1M tokens. Рекомендуемые headers: Authorization: Bearer sk-or-..., HTTP-Referer, X-Title (для app leaderboard attribution).

• Key sprawl: rotation, quota, alert fatigue × N vendors; zombie cron job съедает Anthropic credits пока OpenAI key протухает.
• Hand-rolled fallback: try/catch вокруг трёх SDK — нет unified 429/503 telemetry, нет cost attribution per route.
• Proxy latency tax: extra hop добавляет типично 10–80 ms TTFT; для voice/realtime chat — dealbreaker; для batch agent — OK.
• Compliance gap: данные через US aggregator при DPA только с одним vendor.
• Host layer: Hermes gateway / Claude Code на shared macOS VPS — CPU steal, bandwidth jitter, SSE long-poll disconnect на multi-hour tool loops. Routing layer healthy, transport layer dead.

Free tier: 25+ models, 50 req/day без balance, 1000 req/day после $10 top-up. POC only — не для 7×24 agent fleet.

• Single SDK surface: swap base_url + model — zero client fork.
• Native fallback chains: declarative multi-model без custom orchestrator.
• BYOK economics: your direct keys, no token markup, 1M req/month platform quota free.
• Market telemetry: apps board + model rankings — см. июнь 2026 model share.
• Fast model A/B: Kimi K3, DeepSeek V4, Grok 4.5 — one-line switch без нового vendor account.

1. A
   Register на openrouter.ai, generate sk-or-..., optional BYOK для existing vendor keys.
2. B
   Export OPENROUTER_API_KEY в shell profile на dev machine или cloud node.
3. C
   Smoke test curl ниже → migrate SDK на https://openrouter.ai/api/v1.

curl — non-streaming completion:

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -H "HTTP-Referer: https://yourdomain.com" \
  -H "X-Title: Agent Gateway" \
  -d '{
    "model": "anthropic/claude-sonnet-5",
    "messages": [{"role": "user", "content": "Explain OpenRouter routing in 3 sentences."}]
  }'

Python requests (minimal deps):

import os, requests
resp = requests.post(
    "https://openrouter.ai/api/v1/chat/completions",
    headers={"Authorization": f"Bearer {os.environ['OPENROUTER_API_KEY']}"},
    json={
        "model": "openai/gpt-5.6-sol",
        "messages": [{"role": "user", "content": "ping"}],
    },
    timeout=120,
)
print(resp.json()["choices"][0]["message"]["content"])

Python OpenAI SDK:

from openai import OpenAI
client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key=os.environ["OPENROUTER_API_KEY"],
)
out = client.chat.completions.create(
    model="google/gemini-2.5-pro",
    messages=[{"role": "user", "content": "Summarize in Russian."}],
)
print(out.choices[0].message.content)

Node OpenAI SDK:

import OpenAI from "openai";
const or = new OpenAI({
  baseURL: "https://openrouter.ai/api/v1",
  apiKey: process.env.OPENROUTER_API_KEY,
});
const res = await or.chat.completions.create({
  model: "anthropic/claude-sonnet-5",
  messages: [{ role: "user", content: "test" }],
});
console.log(res.choices[0].message.content);

SSE streaming (Node):

const stream = await or.chat.completions.create({
  model: "openai/gpt-5.6-sol",
  messages: [{ role: "user", content: "stream me" }],
  stream: true,
});
for await (const ev of stream) {
  process.stdout.write(ev.choices[0]?.delta?.content ?? "");
}

Fallback payload:

{
  "models": [
    "anthropic/claude-sonnet-5",
    "openai/gpt-5.6-sol",
    "google/gemini-2.5-pro"
  ],
  "messages": [{"role": "user", "content": "critical path"}],
  "route": "fallback"
}

Model catalog introspection:

curl -s https://openrouter.ai/api/v1/models \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  | jq '.data[] | {id, pricing}' | head -20

BYOK mode: platform relays your OpenAI/Anthropic/Google keys — zero token markup, 1M requests/month free on OpenRouter side. Prepaid credits: displayed token price = provider rate; 5.5% fee только при purchase credits, не per-token.

• Metric 1: log model, provider, input/output tokens, wall time per request — correlate с generation ID из response headers.
• Metric 2: budget alert at 80% threshold до запуска Hermes 7×24 loop.
• Metric 3: IDE integration: OPENROUTER_API_KEY + base URL в Cursor/Claude Code — см. Agent Skills runbook.

• GSC: indexed pages per locale, query/impression mismatch, crawl errors on /ru/blog/.
• CDN/WAF: geo block или aggressive bot filter режет crawler из EU/US.
• hreflang: matrix только на blog index, не на detail pages (NUKCLOUD convention).
• Machine translation: thin/duplicate content penalty — rewrite per locale, не Google Translate pipeline.
• Backlinks: RU anchors с Habr, EN с Dev.to/HN — не zh-only link graph.

Правило: localize search intent, не transliterate zh skeleton. Один slug, восемь независимых тел в blog-data.js.

• URL pattern: https://nukcloud.com/{lang}/blog/2026-openrouter-api-guide-gpt-claude-gemini-20260724.html
• Canonical: self-referencing per detail page; no hreflang on articles.
• Sitemap: regenerate project script post-publish; GSC per property.
• Schema: BlogPosting + nested FAQPage в mainEntity.
• IndexNow: push new slug для Bing/Yandex indexers.

30-day KPIs: GSC impressions/CTR per locale, avg time on page > 4 min, clicks tseny / zakaz, console signups UTM ?utm_source=blog&utm_campaign=openrouter-api.

1. 01
   Account + keys: OpenRouter signup, sk-or-..., BYOK if existing vendor credits, budget cap at 80% alert threshold.
2. 02
   Local smoke: curl + OpenAI SDK; validate streaming SSE + 3-model fallback JSON.
3. 03
   Provision node: NUKCLOUD console — Mac mini M4 32GB+ для parallel sub-agents; region co-located с Git/registry.
4. 04
   SSH baseline: brew install node python@3.12, secrets in ~/.zshrc, RTT to openrouter.ai < 100ms from node.
5. 05
   Agent gateway: launchd plist для Hermes/Claude Code/custom service; SKILL.md для recurring prompts — Hermes install.
6. 06
   Bi-weekly review: OpenRouter $ vs node $, p95 latency, apps board; tune fallback chain + RAM tier на tseny.

Shared macOS VPS trade fast spin-up за bandwidth jitter, CPU oversubscription, SSE disconnect на thousand-tool-call sessions. Для OpenRouter-powered agent gateway 7×24 с auditable tenant boundary — multi-region bare-metal Mac cloud NUKCLOUD: unified memory, Seatbelt, launchd persistence. Hourly trial на zakaz, monthly commit после acceptance.