For a workload with 1 million input tokens and 250,000 output tokens, the representative models on this page range from $0.20 for Gemini 2.5 Flash-Lite to $11.25 for Claude Opus 5 at official list prices. The right provider depends on the tier you compare, not the brand name alone.
How do OpenAI, Anthropic, and Gemini API costs compare?
This table applies the same workload to every model: 1 million fresh input tokens and 250,000 output tokens. It shows first-party list pricing separately from OpenRouter inference pricing. Cached-input discounts are not applied, which keeps the baseline consistent.
| Model | Official in / out | OpenRouter in / out | Example official cost | Context |
|---|---|---|---|---|
| GPT-5.4OpenAI | $2.50 / $15.00 | $2.50 / $15.00 | $6.25 | 1050K |
| GPT-5.4 miniOpenAI | $0.75 / $4.50 | $0.75 / $4.50 | $1.875 | 400K |
| GPT-5.6 LunaOpenAI | $1.00 / $6.00 | $0.50 / $3.00 | $2.50 | 1050K |
| Claude Opus 5Anthropic | $5.00 / $25.00 | $5.00 / $25.00 | $11.25 | See provider |
| Claude Sonnet 5Anthropic | $2.00 / $10.00 | $2.00 / $10.00 | $4.50 | See provider |
| Claude Haiku 4.5Anthropic | $1.00 / $5.00 | $1.00 / $5.00 | $2.25 | See provider |
| Gemini 3.1 Pro PreviewGoogle | $2.00 / $12.00 | $2.00 / $12.00 | $5.00 | See provider |
| Gemini 3.6 FlashGoogle | $1.50 / $7.50 | $1.50 / $7.50 | $3.375 | See provider |
| Gemini 2.5 Flash-LiteGoogle | $0.10 / $0.40 | $0.10 / $0.40 | $0.20 | See provider |
Official sources: OpenAI model documentation, Anthropic pricing, and Gemini API pricing. Some providers use promotional, batch, regional, or long-context tiers that require a separate calculation.
Which provider is cheapest for high-volume workloads?
Budget-tier Gemini models are the lowest-cost representatives in this selected set, while budget OpenAI and Anthropic models occupy different price points. That result should not be generalized to every model or task. Each provider offers several tiers, and the output-to-input ratio can change the order.
High-volume applications should test a small model on real production examples before moving traffic. If the cheaper model causes more retries, longer outputs, or frequent escalation to a premium model, its apparent token-price advantage can shrink.
How should you choose between the three API families?
Broad model ladder
Compare mini and standard tiers when you want to keep one API family while routing simple and difficult tasks to different prices.
Clear Claude tiers
Use Haiku, Sonnet, and premium tiers as a cost ladder, then account for cache use and any model-specific pricing note.
Budget and long context
Flash and Flash-Lite rates can be attractive for volume, but long-context thresholds and modality-specific billing still need review.
When is OpenRouter useful across multiple providers?
OpenRouter can reduce integration work when an application needs several model families or fallbacks behind one API. Direct provider accounts can be preferable when you need first-party features, commercial terms, support, or precise control over a provider relationship.
Compare both routes on total cost, supported parameters, uptime strategy, latency, data retention, and rate limits. A matching token rate does not make the operational experience identical.
OpenAI, Anthropic, and Gemini pricing questions
Which API is cheaper: OpenAI, Anthropic, or Gemini?
For the representative models and example workload on this page, Gemini 2.5 Flash-Lite from Google is the cheapest at $0.20. This is a price comparison, not a claim that the models provide equivalent quality.
Which provider has the cheapest output tokens?
Within the representative set shown here, compare the output half of each input/output price pair. Provider families contain several tiers, so the cheapest model from one provider should not be compared with the premium model from another as if they were direct substitutes.
Does OpenRouter change the model price?
Many OpenRouter inference prices match first-party list rates, while some mapped models differ. OpenRouter also charges separately when credits are purchased. LLM API Prices shows the two channels independently.
How should I compare context windows?
A larger context window can accept longer prompts, but it does not mean the model is cheaper or better for every task. Check long-context price tiers and calculate the tokens you will actually process.
Can I use one API as a fallback for another?
Yes, if your application normalizes prompts, tool definitions, output parsing, and safety behavior across providers. A routing layer can simplify failover, but you should still test model-specific behavior and data policies.