Cost Per Token Explained: GPT vs Claude vs Gemini (2026)

Token pricing is the part of AI API documentation that confuses more people than any other. The numbers are small ($0.15 per million tokens), which makes them feel unimportant until the invoice arrives and the math turns out to be less friendly than expected.

The confusion usually comes from two places. First, most people do not have a clear sense of how many tokens their actual usage generates. “A million tokens” sounds like a lot until you realize that a moderately busy customer support chatbot can burn through several million tokens in a day. Second, the price comparison tables you find online present raw token prices without any context for what those tokens actually produce. $0.075/MTok sounds cheaper than $3.00/MTok until you realize one model requires five calls to complete a task that the other handles in one.

This guide gives you the complete picture: what tokens are, how they are priced, how much GPT, Claude, and Gemini charge at current rates, and how to calculate what your specific application will actually cost.

What is a token?

Language models do not read text the way humans do, word by word. They process text as sequences of tokens, where tokens are the result of splitting text according to a vocabulary that the model was trained with. Common short words are single tokens. Longer or rare words often split into multiple tokens.

The practical conversion rate:

• ✓1,000 tokens is approximately 750 words
• ✓1 million tokens is approximately 750,000 words -- roughly the combined length of the Harry Potter series
• ✓A typical customer support exchange (question + answer): 300-800 tokens
• ✓A short blog post (800 words): approximately 1,067 tokens
• ✓A detailed GPT system prompt: 500-2,000 tokens

Token counting is slightly different across providers because different models use different vocabularies. GPT, Claude, and Gemini each have their own tokenization scheme. In practice, the differences are small (within 10-15%) for standard English text. For code, non-English languages, or heavy use of special characters, token counts can vary more significantly.

Use the Vortenza AI Token Counter to measure the exact token count for your specific prompts across different models before estimating costs.

Why AI companies charge per token

The cost of running an inference (generating an AI response) scales with the number of tokens involved. A 100-token response uses roughly half the compute of a 200-token response on the same model. So per-token pricing is the most accurate way to bill for actual compute consumption.

Why output costs more than input

Output tokens (tokens the model generates) cost significantly more than input tokens (tokens in your prompt). The reason is that generating each output token requires the model to run a full forward pass through the neural network -- it is computationally intensive. Reading input tokens can be done more efficiently through parallel processing.

At GPT-4o: input is $2.50/MTok, output is $10.00/MTok -- a 4x difference. At Claude 3.5 Sonnet: input is $3.00/MTok, output is $15.00/MTok -- a 5x difference. This asymmetry is why output-heavy applications (content generation, long-form answers) cost much more per interaction than input-heavy applications (classification, short Q&A).

Context windows and cost

Every token in your context window costs money on every request. A system prompt of 1,000 tokens gets charged as input on every API call. If you make 100,000 requests per month, that system prompt alone costs 100 million tokens in input. At Gemini Flash pricing ($0.15/MTok), that is $15/month just for the system prompt. At Claude Sonnet pricing ($3.00/MTok), it is $300/month. Prompt caching reduces this to near zero for cached content.

How much does one million tokens cost?

One million tokens is the standard pricing unit across all major AI providers. Here is what each model charges as of June 2026.

What does $1 of tokens buy you?

• ✓At Gemini 2.5 Flash ($0.15 input / $0.60 output): $1 buys approximately 6.7 million input tokens or 1.7 million output tokens.
• ✓At Claude 3.5 Sonnet ($3.00 input / $15.00 output): $1 buys approximately 333,000 input tokens or 67,000 output tokens.
• -At Sonnet pricing, you get about 20x fewer tokens per dollar than at Flash pricing.

GPT token pricing explained

GPT-4o pricing

GPT-4o is OpenAI's primary production model. At $2.50/$10.00 per million tokens, it sits in the mid-premium tier. The 4x output multiplier ($10.00 vs $2.50) means output-heavy applications cost proportionally more.

OpenAI automatically caches input tokens for requests where the same prompt prefix has been seen recently. Cached input tokens cost $1.25/MTok (50% off). This applies automatically -- you do not need to implement caching manually. The practical effect: for applications that repeatedly use the same long system prompt, input costs are often lower than the stated $2.50/MTok.

GPT-4o mini pricing

Mini sits at $0.15 input / $0.60 output -- identical pricing to Gemini 2.5 Flash. For OpenAI ecosystem users who want low cost, mini is the natural choice. Tool and function calling implementation in mini is mature and well-tested.

Batch API pricing

OpenAI's Batch API processes requests asynchronously (results within 24 hours) at 50% off standard pricing. GPT-4o via batch costs $1.25 input / $5.00 output per million tokens. For non-real-time workloads (overnight processing, bulk analysis), batch pricing effectively makes GPT-4o comparable to mid-tier models.

Example cost calculations:

See the full OpenAI pricing breakdown in the OpenAI API Pricing 2026 guide.

Claude token pricing explained

Claude 3.5 Haiku

Haiku is Anthropic's cost-optimized model. At $0.80/$4.00, it is roughly 5x more expensive on input than Gemini Flash but noticeably better at following complex, multi-part instructions. For businesses where chatbot quality directly affects outcomes (sales conversations, nuanced customer support), Haiku's quality edge over Flash often justifies the premium.

Claude 3.5 Sonnet

Sonnet is the expensive option in the mid-tier. $3.00/$15.00 makes it the most expensive model outside the frontier tier (GPT-5, Claude Opus) for typical production use. The premium is justified when output quality is critical enough that errors are costly: production code generation, customer-facing writing that gets used without editing, complex multi-step analysis.

Anthropic's prompt caching

Anthropic's caching is the most aggressive of any major provider. Cached reads on Haiku cost $0.08/MTok (90% off standard input). Cached reads on Sonnet cost $0.30/MTok (also 90% off). Cache writes cost 25% more than standard input (one-time cost to establish the cache). The minimum cacheable block is 1,024 tokens. For applications with large repeated system prompts or document contexts, this brings input costs down dramatically.

Example cost calculations:

See full details in the Claude API Pricing 2026 guide.

Gemini token pricing explained

Gemini 2.5 Flash-Lite

Flash-Lite is the cheapest capable model from any major provider: $0.075/$0.30 per million tokens. Half the cost of Flash. The 1M token context window at this price is unusual -- most budget-tier models have smaller context windows. Flash-Lite handles classification, structured extraction, and simple Q&A well. It produces more errors on complex reasoning and nuanced instruction following.

Gemini 2.5 Flash

Flash at $0.15/$0.60 is the best-value production model in 2026 for most workloads. Same 1M context window as Pro at one-eighth the cost. Suitable for customer support, content generation, RAG applications, and most agentic use cases.

Gemini 2.5 Pro

Pro at $1.25/$5.00 (under 200K context) or $2.50/$10.00 (over 200K context) is appropriate for complex reasoning, advanced code generation, and long-document analysis. The context surcharge above 200K is an important nuance: if your application regularly passes prompts above 200K tokens, input cost doubles.

Context caching

Gemini's caching requires a minimum of 32,768 tokens (32K) to activate. Cached storage costs $1.00/MTok/hour. Cached reads cost approximately 25% of standard input pricing. For large repeated contexts (knowledge bases, long documents), caching can reduce input costs by 75%.

Free tier

Gemini 2.5 Flash offers 15 requests per minute and 1 million tokens per day at no cost. For side projects, internal tools, and early MVPs, the free tier often covers all usage.

Example cost calculations:

See full Gemini pricing in the Gemini API Pricing 2026 guide.

GPT vs Claude vs Gemini token cost comparison

The context window gap is significant

GPT-4o and mini max out at 128K tokens. Claude models support 200K tokens. Gemini models support 1 million tokens. For applications that need to process long documents without chunking, or maintain very long conversation histories, Gemini's context advantage is real and has cost implications beyond the per-token price. A 500K-token document processed in one Gemini Flash call costs $0.075. Processed through multiple GPT-4o calls with chunking and aggregation, the per-token cost and engineering overhead are both higher.

The caching difference matters for production

Anthropic's 90% discount on cached reads is the most aggressive caching discount available. For applications where 50%+ of input tokens are repeated context (common in agentic systems, RAG applications, and chatbots with large knowledge bases), the effective input cost on Haiku drops to roughly $0.08-$0.40/MTok blended, which is competitive with Flash pricing.

Real cost scenario 1: customer support chatbot

Setup: 100,000 conversations per month. Average 8-turn conversation: 400 tokens input + 300 tokens output per turn = 3,200 input + 2,400 output tokens per conversation. Monthly tokens: 320M input + 240M output.

With Anthropic caching on Haiku (500-token system prompt): ~16M tokens at $0.08/MTok + 304M at $0.80/MTok = $244/month vs $256 without caching. The savings are modest here because the system prompt is a small share of total input. For applications with larger cached context, the savings are substantial.

Real cost scenario 2: AI content generation

Setup: 50 blog posts per day (1,500/month). Each post: 1,500 tokens input + 3,500 tokens output. Monthly tokens: 2.25B input + 5.25B output. Output-dominated workload: output tokens are 70% of the token cost.

Content generation is where the output cost asymmetry hits hardest. Output tokens are 3,500 per article vs 1,500 input -- 70% of total tokens. Choosing Gemini Flash instead of Claude Sonnet for content generation at this volume saves $82,012/year. Test your specific brief format on both before committing at scale; for most structured content with detailed prompts, the quality difference is small.

Real cost scenario 3: AI agent workflow

Setup: A 10-step research agent running 10,000 tasks/month. Each task: average 15 LLM calls total (mix of planning, execution, and validation). Average per-call: 2,000 tokens input + 800 tokens output. Monthly tokens: 300M input + 120M output.

Note: this is 15 calls per task, not 1. That is the agent multiplier.

The 15-call multiplier per task is the number that most teams miss. A task that looks like it will cost $0.01 at $0.15/MTok on Flash (with 2,000 input tokens) actually costs $0.15+ when the full 15-call execution is accounted for. For agent workloads, model routing pays off most: use Flash-Lite or Flash for simple planning and validation steps, and a more capable model only for the 2-3 steps that require complex reasoning.

See detailed agent cost analysis in the AI Agent Cost Breakdown 2026.

Why businesses miscalculate token costs

Output inflation

The most common single mistake: estimating output length by what you want, not what the model generates. If you ask a model for a “brief summary” and the model produces 800 tokens when you expected 200, your cost is 4x what you estimated. Measure actual output length on 50 real requests before projecting costs. Most teams find their output estimates are 30-50% low.

Context accumulation

In a multi-turn chatbot, every turn includes all prior turns in the context. A 10-turn conversation does not cost 10x a single turn; it costs considerably more because turn 10 includes the full history of turns 1-9. A conversation with 400 tokens per turn balloons from 400 tokens in turn 1 to 4,000 tokens of input by turn 10. Average input cost is roughly 3x the per-turn estimate.

System prompt overhead

A 2,000-token system prompt adds 2,000 tokens to every single API call. At 100,000 daily calls, that is 200M tokens per day in system prompt overhead alone. At Gemini Flash: $30/day. At Claude Sonnet: $600/day. This overhead is invisible if you only count the user message.

RAG content

Applications using RAG retrieve 3-10 relevant documents and include them in the context with every request. Each retrieved document might be 500-2,000 tokens. At 5 documents x 1,000 tokens each = 5,000 tokens of context per request. This is often 3-5x the user message length and is not in the initial estimate.

Retry overhead

Models that fail or produce incorrect outputs get retried. A 10% retry rate adds 10% to your actual token usage vs your planned usage. Over a year, this compounds.

Tool calls

Agents and function-calling applications add tokens for tool call formatting, tool outputs, and the model's processing of tool results. Each tool call round trip typically adds 500-2,000 tokens to the exchange.

Cost per useful output

This framing matters most for three types of applications.

Classification and extraction

If you are classifying support tickets, extracting structured data, or labeling content at scale, your cost is determined by the number of correct classifications, not the number of API calls. A model at $0.30/MTok input that classifies correctly 95% of the time costs $0.00016 per correct classification (assuming 500 input tokens). A model at $0.075/MTok that classifies correctly 70% of the time costs $0.000107 per API call but $0.000153 per correct classification once you retry failures. The cheaper model is close to the same price per correct result, and worse in quality.

Code generation

Claude 3.5 Sonnet at $3.00/$15.00 generates code that developers accept without modification perhaps 80% of the time on typical tasks. GPT-4o mini at $0.15/$0.60 generates code that requires significant editing perhaps 50% of the time. If a developer costs $100/hour and review takes 10 minutes per accepted generation, the time cost per failed generation is $16.67. At 100 daily generations, the developer review cost on mini is $833/day. On Sonnet, it is $333/day. The API cost difference is small compared to the labor cost difference.

Customer support resolution

A chatbot that resolves support queries on the first response 70% of the time vs 85% of the time has a measurable difference in human escalation cost. Each escalation to a human agent costs $2-$5 in human time. The model quality differential translates directly to escalation rate difference, which translates to monthly human cost.

How to reduce token costs

Optimize prompt length

• ✓Audit system prompts for redundancy; typical first-draft prompts are 30-40% longer than necessary
• ✓Replace multi-sentence instructions with examples (usually fewer tokens, better results)
• ✓Remove repeated instructions that restate the same point differently
• ✓Expected savings: 15-30% on input tokens

Control output length

• ✓Add explicit output length constraints to prompts
• ✓Use structured outputs (JSON, specific format templates) to reduce preamble
• ✓Output tokens cost 4-10x more than input; shorter outputs reduce costs directly
• ✓Expected savings: 20-40% on output costs

Implement prompt caching

• ✓Cache system prompts, knowledge base documents, and other repeated context
• ✓Anthropic: 90% off cached reads; Google: 75% off; OpenAI: 50% automatic
• ✓Expected savings: 30-70% on input costs for applications with large repeated context

Reduce context accumulation

• ✓Summarize conversation history every 5-8 turns instead of passing full history
• ✓Trim RAG retrieved content to relevant sections rather than full documents
• ✓Set explicit context budget limits per conversation
• ✓Expected savings: 20-40% on context-heavy applications

Route by model tier

• ✓Use Flash-Lite or GPT-4o mini for classification, routing, and simple formatting
• ✓Use Flash or GPT-4o mini for standard conversational and content tasks
• ✓Reserve Haiku, Sonnet, or GPT-4o for tasks where quality directly affects outcomes
• ✓Expected savings: 40-70% when 60-80% of queries go to cheap model tiers

Use batch APIs for offline workloads

• ✓OpenAI Batch API: 50% off GPT-4o pricing
• ✓Anthropic batch API: significant discounts for asynchronous workloads
• ✓Applies to: bulk analysis, overnight processing, content generation queues
• ✓Expected savings: 40-50% on applicable workloads

For a broader model-by-model cost comparison, see our LLM Cost Comparison 2026 guide.

Token cost calculator framework

The formula for estimating monthly token costs:

Monthly LLM Cost =
  (Daily Input Tokens x 30 x Input Price / 1,000,000)
  + (Daily Output Tokens x 30 x Output Price / 1,000,000)

Daily Input Tokens =
  (calls/day) x (system prompt tokens
              + avg user message tokens
              + avg context tokens)

Daily Output Tokens =
  (calls/day) x (avg response tokens)

Worked example: SaaS customer support chatbot

The numbers to measure before estimating:

• ☐Actual token count in your system prompt (use Vortenza AI Token Counter)
• ☐Average token count in user messages across 50-100 real examples
• ☐Average token count in model responses across 50-100 real examples
• ☐Average conversation context length at your typical turn depth

Use the Vortenza AI Prompt Cost Estimator to paste real prompt examples and get accurate token counts and cost projections across multiple models simultaneously.

Which model offers the best token value?

The honest summary: Gemini Flash wins on pure per-token value for most workloads. Claude Haiku wins when instruction following quality justifies 5x the price. Sonnet wins when the output goes directly to users or downstream systems without human review. GPT-4o mini wins for OpenAI ecosystem teams who want Flash-level pricing with OpenAI tooling familiarity.

One-minute token cost audit

Use before launching any AI application or when an invoice is higher than expected.

Understanding your token structure

• ☐Do you know your current split between input tokens and output tokens per request?
• ☐Is your output share above 60% of total token cost? (Optimize output length)
• ☐Have you measured actual output lengths on 50+ real requests, or are you estimating?

Identifying the big inputs

• ☐How many tokens is your system prompt?
• ☐How many tokens of conversation history do you pass per request?
• ☐If you use RAG, how many tokens of retrieved content per request?
• ☐Have you measured all of these with Vortenza AI Token Counter?

Caching status

• ☐Is prompt caching implemented for your system prompt?
• ☐Is RAG document context being cached where applicable?
• ☐Are you taking advantage of OpenAI's automatic caching?

Model routing

• ☐Are you using the same model for all request types regardless of complexity?
• ☐What percentage of requests could be handled by a 5-10x cheaper model?

Quick answers

Optimized for ChatGPT, Gemini, Perplexity, Claude, and Google AI Overviews.

Frequently asked questions