AI Model Pricing Comparison 2026 GPT vs Claude vs Gemini vs DeepSeek

AI model pricing changes fast. New models launch monthly, prices drop, old models get deprecated. This guide cuts through the noise with verified, side-by-side pricing for every major LLM provider as of March 2026 — so you can pick the right model for your budget and use case.

TL;DR — Quick Comparison Table

All prices per 1 million tokens. Sorted by input cost, cheapest first.

OpenAI Pricing Breakdown

OpenAI's current lineup centers on the GPT-5.4 family, which replaced GPT-4o and o-series models. The three tiers cover everything from lightweight classification to frontier-class reasoning.

Key Details

• Context window: All GPT-5.4 models support up to 270K tokens at standard pricing.
• Batch API: 50% discount on both input and output when using the asynchronous Batch API (24-hour turnaround).
• Flex processing: Lower costs for non-production workloads in exchange for slower response times.
• Caching: 90% discount on cached inputs across all models. Automatic — no configuration needed.
• Data residency: Regional processing endpoints add 10% for models released after March 5, 2026.

Anthropic (Claude) Pricing Breakdown

Anthropic's current generation is the Claude 4.6 series (Opus and Sonnet) plus Claude Haiku 4.5 as the fast/cheap option. Claude is known for strong coding, long-context, and instruction following.

Key Details

• Extended thinking: Available on all current models. Thinking tokens are billed at output rates.
• Prompt caching: 90% discount on cached inputs. Cache writes cost 25% more than base input price. Minimum 1,024 tokens to cache.
• Batch API: 50% discount on both input and output tokens. Results within 24 hours.
• 1M context: Opus 4.6 and Sonnet 4.6 support 1M tokens natively. Haiku 4.5 tops out at 200K.
• Available on: Anthropic API, AWS Bedrock, Google Vertex AI.

Legacy Models Still Available

Google (Gemini) Pricing Breakdown

Google runs two current generations: the new Gemini 3 series (preview) and the stable Gemini 2.5 series. Pricing is through Google AI Studio (consumer) or Vertex AI (enterprise). Prices below are Vertex AI standard tier.

Key Details

• Long context surcharge: Inputs over 200K tokens are charged at 2× the base rate for Pro models.
• Batch/Flex pricing: 50% discount on both input and output for asynchronous batch processing.
• Priority tier: 80% premium for guaranteed high-speed processing.
• Free tier: Google AI Studio offers generous free usage for Gemini 2.5 Flash (rate-limited). Great for prototyping.
• Grounding: Web search grounding is $14 per 1,000 queries after 5,000 free/month.

DeepSeek Pricing Breakdown

DeepSeek continues to be the price disruptor of the LLM market. Their V3.2 model unifies both chat and reasoning under a single endpoint, with the cheapest per-token pricing of any frontier-class model.

Key Details

• Same pricing: Both chat and reasoning modes cost exactly the same — $0.28/$0.42.
• Cache discount: 90% off for cache hits, matching OpenAI and Anthropic's caching ratios.
• Reasoning mode: Supports up to 64K output tokens; chat mode defaults to 4K (max 8K).
• FIM completion: Fill-in-the-middle available for code editing (chat mode only).
• Availability: DeepSeek API can have availability issues during peak demand. Consider third-party hosts for production.

Mistral Pricing Breakdown

Mistral has rebranded around their Magistral (reasoning), Mistral Small (general), and Devstral (code) families. They offer competitive pricing, especially at the small model tier.

Key Details

• Mistral Small 3.1: At $0.10/$0.30, this is one of the cheapest capable models available. Open weights (Apache 2.0), so you can self-host too.
• Ministral family: Purpose-built for edge and embedded use cases. Ministral 3B at $0.04/1M tokens is effectively free for most use cases.
• Batch API: Available on all models for async processing at reduced rates.
• Open source: Many Mistral models are open-weight, meaning you can self-host to avoid per-token costs entirely.

Open Source Models (Llama & Others)

Open-source models don't have a single "official" price — it depends on your hosting provider. Here's what major inference providers charge for Meta's Llama 4 and other popular open models:

Self-Hosting vs. Inference APIs

• Inference APIs (Together, Fireworks, Groq, etc.): Pay per token. No infrastructure management. Best for variable or moderate workloads.
• Self-hosting (vLLM, TGI, Ollama): Pay for compute (GPU rental). Cost-effective at high volumes — a single A100 running Llama 3.3 70B can serve thousands of requests/hour for ~$2/hr.
• Break-even point: Self-hosting typically beats API pricing at roughly 50M+ tokens/day for a 70B model.

Cost Optimization Tips

The model you choose is only half the equation. How you use it matters just as much. These strategies can cut your LLM costs by 50–90%:

1. Prompt Caching

Every major provider now offers automatic prompt caching. If your requests share a common system prompt or context prefix, cached tokens are 90% cheaper. This is the single biggest cost lever for most applications.

2. Model Routing

Don't use a $15/1M output model for tasks a $0.30/1M model can handle. Route requests to the cheapest capable model for each task:

• Classification, extraction, simple Q&A → GPT-5.4 nano, Mistral Small, or Gemini Flash
• Code generation, analysis, drafting → Claude Sonnet 4.6, GPT-5.4 mini, or Gemini 2.5 Pro
• Complex reasoning, research, novel problems → Claude Opus 4.6, GPT-5.4, or Gemini 3 Pro

3. Batch Processing

If you don't need real-time responses, use batch APIs. OpenAI, Anthropic, and Google all offer 50% off for async batch jobs:

• OpenAI Batch API — 50% discount, 24-hour window
• Anthropic Message Batches — 50% discount, 24-hour window
• Google Vertex Batch/Flex — 50% discount

4. Prompt Engineering

• Be concise: Every token costs money. Remove filler from prompts.
• Use structured output: JSON mode prevents rambling responses — fewer output tokens.
• Set max_tokens: Cap output length to prevent runaway generation.
• Compress context: Summarize conversation history instead of sending full transcripts.

5. Avoid Over-Thinking

Models with reasoning/thinking (Claude extended thinking, Gemini's thinking, GPT-5.4 reasoning) produce internal reasoning tokens billed at output rates. For simple tasks, disable thinking mode to avoid paying for unnecessary chain-of-thought tokens.

Which Model Should You Use?

There's no single "best" model. It depends on what you're building, how much you're willing to spend, and what trade-offs matter. Here's a decision framework:

Decision Cheat Sheet

Interactive AI Cost Calculator

Comparing prices in a table only gets you so far. Real costs depend on your specific usage patterns: how many requests per day, average prompt size, output length, caching hit rate, and batch vs. real-time split.