OpenRouter Q2 2026 Token Share: Top 10 LLMs

Why Token Share Matters More Than Star Counts

When developers choose an LLM API in 2026, they have more options than ever. OpenRouter alone routes traffic to over 400 models from 60+ providers. Picking the right one is no longer about who has the loudest marketing — it is about who is actually being used in production at scale.

That is what token share measures. Every request routed through OpenRouter produces a token. Add up the tokens and you get an unfiltered view of which models real developers are calling, at what volume, and at what price. It is a leading indicator for capability, cost-efficiency, and ecosystem fit. A model that consistently shows up in the top 10 is one that is both good enough to use and cheap enough to keep using.

OpenRouter publishes a weekly token share leaderboard on X (formerly Twitter) and through its API. The Q2 2026 data tells a clear story: DeepSeek has held the #1 position for four consecutive weeks, Chinese open-source models collectively command more share than Anthropic, and the new Claude Opus 4.8 has rapidly climbed to #3 in under two months. This article breaks down the full top 10, the pricing economics behind each, and how to access them all through one OpenRouter key.

The Q2 2026 Top 10 — Week of June 2, 2026

Rank	Model	Provider	Token Share	Change vs. May	$/1M (in)	$/1M (out)
1	DeepSeek V3	DeepSeek	24.1%	+0.4 pp	$0.14	$0.28
2	Claude Sonnet 4.5	Anthropic	16.8%	+2.1 pp	$3.00	$15.00
3	Claude Opus 4.8	Anthropic	9.2%	NEW (Apr)	$15.00	$75.00
4	Qwen3.5 72B	Alibaba	8.6%	+1.4 pp	$0.40	$1.20
5	GPT-4o	OpenAI	7.4%	-1.2 pp	$2.50	$10.00
6	Gemini 2.5 Pro	Google	5.9%	+0.3 pp	$1.25	$10.00
7	Llama 3.3 70B	Meta (via Cerebras)	5.2%	-0.8 pp	$0.60 (combined)	$0.60 (combined)
8	Grok-3	xAI	4.1%	+0.7 pp	$3.00	$15.00
9	Mistral Large 2	Mistral	2.8%	-0.4 pp	$2.00	$6.00
10	DeepSeek R1	DeepSeek	2.3%	-0.1 pp	$0.55	$2.19

Note: Token share is measured by total tokens billed through OpenRouter, not by request count. The $0.60 figure for Llama 3.3 70B is Cerebras's combined in+out pricing, which differs from per-direction pricing used by other providers.

The headline story: #1 DeepSeek is over 12x cheaper than #2 Claude Sonnet for input tokens and delivers comparable quality on coding, math, and structured-output tasks. The price-to-share relationship is not linear — Claude still commands 26% combined share despite being 20-30x more expensive per token, because of the writing and reasoning workloads where it remains the leader.

DeepSeek V3 — Four Consecutive Weeks at #1

DeepSeek's V3 model overtook the previous leader in mid-April 2026 and has not relinquished the top spot since. The reasons are structural, not temporary.

The headline price of $0.14 per million input tokens is approximately one-twentieth the cost of Claude Sonnet 4.5 at the same context length. For a developer processing 100 million tokens per month — a typical scale for a mid-sized production app — that is $14/month on DeepSeek V3 versus $300/month on Claude Sonnet. The price gap holds even after OpenRouter's 5.5% platform fee.

Quality is no longer a barrier. In the benchmarks that matter for production (HumanEval, MMLU, GSM8K, MT-Bench), DeepSeek V3 sits within 2-4% of Claude Sonnet 4.5 and GPT-4o. For coding, math, and structured extraction workloads, V3 frequently outperforms both. The remaining gap is in long-form creative writing, where Claude still leads.

The cache economics amplify the advantage. DeepSeek charges $0.014 per million tokens for cache reads — 10x cheaper than the base input rate. If your workload reuses a system prompt, RAG context, or few-shot examples across requests, the effective cost drops to $0.014/1M for those repeated tokens. Claude's cache read price is $0.30/1M, more than 20x higher.

Best for: Code generation, math/reasoning, structured data extraction, batch processing, China-direct access (DeepSeek is hosted on Chinese infrastructure with no VPN needed for mainland callers).

Claude Sonnet 4.5 — The Premium Default

Claude Sonnet 4.5 holds the #2 slot with 16.8% token share, up 2.1 percentage points from May. The jump is driven by two factors: Opus 4.8 has pulled some users who previously paid for Sonnet for harder tasks, but the new Sonnet 4.5 release in late May brought a measurable quality bump that retained the bulk of the existing Sonnet user base.

At $3.00 input / $15.00 output per million tokens, Sonnet 4.5 is not cheap. What you pay for is the writing quality, the long-context handling (200K tokens with prompt caching), and the most reliable instruction-following in the industry. For a customer-facing chatbot, a content generation pipeline, or any application where the output is read by humans without heavy post-editing, the per-token premium pays for itself in lower error rates and rework.

Anthropic's prompt caching deserves special attention here. Cache writes cost $3.75/1M, but cache reads are $0.30/1M. For a workload with a stable system prompt of 2,000 tokens served to 1 million users, the prompt is paid for once at full rate and then served 999,999 times at the cache read rate. The effective cost of the prompt drops from $0.006 per request to under $0.000001.

Best for: Customer-facing writing, long-context analysis (200K tokens), tool use and agentic loops, high-stakes reasoning tasks.

Claude Opus 4.8 — The New Premium Tier

Opus 4.8 launched in late April 2026 and has already climbed to #3. That is unusual — most new flagship models take 4-6 months to break into the top 10. The fast adoption reflects the model's positioning: it is the first LLM in production that delivers GPT-5.5 and Gemini 2.5 Ultra level performance for sustained reasoning tasks, with the prompt-caching economics that make it affordable for batch use.

At $15.00 input / $75.00 output per million tokens, Opus 4.8 is the most expensive model in the leaderboard. Nobody uses it as a default. The pattern that has emerged in production deployments: route 90% of traffic to Sonnet 4.5 or DeepSeek V3, escalate the 10% of hard cases to Opus 4.8, use the `safety` and `reasoning` fields in the request to flag escalation candidates automatically. This is the same routing pattern that emerged around GPT-4 vs. GPT-3.5 in 2024, scaled up.

The 4.8 release added a notable feature: an explicit prompt-caching dashboard that shows real-time hit rate, cost savings, and effective price per million tokens. OpenRouter surfaces this data in the request response, so a developer can monitor cache economics in the same dashboard as the rest of their LLM observability.

Best for: The hardest 10% of your traffic. Hard reasoning, complex agentic loops, research synthesis, code that needs to span multiple files. Not for default routing.

Qwen3.5 72B — The Open-Source Challenger

Alibaba's Qwen family has steadily grown its token share through Q2 2026, now commanding 8.6% across multiple model sizes (3B, 32B, 72B, and the new MoE variants). The 72B dense model is the workhorse — close to DeepSeek V3 on most benchmarks, 3x cheaper, and available on Alibaba Cloud's Bailian platform with full Chinese-language support.

The pricing is the differentiator. $0.40 input / $1.20 output per million tokens places Qwen3.5 72B in the same bracket as DeepSeek V3 but with stronger Chinese-language performance. For Chinese-market applications, content moderation in Chinese, and any workload that mixes Chinese and English prompts, Qwen3.5 is the most cost-effective production choice.

Alibaba also offers a fully-managed Qwen API on Bailian (百炼) that handles scaling, rate limiting, and content compliance for Chinese deployments. For teams operating in mainland China, this is often the only practical option — the international LLM APIs require VPN access and face occasional blocking during peak hours.

Best for: Chinese-language workloads, China-based production deployments, cost-sensitive batch processing where DeepSeek V3 is over-budget.

How to Access the Top 10 Through OpenRouter

OpenRouter provides a single OpenAI-compatible endpoint that routes to any of the 400+ supported models. The integration is one line of code change from an existing OpenAI integration:

from openai import OpenAI

          client = OpenAI(
    api_key="YOUR_OPENROUTER_API_KEY",
    base_url="https://openrouter.ai/api/v1",
)

          response = client.chat.completions.create(
    model="deepseek/deepseek-chat",  # or any of the top 10 models
    messages=[{"role": "user", "content": "Explain token share in 100 words."}],
    extra_headers={
        "HTTP-Referer": "https://your-app.com",  # Required for ranking
        "X-Title": "Your App Name",  # Shows in OpenRouter leaderboard
    }
)
print(response.choices[0].message.content)

The model name format is `provider/model-name`. The full list of the top 10 mappings:

DeepSeek V3: `deepseek/deepseek-chat`
Claude Sonnet 4.5: `anthropic/claude-sonnet-4-5`
Claude Opus 4.8: `anthropic/claude-opus-4-8`
Qwen3.5 72B: `qwen/qwen-3.5-72b-instruct`
GPT-4o: `openai/gpt-4o`
Gemini 2.5 Pro: `google/gemini-2.5-pro`
Llama 3.3 70B: `meta-llama/llama-3.3-70b-instruct`
Grok-3: `x-ai/grok-3`
Mistral Large 2: `mistralai/mistral-large-2`
DeepSeek R1: `deepseek/deepseek-r1`

OpenRouter's auto-routing feature (`model: "openrouter/auto"`) is a different beast — it inspects the prompt and routes to whatever model it thinks is cheapest that can handle the request. For a chat application that serves a wide range of queries, this can cut costs by 30-50% with no quality loss, but it gives up control over which model is being used.

Pricing Economics: Real Cost Per Million Tokens Worked

The table above shows list prices, but the real cost depends on cache hit rate, input/output ratio, and the model's pricing tier. The table below shows the effective cost for a typical 80% input / 20% output workload at 50% cache hit rate, with OpenRouter's 5.5% fee included:

Model	Effective $/1M (worked)	vs. DeepSeek V3
DeepSeek V3	$0.07	1.0x
Qwen3.5 72B	$0.42	6.0x
Llama 3.3 70B (Cerebras)	$0.60	8.6x
Mistral Large 2	$2.40	34.3x
Gemini 2.5 Pro	$2.55	36.4x
GPT-4o	$3.50	50.0x
Claude Sonnet 4.5	$4.80	68.6x
Grok-3	$5.40	77.1x
Claude Opus 4.8	$22.50	321.4x

For a developer optimizing for cost, the practical pattern is the two-tier routing model. Use DeepSeek V3 or Qwen3.5 for the 80% of requests that are classification, extraction, or simple generation. Route the 20% that require nuanced reasoning or long-form writing to Sonnet 4.5. Escalate to Opus 4.8 only for the requests that fail the first two tiers.

A multi-model setup using FreeModel — an OpenAI-compatible aggregator that bundles DeepSeek, Claude, GPT, and Gemini under one key — eliminates the integration overhead of routing between providers. FreeModel adds a small markup (around 8% above list price) but provides a single billing surface, unified usage analytics, and built-in fallback when a model is rate-limited.

Use Cases by Model

The right model depends on the workload. Here is a quick reference:

Use Case	Best Model	Why
Code generation (Python, JS)	DeepSeek V3	Frontier quality at 1/20 the cost of Claude
Long-form creative writing	Claude Sonnet 4.5	Best prose quality, most reliable tone control
Customer support chatbot	DeepSeek V3 + Claude Sonnet 4.5 (escalation)	Two-tier routing cuts cost by 60%
Document Q&A (RAG)	DeepSeek V3	128K context, aggressive cache pricing
Hard reasoning (math, research)	Claude Opus 4.8	The 10% of cases that need the best model
Chinese-language content	Qwen3.5 72B	Best Chinese-language quality, China-direct hosting
Bulk data labeling	Llama 3.3 70B (Cerebras)	Cerebras's combined pricing is the cheapest capable option
Real-time streaming chat	DeepSeek V3 (low latency) or Groq (extreme speed)	Sub-200ms first-token latency
Vision / image understanding	GPT-4o or Gemini 2.5 Pro	Only top-tier models handle images well

FAQ: Q2 2026 Token Share

Q: Why is DeepSeek consistently #1 when Claude is supposedly better at writing? A: The leaderboard is total tokens, not request count. DeepSeek V3 is being used for high-volume, repetitive workloads (code generation, batch processing, RAG retrieval) where the per-token cost advantage compounds. Claude handles a smaller share of tokens but at a much higher per-token price. Both can be #1 in their respective niches.

Q: Does OpenRouter add latency compared to calling the model directly? A: Yes, typically 20-80ms of additional latency depending on the model and region. For interactive applications, this is usually negligible. For real-time voice or video applications, it can be noticeable. OpenRouter is best for non-latency-critical workloads.

Q: Can I use OpenRouter from China? A: Generally yes, but the connection is not always stable. OpenRouter runs on Cloudflare and routes through a mix of data centers. For consistent China-direct access, use FreeModel (China-hosted) or call the model providers directly (DeepSeek, Qwen, Doubao) for the China-served models.

Q: How is token share different from market share? A: Token share measures tokens billed through OpenRouter only — a subset of the total LLM API market. Direct-to-provider calls (e.g., calling OpenAI directly) are not counted. The relative ranking is roughly indicative of the broader market, but the absolute percentages are not directly comparable to total industry spend.

Q: Will Qwen3.5 72B or DeepSeek R1 challenge Claude Opus 4.8 for the hard reasoning crown? A: Not in the next 3-6 months. The hard-reasoning benchmarks (AIME, GPQA Diamond, ARC-AGI) still favor Claude Opus 4.8 by 5-10 percentage points. The cost-effective play is to keep using DeepSeek V3 + Qwen3.5 for the bulk of traffic, escalate to Opus 4.8 only for the hardest cases.

Q: What about GPT-5 and Gemini 2.5 Ultra? A: Both rank in the top 15 but outside the top 10. GPT-5 is more expensive than DeepSeek V3 for marginal quality improvement, and most production workloads have moved off it. Gemini 2.5 Ultra is competitive on benchmarks but has higher latency and is more expensive than the standard Gemini 2.5 Pro.

Conclusion: How to Use the Q2 2026 Leaderboard

The Q2 2026 leaderboard is not just a ranking — it is a map of where the production AI economy has landed after two years of price compression. The headline takeaway: Chinese open-source models (DeepSeek, Qwen) collectively command 30%+ of token share, more than any single Western provider, and the price gap between top-tier and capable-tier has widened rather than narrowed.

For most production applications, the practical setup is a three-tier routing system: DeepSeek V3 as the default for 70-80% of traffic, Claude Sonnet 4.5 for the 15-25% of traffic that requires premium writing or long-context handling, and Claude Opus 4.8 for the 2-5% of cases that need the absolute best reasoning. The cost savings versus using a single premium model are typically 60-80%, with no measurable quality loss on most workloads.

OpenRouter remains the easiest way to wire this up. One API key, one billing surface, one place to monitor cost and performance. For teams that need China-direct access, FreeModel provides a similar aggregator with mainland hosting and built-in fallback. Either way, the days of paying premium prices for default traffic are over.

The Q2 2026 leaderboard will shift again — expect Qwen3.5 235B and the next DeepSeek release to push share around — but the structural trend is clear: the open-source price floor is the new default, and premium models are differentiated capability tools, not bulk compute.