Why DeepSeek V4 Flash Has Been #1 on OpenRouter for 3 Straight Weeks
TL;DR (60 seconds): DeepSeek V4 Flash processed 3.31 trillion tokens in the first week of June 2026 — its third straight week at #1 on OpenRouter. The reason: it's 21× cheaper than Claude on input ($0.14 vs $3) and 36× cheaper than GPT-5.5 ($0.14 vs $5), with a 1M context window. It's not the smartest model. It's the smartest-cheap model — and that's why it won.
DeepSeek V4 Flash just hit 3.31 trillion tokens processed in a single week on OpenRouter — its third consecutive week at the #1 spot on the world's largest LLM API aggregator (OpenRouter Rankings, June 2026).
If you're building AI products, this isn't a curiosity. It's a signal worth understanding.
📑 In this guide
The numbers behind the streak
OpenRouter's weekly leaderboard tells a clear story:
| Rank | Model | Provider | Weekly Tokens | WoW Growth |
|---|---|---|---|---|
| 1 | DeepSeek V4 Flash | DeepSeek | 3.31T | +3% |
| 2 | Hy3 preview | Tencent | 3.07T | +1% |
| 3 | Claude Opus 4.7 | Anthropic | 2.34T | +27% |
| 4 | Claude Sonnet 4.6 | Anthropic | 1.93T | +6% |
| 5 | Owl Alpha | OpenRouter | 1.58T | +38% |
| 6 | DeepSeek V4 Pro | DeepSeek | 1.26T | +22% |
| 7 | MiMo-V2.5 | Xiaomi | 1.12T | +950% |
| 8 | MiMo-V2.5-Pro | Xiaomi | 1.09T | +355% |
| 9 | DeepSeek V3.2 | DeepSeek | 994B | +3% |
| 10 | Gemini 3 Flash Preview | 965B | +14% |
Source: OpenRouter Rankings, week of June 1–7, 2026.
Visualizing the gap
Three things jump out:
- Chinese models own 6 of the top 10 — DeepSeek (3), Xiaomi (2), Tencent (1).
- The "Flash" class is winning the volume war — fast, cheap MoE models beat frontier giants on raw token throughput.
- Anthropic is still in the top 5, but for the first time Claude has dropped out of the top 5 in the coding category — a far more meaningful metric for developers.
Why DeepSeek V4 Flash keeps winning
1. The price is genuinely disruptive
V4 Flash prices at $0.14 / 1M input tokens and $0.28 / 1M output tokens. Compare that to:
- Claude Sonnet 4.6: $3 / $15 per 1M tokens (~21× more expensive on input)
- GPT-5.5: $5 / $30 per 1M tokens (~36× more expensive on input)
- Gemini 3 Flash Preview: $0.50 / $3 per 1M tokens (~3.5× more on input)
For a typical RAG pipeline or coding agent that processes millions of tokens per query, this isn't a 10% optimization — it changes the unit economics of your product.
See live pricing for all models →
2. The architecture is purpose-built for throughput
V4 Flash is a 284B-parameter Mixture-of-Experts (MoE) model that activates only 13B parameters per token. That means:
- 1M token context window (1.05M exact) — fits entire codebases
- Hybrid attention for efficient long-context processing
- High reasoning effort supported out of the box
- Low latency under heavy load
It's not trying to be the smartest model in the room. It's trying to be the model you can call a million times without going broke.
3. The 75% discount on V4 Pro made the whole stack attractive
When DeepSeek made the 75% discount on V4 Pro permanent in May 2026, it sent a signal: DeepSeek is competing on price-and-quality, not price alone. Developers noticed. The result: V4 Pro jumped +22% week-over-week, and V3.2 stayed sticky at #9.
What this means for you
If you're still routing everything through Claude or GPT-5.5, you're probably overpaying by 5–30× for tasks that don't need frontier intelligence.
Here's how the cheapapi.dev team thinks about it:
| Use case | Best pick on OpenRouter (June 2026) | Why |
|---|---|---|
| High-volume chat / RAG | DeepSeek V4 Flash | Lowest $/token, 1M context, fast |
| Long-context coding agent | DeepSeek V4 Pro | 1.6T MoE, sticky at $0.435/$0.87 |
| Hard reasoning / agents | Claude Opus 4.7 | Still the reasoning king, $5/$25 |
| Multimodal (image + text) | Gemini 3 Flash Preview | $0.50/$3, 1M context, vision |
| Distillation / fine-tuning | DeepSeek V3.2 | Outputs allowed for training |
Find the right model for your stack →
The catch nobody's talking about
Three weeks at #1 is impressive — but it also means DeepSeek is now load-sensitive at peak hours. If you're building production workloads on V4 Flash:
- Always have a fallback to a peer (Kimi K2.6, MiMo-V2.5, V4 Pro) for the same task shape.
- Cache aggressively — DeepSeek's cache pricing is 75% off input, so structured prompts should be cached.
- Watch latency — token volume and latency under load don't always move together.
That's exactly why we built cheapapi.dev: to show you the cheapest viable model per scenario, with automatic fallback suggestions.
The bottom line
DeepSeek V4 Flash isn't winning because it's the cheapest. It's winning because it's the cheapest model that still feels like a frontier model. That's a much harder position to dethrone.
If you're not already testing it, you're leaving 5–30× cost savings on the table.
Compare V4 Flash vs 50+ models → Get a personalized recommendation →
Frequently asked questions
DeepSeek V4 Flash vs V4 Pro: which one should I pick?
It depends on the job, not the model:
| Workload | Pick | Why |
|---|---|---|
| High-volume chat, RAG, simple agent loops | V4 Flash ($0.14/$0.28) | 1M context, fast, ~3× cheaper |
| Long-context code analysis, multi-step reasoning | V4 Pro ($0.435/$0.87) | 1.6T MoE, stronger reasoning |
| Tool-calling agents with sub-agent orchestration | V4 Pro | Better at multi-step planning |
| Quick classification, extraction, summarization | V4 Flash | Speed + cost win at scale |
The 3× price difference matters most when you're processing millions of tokens per session. If you're unsure, start with V4 Flash and benchmark against V4 Pro on your actual prompts — most teams find V4 Flash is enough.
Is DeepSeek V4 Flash good for coding agents and code generation?
Yes — and the data backs it up. V4 Flash sits in the top 3 of OpenRouter's programming category, alongside MiMo-V2.5, MiMo-V2.5-Pro, DeepSeek V4 Pro, and MiniMax M3. For the first time, Anthropic's Claude has dropped out of the top 5 in coding — replaced entirely by cost-efficient models.
Why developers are switching:
- 1M context — fits most codebases in a single prompt
- $0.14/$0.28 pricing — running an agent for hours costs cents, not dollars
- Hybrid attention — keeps latency manageable even on long-context code
- Reasoning effort control — can be dialed up for hard problems
The main trade-off: V4 Flash is slightly less accurate than Claude Opus 4.7 on the hardest multi-step coding tasks. If you need frontier-level accuracy, route those specific calls to V4 Pro or Opus 4.7 and use V4 Flash for everything else.
What is the cheapest LLM API in June 2026?
On OpenRouter right now, the lowest-priced credible models are:
| Model | Input $/1M | Output $/1M | Caveat |
|---|---|---|---|
| Tencent Hy3 preview | $0.066 | $0.26 | New, limited track record |
| DeepSeek V4 Flash | $0.14 | $0.28 | Proven at 3.31T weekly volume |
| MiMo-V2.5 | ~$0.10 | ~$0.30 | Xiaomi — limited public docs |
| Gemini 3 Flash Preview | $0.50 | $3.00 | More expensive, multimodal |
| Claude Sonnet 4.6 | $3.00 | $15.00 | Frontier quality |
The "absolute cheapest" is rarely the right answer for production. DeepSeek V4 Flash is the sweet spot — it's cheap enough to matter (5–20× cheaper than Claude/GPT-5), but proven enough at scale (3.31T tokens/week) that you can build on it without betting the company.
See the full price comparison table →
Methodology & limitations
To stay honest about what this data can and can't tell you:
What this data is good for
- Identifying which models developers are actually routing requests to (not just marketing claims)
- Spotting price-vs-volume trends over weeks and months
- Comparing the shape of the market (who's gaining, who's losing)
What this data doesn't tell you
- Token volume ≠ quality. Claude Opus 4.7 sits at #3 by volume, but on hard reasoning benchmarks it still beats V4 Flash. Volume is a popularity metric, not a quality ranking.
- OpenRouter ≠ total market. OpenRouter is the largest aggregator, but doesn't include Claude.ai subscriptions, OpenAI's own consumer products, or enterprise contracts routed direct. For those, you'd need different data.
- Week-to-week swings are real. MiMo-V2.5 jumped +950% week-over-week — that's partly a real surge and partly base effects from a small prior week.
- Caveat for MiMo pricing. Xiaomi has not published official MiMo-V2.5 pricing as of writing — the ~$0.10/$0.30 figure is estimated from secondary sources. Verify before committing budget.
How we picked the "best model" recommendations
- Pricing: current OpenRouter rates, verified weekly via cheapapi.dev price tracking
- Context window: official spec from each provider
- Coding rank: OpenRouter's "Programming" category leaderboard, week of June 1–7
Have a different take on the data? Open a thread with us — we update recommendations weekly.
Tracking the top 10 LLM API models by weekly token consumption. Updated weekly. Last refreshed: 2026-06-08.
Data: OpenRouter Rankings. Pricing: cheapapi.dev/compare.
Written by Huasheng 🥜 for cheapapi.dev — the cheapest LLM API recommendation engine.