LLM API Pricing Comparison 2026: Every Model, Every Dollar
LLM API pricing changes constantly. New models drop, old ones get cheaper, and the pricing page you bookmarked last month is already outdated. This post is a comprehensive, no-marketing-fluff comparison of what you'll actually pay across the major providers as of April 2026.
We're covering OpenAI, Anthropic, Google, Mistral, and open-source models via hosting providers. For each, we include per-token pricing and realistic cost-per-task estimates so you can compare apples to apples.
The Pricing Table: All Major Models
Prices are per 1 million tokens (1M tokens). All figures are from official provider pricing pages.
| Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|
| OpenAI | |||
| GPT-4o | $2.50 | $10.00 | 128K |
| GPT-4o mini | $0.15 | $0.60 | 128K |
| GPT-4.1 | $2.00 | $8.00 | 1M |
| GPT-4.1 mini | $0.40 | $1.60 | 1M |
| GPT-4.1 nano | $0.10 | $0.40 | 1M |
| o3 | $10.00 | $40.00 | 200K |
| o3 mini | $1.10 | $4.40 | 200K |
| o4-mini | $1.10 | $4.40 | 200K |
| Anthropic | |||
| Claude Opus 4 | $15.00 | $75.00 | 200K |
| Claude Sonnet 4 | $3.00 | $15.00 | 200K |
| Claude Haiku 3.5 | $0.80 | $4.00 | 200K |
| Gemini 2.5 Pro | $1.25 | $10.00 | 1M |
| Gemini 2.5 Flash | $0.15 | $0.60 | 1M |
| Gemini 2.0 Flash | $0.10 | $0.40 | 1M |
| Mistral | |||
| Mistral Large | $2.00 | $6.00 | 128K |
| Mistral Small | $0.10 | $0.30 | 32K |
| Codestral | $0.30 | $0.90 | 256K |
| Open Source (via hosting) | |||
| Llama 4 Maverick (Together) | $0.27 | $0.85 | 1M |
| Llama 4 Scout (Together) | $0.18 | $0.59 | 10M |
| DeepSeek V3 (DeepSeek API) | $0.27 | $1.10 | 128K |
| DeepSeek R1 (DeepSeek API) | $0.55 | $2.19 | 128K |
Key takeaway: Output tokens cost 3-5x more than input tokens across every provider. This isn't a one-provider quirk — it's an industry-wide pattern. If you haven't read our output token cost deep dive, start there.
Real-World Cost Per Task
Raw per-token pricing is meaningless without context. Here's what common tasks actually cost based on typical token counts:
| Task | Input Tokens | Output Tokens | GPT-4o | Claude Sonnet 4 | Gemini 2.5 Flash |
|---|---|---|---|---|---|
| RAG query | 3,000 | 500 | $0.013 | $0.017 | $0.001 |
| Chatbot turn | 2,000 | 800 | $0.013 | $0.018 | $0.001 |
| Document summary (10 pages) | 8,000 | 1,500 | $0.035 | $0.047 | $0.002 |
| Code generation (function) | 1,500 | 2,000 | $0.024 | $0.035 | $0.001 |
| Email classification | 500 | 50 | $0.002 | $0.002 | $0.000 |
| Long doc analysis (100 pages) | 80,000 | 3,000 | $0.230 | $0.285 | $0.014 |
At first glance, Gemini 2.5 Flash looks unbeatable on price. And for many tasks, it is. But price is only one variable — quality, latency, and reliability all factor into total cost of ownership.
The Hidden Costs Nobody Talks About
The per-token price on a pricing page doesn't tell the full story. Here's what actually inflates your bill:
- Retries and fallbacks. If a model hallucinates or times out, you pay for the failed attempt and the retry. A model that's 5% cheaper but fails 10% more often is actually more expensive.
- Context window waste. Stuffing 100K tokens of context into a query when 10K would suffice wastes 10x on input costs. Most teams over-retrieve in their RAG pipelines.
- Verbose output. Models like Claude tend to produce longer, more detailed responses. If you're paying $15/M for output tokens, every extra paragraph adds up. Use
max_tokensaggressively. - Prompt caching miss rates. OpenAI and Anthropic offer prompt caching at 50-90% discount, but only if your prompts have stable prefixes. Dynamic system prompts kill your cache hit rate.
- Rate limit throttling. When you hit rate limits, your queue backs up. The cost isn't the tokens — it's the latency and the engineering time building retry queues.
Cost Optimization Strategies by Provider
OpenAI: Use GPT-4.1 nano for classification and routing, GPT-4.1 mini for most generation tasks, and GPT-4o only when you need the highest quality. Batch API gives 50% off for non-realtime workloads. Prompt caching (automatic in the API) saves 50% on input tokens for repeated prefixes.
Anthropic: Claude Haiku 3.5 is the workhorse — use it for everything that doesn't need Sonnet-level reasoning. Extended thinking mode on Sonnet/Opus increases output tokens significantly, so only enable it when the task genuinely requires chain-of-thought. Prompt caching gives 90% off cached input tokens.
Google: Gemini 2.5 Flash is the best price/performance ratio in the market right now. Use it as your default and only escalate to Pro for complex reasoning. The 1M context window means you don't need to chunk long documents.
Open Source (Llama, DeepSeek): Hosting costs are the wildcard. Together AI, Fireworks, and Groq offer competitive per-token rates, but self-hosting on GPU instances can be cheaper at scale (10M+ tokens/day). Factor in engineering time for infrastructure management.
The Multi-Model Strategy
The smartest teams in 2026 don't use one model — they use three or four:
- Classifier/Router (cheapest): GPT-4.1 nano, Gemini 2.0 Flash, or Mistral Small. Decides which model should handle the request. Cost: $0.10-0.15/M input.
- Workhorse (balanced): GPT-4.1 mini, Claude Haiku 3.5, or Gemini 2.5 Flash. Handles 80% of requests. Cost: $0.15-0.80/M input.
- Premium (quality): GPT-4o, Claude Sonnet 4, or Gemini 2.5 Pro. Reserved for complex tasks that the workhorse can't handle. Cost: $2.00-3.00/M input.
- Reasoning (when needed): o3-mini, o4-mini, or Claude Opus 4. For tasks requiring deep reasoning or chain-of-thought. Cost: $1.10-15.00/M input.
A well-implemented router can reduce your average cost per request by 60-80% compared to routing everything to a premium model.
Scaling Math: What 1M API Calls Actually Costs
Let's say you're building a customer support chatbot that handles 1 million conversations per month. Average 2,000 input tokens and 800 output tokens per conversation.
| Model | Monthly Cost | Annual Cost |
|---|---|---|
| GPT-4o | $13,000 | $156,000 |
| Claude Sonnet 4 | $18,000 | $216,000 |
| GPT-4.1 mini | $2,080 | $24,960 |
| Claude Haiku 3.5 | $4,800 | $57,600 |
| Gemini 2.5 Flash | $780 | $9,360 |
| GPT-4.1 nano | $520 | $6,240 |
| Multi-model (router) | ~$1,500 | ~$18,000 |
The difference between the most expensive option ($216K/year) and the cheapest single-model option ($6.2K/year) is 35x. Even the multi-model router approach saves 90%+ over routing everything to a frontier model.
How to Track This in Practice
Pricing comparisons are useful for planning, but your actual costs depend on your specific usage patterns. The only way to know for sure is to measure.
That's exactly what LLMeter does. Paste your API key, and it pulls your actual usage data from your provider — no proxy, no code changes, no SDK. You see your real costs broken down by model, by day, and by usage pattern.
- Per-model cost breakdown — see exactly which model is eating your budget
- Usage trend charts — catch cost spikes before they become a problem
- Cost alerts — get notified when daily spend exceeds your threshold
- Multi-provider support — compare OpenAI, Anthropic, and Google in one dashboard
Start tracking your LLM costs for free →
Bottom Line
The LLM pricing landscape in 2026 is more competitive than ever. Prices have dropped 50-80% from 2024 levels, but that doesn't mean costs are under control — usage has grown even faster. The teams that win are the ones that match the right model to the right task and measure everything.
Don't guess. Measure.