LLM API Pricing in 2026: GPT, Claude and Gemini per Million Tokens
As of mid-2026, flagship LLM APIs cost $2–5 per million input tokens (GPT-5.5 $5/$30, Claude Opus 4.8 $5/$25, Gemini 3.1 Pro $2/$12), while budget tiers run as low as $0.10/$0.40. Prompt caching discounts of 75–90% and 50% batch pricing change the real math.
What do the major LLM APIs cost right now?
Prices below are USD per million tokens (input / output), standard pay-as-you-go, as compiled from vendor documentation in May–July 2026. Always verify against the vendor's live pricing page before committing spend.
| Tier | Model | Input / Output per 1M tokens |
|---|---|---|
| Flagship | OpenAI GPT-5.5 | $5.00 / $30.00 |
| Flagship | Anthropic Claude Opus 4.8 | $5.00 / $25.00 |
| Flagship | Google Gemini 3.1 Pro | $2.00 / $12.00* |
| Mid | Claude Sonnet 4.6 | $3.00 / $15.00 |
| Mid | GPT-5.4 | $2.50 / $15.00 |
| Mid | Gemini 3.5 Flash | $1.50 / $9.00 |
| Budget | GPT-4.1 Nano | $0.10 / $0.40 |
| Budget | Gemini 2.5 Flash-Lite | $0.10 / $0.40 |
| Budget | DeepSeek V4-Flash | $0.14 / $0.28 |
*Gemini rates roughly double for prompts above 200K tokens.
The discounts matter more than the list price
Three levers reduce real-world costs dramatically. Prompt caching cuts repeated-context input costs by 75–90% across all three major vendors. Batch APIs price asynchronous workloads at roughly 50% of standard rates. And model routing — using budget models for high-volume steps and mid-tier models only where quality demands — routinely cuts blended costs by an order of magnitude.
What does a real workload cost?
A content-generation pipeline consuming about 60K input and 15K output tokens per article costs roughly $0.012 per article on budget-tier models, or about $0.41 on mid-tier models. At these prices, LLM inference is rarely the binding cost in content operations — human review time is.
Practical guidance
Match the model to the step, not the brand to the company. High-volume extraction and classification belong on the $0.10 tier; synthesis and long-form drafting are comfortable on the $1.50–3.00 tier; reserve $5.00-tier flagships for tasks where reasoning depth measurably changes outcomes.
Frequently asked questions
- Which major LLM API is cheapest in 2026?
- At the budget tier, GPT-4.1 Nano and Gemini 2.5 Flash-Lite both list at $0.10 per million input tokens and $0.40 per million output tokens; DeepSeek V4-Flash is comparable at $0.14/$0.28. For most high-volume tasks these are functionally interchangeable.
- How much does prompt caching save?
- All three major vendors discount cached or repeated input context by roughly 75–90%. For pipelines that reuse large system prompts or reference documents, caching often halves total spend on its own.
- What does one AI-generated article cost in tokens?
- A multi-step pipeline using about 60K input and 15K output tokens per article costs roughly $0.01 on budget-tier models and about $0.41 on mid-tier models like Claude Sonnet — small compared to the editorial time the draft still requires.