LLM Token Cost by Model: 2026 Pricing Data and Optimization Tips

LLM token costs range from about $0.10 to $10 per million input tokens and up to $50 per million output tokens. Output tokens typically cost three to eight times more than input tokens, with five times the most common ratio.

The price of an LLM token can range from free to tens of dollars per million, and the difference between choosing the right model and the wrong one can multiply your AI bill by 10x without improving outcomes. Output tokens cost three to eight times more than input tokens, cached tokens cost a fraction of fresh ones, and reasoning models bill for thinking you never see.

This guide breaks down current pricing across major providers, explains the mechanics that drive token costs, and covers practical ways to reduce spend without sacrificing quality.

What Is an LLM Token and Why Token Cost Matters

LLM API pricing ranges from free (for self-hosted open-weight models) to $10 per million input tokens for frontier models, with output tokens typically costing three to eight times more. Providers charge per million tokens, where one token equals roughly 0.75 English words. A million tokens translates to about 750,000 words, or around 1,500 pages of text.

A token is a chunk of text the model processes. It might be a word fragment, a full word, or a punctuation mark. Every API call bills based on tokens consumed, making the token the fundamental unit of AI spend.

The pricing spread across models is dramatic. Budget models like GPT-6 Luna or Gemini 3.5 Flash-Lite cost fractions of a cent per thousand tokens. Frontier models like GPT-6 Astra or Claude Fable 5.1 cost $10 per million input tokens and $50 per million output tokens. Choosing the wrong model for a workload can multiply costs by 10x or more without improving outcomes.

The Four Token Types That Set Your Bill

Not all tokens cost the same. Providers bill separately for different token types, and understanding the distinctions matters for accurate cost estimation.

Input Tokens

Input tokens are the text you send to the model. This includes your prompt, system instructions, conversation history, and any context you provide. Input tokens are typically the cheaper token type, ranging from about $0.10 to $10 per million depending on the model.

Output Tokens

Output tokens are the text the model generates back to you. Generation requires more active compute than reading, which is why output tokens cost three to eight times more than input tokens. A verbose response can quickly become expensive, especially at scale.

Cache Read Tokens

Many providers offer significant discounts when you reuse identical input text within a short time window. If your system prompt stays the same across requests, cached input pricing can reduce effective input costs by 50% to 90%. On current flagship models from OpenAI, Anthropic, and Google, cache reads are billed at about 90% off the standard input rate. This is one of the most powerful optimization levers available.

Cache Write Tokens

Writing to cache can cost more than standard input tokens, but the investment pays off when that cached content is read multiple times. Anthropic charges 1.25x the input rate for a 5-minute cache write and 2x for a 1-hour write. OpenAI now lists cache writes separately on newer models too: GPT-6 Astra cache writes cost $12.50 per million versus $10 for standard input. Not all providers expose cache write as a separate line item, though the economics still apply.

LLM Token Cost by Model and Provider

Current pricing varies significantly across providers and model tiers. The table below shows standard API list rates as of October 2026.

* Gemini 3.8 Flash rises to $1.50 input and $7.50 output per million on January 1, 2027.
** Long prompts cost more: Gemini 3.1 Pro is $4/$18 above 200K tokens, and Grok 4.7 is $4/$12 at 200K tokens and above.
*** Peak-hour rate. DeepSeek charges half price off-peak.
Batch processing is 50% off at OpenAI, Anthropic, Google, and Mistral.

OpenAI GPT Token Pricing

OpenAI's model family spans from GPT-6 Luna at budget rates to GPT-6 Astra at frontier pricing, with the GPT-5.6 Sol, Terra, and Luna tiers still available alongside them. Reasoning in these models bills internal "thinking" tokens as output, so they appear on your invoice but not in the visible response. Prompts longer than 272K tokens on GPT-6 Sol bill at 2x input and 1.5x output. Older models including o1, o3-mini, GPT-4 Turbo, and GPT-3.5 Turbo shut down on October 23, 2026, so workloads still pinned to them need to move. Batch API discounts of up to 50% are available for non-time-sensitive workloads.

Anthropic Claude Token Pricing

Claude models range from Haiku for high-volume, cost-sensitive tasks to Opus 5.5 and Fable 5.1 for complex reasoning. Anthropic offers competitive caching discounts, with cache reads at about 90% off input. Two details affect real cost: Claude 4.7 and later models use a tokenizer that produces about 30% more tokens for the same text, and US-only inference adds 10% on Claude 4.6 and later. Claude is also available through AWS Bedrock and Google Vertex AI, but rates and retirement dates are set per channel, so check each platform's pricing page rather than assuming they match Anthropic's direct API.

Google Gemini Token Pricing

Gemini 3.8 Flash and Gemini 3.5 Flash-Lite are positioned as the fast, affordable options for production workloads. Gemini pricing is competitive in the mid-tier segment, and the long context windows available on some Gemini models can reduce the need for retrieval infrastructure. Budget for two changes: Gemini 3.8 Flash's introductory rate doubles on January 1, 2027, and Gemini 2.5 models are now available to existing users only.

Meta Llama and Open-Weight Model Pricing

Open-weight models like Llama are free to download but cost money when accessed via hosted APIs. "Free" depends entirely on whether you self-host or use a provider like Together, Replicate, or a cloud marketplace.

Mistral, Cohere, and Other Provider Pricing

Smaller providers often compete on price for specific use cases. Mistral offers strong price-performance ratios for European deployments: Mistral Large 3 costs $0.50 per million input tokens and $1.50 per million output tokens. Cohere focuses on enterprise search and RAG applications. xAI's Grok 4.7 has one of the lowest output-to-input ratios among mid-tier models, at $2/$6, and DeepSeek halves its rates during off-peak hours. Comparing across providers for your specific workload can surface meaningful savings.

How to Calculate LLM Token Cost for Your Workload

Estimating costs before committing to a model or architecture helps avoid surprises. Here is a practical approach.

1. Count Input and Output Tokens per Request

Start by measuring tokens in a typical request. Most providers offer tokenizer tools to get exact counts. Remember that system prompts count as input tokens on every request, which adds up quickly at scale. Count with the tokenizer for the exact model you plan to use, since token counts for the same text can differ by model generation.

2. Apply the Right Rate to Each Token Type

Separate input, output, and cached tokens when calculating. A common mistake is using a blended rate, which understates costs when output-heavy or overstates them when cache-heavy.

3. Multiply by Request Volume and Cache Hit Rate

Estimate your monthly request volume, then factor in cache hit rate. If 80% of your requests reuse the same system prompt and context, your effective input cost drops significantly.

4. Add Reasoning and Agent Loop Overhead

Reasoning models generate hidden thinking tokens that appear on the bill but not in the response. Agent workflows multiply model calls from a single user action. A straightforward calculation often understates real spend by 2x to 5x for agentic applications.

What Drives the Price of a Token

Understanding why prices vary helps you make better cost-quality tradeoffs.

• Model size and architecture: Larger models with more parameters require more compute per token. Bigger models generally produce better outputs but cost more.
• Context window length: Longer context windows allow more input but cost more to process. Some providers charge premium rates for extended context beyond standard limits. For example, Gemini 3.1 Pro doubles its input rate above 200K tokens, and GPT-6 Sol bills 2x input above 272K.
• Reasoning and hidden thinking tokens: Reasoning modes in models like GPT-6 or Claude's extended thinking bill for internal processing the user never sees. This can multiply costs unexpectedly.
• Cache hit rate: A high cache hit rate dramatically reduces effective input costs. A broken cache can turn a cheap workflow into an expensive one overnight.
• Billing channel and model lifecycle: The same model accessed through different channels, such as direct API versus AWS Bedrock versus Azure OpenAI, can have different rates. Deprecated models may face price increases or feature lockouts. On AWS Bedrock, for example, Claude 3.5 Sonnet now runs $6 per million input tokens and $30 per million output tokens under paid extended access, double its original $3/$15 list price.

How to Reduce LLM Token Cost Without Losing Quality

Optimization requires balancing cost against output quality. The following approaches help maintain quality while reducing spend.

1. Route Work to the Cheapest Model That Passes Evaluation

Use a frontier model only when a cheaper model fails quality checks. Many tasks, including summarization, classification, and simple Q&A, perform well on budget models. Model routing based on evaluation metrics rather than assumptions can cut costs by 50% or more.

2. Cache Aggressively and Measure Cache Hit Rate

Maximize cache utilization by structuring prompts to reuse identical prefixes. Monitor cache hit rate as a key metric. Note that routing decisions interact with caching: changing models can invalidate cached context.

3. Shorten Prompts and Trim Context

Remove redundant instructions. Summarize long documents before including them. Only include the context the model actually needs. Every token in your prompt costs money on every request.

4. Cap Output Tokens and Enforce Structured Responses

Use the max_tokens parameter to prevent verbose responses. Structured output formats like JSON schemas constrain the model to produce only what you need, directly reducing output token spend.

5. Allocate Token Spend to a Team, Feature, or Customer

Optimization requires knowing who consumes what. Without allocation, teams cannot identify which features or customers drive spend. Mapping token costs to business dimensions enables accountability and targeted optimization.

Why Token Cost Understates Your Real AI Spend

Tokens are the most visible and easily metered portion of AI cost, but they are not the whole picture. Treating "AI cost" and "token cost" as interchangeable leads to incomplete budgets and surprise bills.

• Orchestration overhead: Agent frameworks and middleware add latency and compute costs outside the token line
• Retrieval and embeddings: RAG pipelines require embedding generation and vector database queries
• Evaluations and guardrails: Quality checks and safety filters consume additional tokens
• Infrastructure: GPU instances for self-hosted models, storage, and networking
• Labor: The people building, tuning, and maintaining AI systems

A complete view of AI economics includes all of the above, not just the per-token API charges.

Bringing Token Cost Under FinOps Governance

Managing token cost at scale requires the same FinOps practices used for cloud spend: visibility, allocation, budgeting, anomaly detection, and optimization. The FinOps Foundation's 2026 State of FinOps report found 98% of practitioners now manage AI spending, up from 31% in 2024. Finout is an enterprise-grade FinOps platform for cloud and AI spend that treats AI costs the same as infrastructure costs.

Virtual Tags map token spend to teams, features, or customers without changing provider configurations. The Canonical AI Taxonomy normalizes model names across providers, making the same model through Bedrock and direct API appear as one line. Billy lets teams ask natural-language questions about AI spend and get immediate answers. 

AI Agents take that governance further, watching multi-step agentic workflows and surfacing the true cost of agent loops that multiply token consumption behind a single user action. Detector flags the anomaly the moment token spend breaks its normal pattern. Investigator traces it back through the actual chain of calls to find the real cause, not just a summary. Orchestrator matches the finding to the right fix and drives it through your systems, executing reversible actions automatically and requiring a human sign off before anything destructive runs.

For teams that need to answer "what does this AI feature actually cost?" on demand, book a demo to see how Finout brings AI spend under the same governance as cloud infrastructure.