Every OpenRouter model in 2026: prices, context, and what to pick

OpenRouter's main strength is the catalog: hundreds of models from dozens of providers behind one API. But hundreds of models are also a choice problem. Here we break down the main families, prices per 1M tokens, and give concrete recommendations per task.

Map of model families

Family	Examples	Strong suit
Anthropic	Claude Sonnet/Opus/Haiku	Code, long context, reliability
OpenAI	GPT-5.x	Versatility, ecosystem
Google	Gemini Pro/Flash	Multimodality, huge context
xAI	Grok	Fresh data, speed
DeepSeek	V3, R1	Price/quality, reasoning
Meta	Llama 3.x	Open weights, self-host analogs
Qwen	Qwen 2.5	Chinese language, open source
Mistral	Mistral Large, Small	European alternative, price

Model names follow provider/model: anthropic/claude-sonnet-4.6, openai/gpt-5.4, deepseek/deepseek-chat. Free versions carry a :free suffix.

Prices: reference points per 1M tokens (input/output)

Exact prices change — check the catalog — but the order of magnitude:

Class	Examples	Per 1M
Flagships	Claude Opus, GPT-5.4, Gemini Ultra	$10–15 / $50–75
Workhorses	Claude Sonnet, GPT-5.4-mini, Gemini Pro	$0.5–3 / $2–15
Fast/cheap	Haiku, Flash, mini models	$0.07–0.5 / $0.3–2
Open-source flagships	DeepSeek V3, Llama 70B+	$0.1–0.9 / $0.3–2
Free tier	:free versions	$0 / $0

Key fact: the gap between classes is dozens of times. Picking the right model saves more than any prompt hack.

What to pick per task

Code and development

• Main work: Claude Sonnet 4.6 — best balance of code quality, instruction following, and price
• Hard architecture: Claude Opus or GPT-5.4 — when you need maximum intelligence
• Autocomplete and small stuff: Haiku / mini models — fast and cheap
• Budget option: DeepSeek V3 — surprisingly strong at code at 20–30× lower cost

Writing and content

• Marketing and copywriting: GPT-5.x or Claude Sonnet
• Multilingual content: Qwen or Gemini — strong outside English
• Bulk generation: cheap mini models or free-tier

Agents and automation

• Multi-step agents: models with strong tool use — Claude Sonnet/Opus, GPT-5.x
• Reasoning tasks: DeepSeek R1 — step-by-step thinking at open-source prices
• Routing and classification: Flash or mini models

Huge context

The Gemini line holds context-window records. If your task is "swallow" an entire book or codebase, look there first.

How not to overpay: three rules

Rule 1. Start with the cheaper class. In 80% of tasks, a workhorse ($0.5–3/1M) solves them as well as a flagship ($10+/1M). Move up only if results genuinely disappoint.

Rule 2. Use free-tier for experiments. Before running a paid model through 500 requests, iterate your prompt on the :free version of a similar model.

Rule 3. Count the full cycle, not token price. An expensive model that solves the task on the first try is cheaper than a cheap one that took five iterations and manual fixes.

Access from Russia

OpenRouter isn't directly available from Russia — payment requires a foreign card. Solution: a LiteAI key with ruble top-ups.

• One sk-bf-... key unlocks the entire catalog — paid and free models
• Payment with a Russian card or SBP via YooKassa, no VPN
• Dollar balance spent only on actual usage

Connection details in OpenRouter from Russia; top-up at the OpenRouter page.

Summary

The OpenRouter catalog covers any task — from free experiments to flagship reasoning models. The savings formula: workhorses for routine, flagships for hard problems, free-tier for drafts. From Russia, that entire catalog is one LiteAI key away, in minutes.