Welcome to the March edition of our monthly LLM rankings. Each month, we pull real-world usage data from OpenRouter — a platform routing millions of API calls across every major LLM — and rank the models developers are actually paying for.
Not benchmarks. Not press releases. Real tokens, real money, real usage.
The Top 10 — March 2026
| Rank | Model | Provider | Move | Input / 1K tokens | Output / 1K tokens |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 4.6 | Anthropic | ↑ | $0.003 | $0.015 |
| 2 | Claude 3.5 Sonnet | Anthropic | ↓ | $0.003 | $0.015 |
| 3 | GPT-5.4 | OpenAI | 🆕 | $0.002 | $0.01 |
| 4 | Gemini 3 Flash | ↑ | $0.0001 | $0.0004 | |
| 5 | GPT-5.2 | OpenAI | ↓ | $0.00175 | $0.014 |
| 6 | DeepSeek V3 | DeepSeek | ● | $0.0003 | $0.0009 |
| 7 | Gemini 3.1 Pro | 🆕 | $0.00125 | $0.005 | |
| 8 | Claude Opus 4.6 | Anthropic | ● | $0.015 | $0.075 |
| 9 | GPT-5.4 Mini | OpenAI | 🆕 | $0.0003 | $0.0012 |
| 10 | DeepSeek R1 | DeepSeek | ↓ | $0.0006 | $0.002 |
New This Month
GPT-5.4 (Feb 20) — OpenAI's spring flagship. 500K context window, native image understanding, and a step-change in reasoning that finally closes the gap with Claude on coding tasks. At $0.002/1K input, it's aggressively priced. The full GPT-5.4 family launched simultaneously: base, Pro ($0.01/$0.05), Mini ($0.0003/$0.0012), and a limited-access 5.4 Pro for enterprise.
GPT-5.4 Mini (Feb 20) — The price fighter. At $0.0003 input, it directly undercuts DeepSeek V3 while matching GPT-5.2 quality. This model is going to eat an enormous amount of low-cost inference volume.
Gemini 3.1 Pro (Feb 18) — Google's quiet upgrade. Improved multilingual performance, better structured output, and a 2M token context window that makes it the go-to for massive document processing. Slots in at #7 with room to grow.
Grok 4.20 (Feb 25) — xAI's latest didn't crack the top 10 but deserves a mention. Strong reasoning, real-time web access, and no content restrictions make it popular for specific use cases. Sitting at #12 with growing adoption.
The Story
The crown passes. Claude Sonnet 4.6 officially overtakes Claude 3.5 Sonnet to claim the #1 spot. It's the first time since mid-2025 that the top model has changed. The transition was remarkably smooth — same pricing, better capabilities, and Anthropic's migration tools made the switch painless for most developers.
But the real story is OpenAI's comeback. GPT-5.4 launched with four variants at once, and three of them appear in this month's rankings. OpenAI now holds positions #3, #5, and #9. Their strategy is clear: bracket the market. GPT-5.4 Pro for enterprises that need maximum quality. GPT-5.4 base for the mid-tier. GPT-5.4 Mini for price-sensitive, high-volume workloads. It's aggressive, and it's working.
Gemini 3 Flash continues its relentless climb — up two spots to #4. Google isn't trying to win the quality crown; they're trying to own the floor. At $0.0001/1K tokens, Flash processes are essentially free at scale. Combined with Gemini 3.1 Pro at #7, Google has a complete product lineup for the first time.
Claude 3.5 Sonnet drops to #2 but retains massive usage. Legacy integrations, corporate contracts, and risk-averse teams will keep it in the top 5 for months. But the trend is clear.
Market Share
| Provider | Share | Trend |
|---|---|---|
| Anthropic | ~32% | Stable |
| OpenAI | ~24% | Recovering |
| ~22% | Growing | |
| DeepSeek | ~10% | Stable |
| Alibaba | ~4% | Stable |
| Meta | ~3% | Declining |
| xAI | ~2% | New entry |
| Mistral | ~2% | Declining |
| Others | ~1% | — |
The Price War in Numbers
The cost of intelligence is falling off a cliff. Here's what $1 buys you in March 2026 versus January:
| Model Tier | Jan 2026 (per $1 input) | Mar 2026 (per $1 input) | Change |
|---|---|---|---|
| Flagship | 200K tokens (GPT-4o) | 333K tokens (Claude Sonnet 4.6) | +67% |
| Mid-tier | 571K tokens (GPT-5.2) | 500K tokens (GPT-5.4) | -12%* |
| Budget | 3.3M tokens (DeepSeek V3) | 3.3M tokens (GPT-5.4 Mini) | Equal |
| Speed | 10M tokens (Gemini Flash) | 10M tokens (Gemini 3 Flash) | Equal |
*GPT-5.4 is slightly more expensive than 5.2 but significantly more capable — the value per token actually improved.
Looking Ahead
April should be quieter after March's avalanche. Watch for Claude's response to GPT-5.4 — Anthropic rarely lets a competitive gap persist for more than one cycle. DeepSeek V4 rumors continue to build. And Meta's Llama 4 is overdue — the open-source champion hasn't shipped a competitive model in months.
Monthly LLM rankings by Demand Signals, sourced from OpenRouter usage data. Subscribe to our blog for monthly updates.
Get a Free AI Demand Gen Audit
We'll analyze your current visibility across Google, AI assistants, and local directories — and show you exactly where the gaps are.