InferenceToken economics

How much is 1 million tokens in an LLM? Pricing guide

Discover how much 1 million AI tokens cost in 2026. Compare LLM API pricing for ChatGPT, Claude, and open-weight models to optimise your inference bill.

, AI Researcher8 min read
In this article8
  1. What is 1 million tokens in AI?
  2. How much is 1 million tokens in ChatGPT and Claude?
  3. Input vs. output tokens: the pricing gap
  4. How open-weight models change the token economics
  5. Hidden drivers of the final token bill
  6. Optimising enterprise token costs with Lyceum Serverless Inference
  7. Frequently asked questions
  8. Sources
A letterpress type case holding hundreds of pieces of movable type
A type case at the Gutenberg Museum in Mainz. Text has been priced by the piece before. Photo: dronepicr, CC BY 2.0, via Wikimedia Commons

What is 1 million tokens in AI?

In natural language processing, a token represents the basic unit of text that a machine learning model reads and generates. For the English language, standard subword tokenizers (such as Byte-Pair Encoding) divide written material into character clusters, where one token corresponds to approximately 0.75 words, or roughly four characters of text[1]. Consequently, a metric volume of 1 million tokens equals roughly 750,000 English words.

When sizing infrastructure budgets or evaluating LLM cost per token, abstract token counts translate directly into tangible operational volumes. For enterprise workflows, 1 million tokens represents a substantial amount of information density across varied text formats:

  • Full-length books: Roughly the text of seven to ten average-length published novels.
  • Corporate correspondence: Several thousand standard workplace emails of a few hundred words each.
  • Technical documentation: On the order of a thousand-plus pages of single-spaced analytical reports or operational manuals.
  • Customer service dialogues: A couple of thousand multi-turn support transcripts, depending on context retention depth.

Understanding this baseline volume is critical because model providers do not charge flat software subscriptions for programmatic access. Instead, API usage is metered purely on consumption, billing enterprises for every million tokens ingested as input or delivered as output.

How much is 1 million tokens in ChatGPT and Claude?

API pricing across proprietary frontier models spans a wide cost range depending on model parameter size and reasoning capability. As documented in verified LLM API pricing statistics, commercial providers bill per million tokens (MTok) and charge distinct rates for prompt ingestion (input) versus text generation (output).

On OpenAI's API platform, input rates range from $0.20 per million tokens on lightweight tiers like GPT-5.4 Nano up to $5.00 per million tokens on the flagship GPT-5.6 Sol, while output rates reach $30.00 per million tokens on GPT-5.6 Sol and up to $180.00 on specialised high-compute tiers like GPT-5.5 Pro[2]. Anthropic's Claude API operates across a comparable spread, charging $1.00 per million input tokens ($5.00 output) on Claude Haiku 4.5, $2.00 per million input tokens ($10.00 output) on Claude Sonnet 5, $5.00 per million input tokens ($25.00 output) on Claude Opus 5 and Opus 4.8, and $10.00 per million input tokens ($50.00 output) on Claude Fable 5[3].

Provider & Model FamilyModel TierInput Price (per 1M Tokens)Output Price (per 1M Tokens)
OpenAI GPT-5.6Sol (Flagship)$5.00$30.00
OpenAI GPT-5.6Terra (Balanced)$2.00$12.00
OpenAI GPT-5.6Luna (High-Volume)$1.00$6.00
OpenAI Legacy / EfficientGPT-5.4 Nano$0.20$1.25
Anthropic ClaudeFable 5 (Frontier)$10.00$50.00
Anthropic ClaudeOpus 5 / Opus 4.8$5.00$25.00
Anthropic ClaudeSonnet 5 (Standard)$2.00$10.00
Anthropic ClaudeHaiku 4.5 (Fast)$1.00$5.00

Because the rate gap between entry tiers and flagship models spans a 25-fold to 50-fold difference on input tokens, choosing an oversized model for routine classification or data extraction rapidly multiplies operational spend.

If those per-token rates are too high for your volume, open-weight models through a hosted API cost less: on Lyceum's serverless inference, an OpenAI-compatible API, DeepSeek V4 Flash costs $0.30 per 1M output tokens.

Input vs. output tokens: the pricing gap

A standard feature across all major AI model rate cards is that output tokens cost significantly more than input tokens. Across commercial providers, output generation is typically priced 3 to 6 times higher than prompt input. On the GPT-5.6 family, output costs exactly 6x input, so Sol bills $5.00 per million input tokens against $30.00 per million output tokens[2]. Claude maintains a 5x output-to-input differential across its entire current lineup[4], running from Haiku 4.5 ($1.00 vs. $5.00) through Sonnet 5 ($2.00 vs. $10.00) and Opus 5 ($5.00 vs. $25.00)[3].

This asymmetry stems directly from the hardware mechanics of transformer architectures. Understanding the token cost trajectory requires separating the two core operational phases:

  1. Prompt ingestion (prefill phase): Input tokens are processed in parallel across GPU compute cores. The entire prompt context is ingested simultaneously in a single compute-bound matrix operation, resulting in high GPU utilization efficiency and low compute time per token.
  2. Text generation (decoding phase): Output tokens must be generated autoregressively, one sequential token at a time. Each generated token requires reading the entire model weight matrix and previous key-value (KV) cache from high-bandwidth memory, making decoding memory-bandwidth bound and computationally resource-intensive.

Because sequential decoding consumes GPU memory bandwidth over prolonged durations, serving infrastructure must dedicate accelerator capacity per active output stream. Workloads characterized by verbose responses or iterative reasoning loops therefore incur the highest relative costs.

How open-weight models change the token economics

The emergence of high-capability open-weight models hosted on serverless infrastructure has established a competitive alternative to proprietary APIs. By running open architectures on optimized inference engines, enterprises achieve substantial unit cost reductions while preserving parity on standard enterprise benchmarks.

When evaluating cheapest LLM API token pricing , open-weight models deployed on specialized platforms offer steep price advantages over closed-source equivalents. Qwen3.5 9B, for instance, is priced at $0.15 per million input tokens and $0.20 per million output tokens, a 94% reduction against closed-source alternatives such as Claude Haiku 4.5 at $1.00 input and $5.00 output .

Open-Weight ModelOpen Input / Output (per 1M)Closed-Source Benchmark TierClosed Input / Output (per 1M)Token Cost Differential
Qwen3.5 9B$0.15 / $0.20Claude Haiku 4.5$1.00 / $5.00-94% on tokens
MiniMax M3$0.40 / $2.00Claude Sonnet 4.5$3.00 / $15.00-87% on tokens
DeepSeek V4 Pro$1.75 / $3.50Claude Opus 4.6$5.00 / $25.00-82% on tokens
GLM-5.2$1.50 / $4.50Claude Opus 4.8$5.00 / $25.00-80% on tokens
Kimi K3$3.00 / $15.00Claude Fable 5$10.00 / $50.00-70% on tokens

For engineering teams running high-throughput production workloads, shifting from proprietary APIs to open-weight architectures allows inference volume to scale by a multiple within an identical budgetary envelope, since the per-token rate rather than the request count is what changes.

Hidden drivers of the final token bill

Evaluating an AI infrastructure invoice requires looking beyond nominal per-million token rates. Several secondary mechanisms significantly influence the gross token total and the effective unit price paid across enterprise production environments.

Modern inference providers offer technical controls and volume modifiers that alter baseline token expenditure:

  • Prompt and context caching: When applications repeatedly pass static system instructions, reference schemas, or large document corpora, prompt caching stores the intermediate KV states. Anthropic, for example, bills cache read hits at 0.1x of standard input token rates (10% of the standard input price), while 5-minute cache writes incur a 1.25x multiplier[3].
  • Asynchronous batch processing: For non-real-time jobs such as offline document classification, data extraction, or backfill evaluations, batch processing endpoints apply a 50% discount across both input and output tokens for responses returned within 24 hours[4].
  • Tool use and schema overhead: Enabling function calling or native web search automatically injects system definitions into every API request. Tool definitions add several hundred prompt tokens per call depending on the model and tool specification, accumulating continuous input costs across multi-turn sessions.
  • Long-context surcharges: Certain proprietary providers implement separate billing meters for large context windows. OpenAI doubles input rates and raises output rates by 50% for the full request once a prompt exceeds roughly 272K input tokens, transforming deep document retrieval into a higher cost tier[5].

Careful architecture design, such as structuring prompts to maximize cache hits and routing non-urgent jobs to batch queues, often yields larger net cost reductions than renegotiating baseline rate cards.

Optimising enterprise token costs with Lyceum Serverless Inference

Managing production inference spend at scale requires combining predictable per-token pricing with dynamic workload management. Dedicated serverless platforms eliminate the necessity of provisioning, managing, and paying for idle GPU hardware capacity.

To address these enterprise requirements, Lyceum Serverless Inference runs open-weight models through a drop-in OpenAI-compatible API, with a per-token price published for every model it serves. Billing is pay-per-token throughout, so engineering teams deploy production models without capital commitments or fixed cluster reservations.

Key technical capabilities of Lyceum Serverless Inference for controlling enterprise inference costs include:

  • Open-weight catalogue access: direct programmatic access to the models on Lyceum's published list, among them GLM-5.3, DeepSeek V4 Flash and Qwen3.5 9B, each with its own per-token rate.
  • Smart routing: a single endpoint that selects the model per request from the prompt it receives, instead of hard-coding one model for every task.
  • OpenAI-compatible integration: Standardized API endpoints supporting chat completions, streaming, function calling, and structured JSON outputs with zero changes to existing client codebases.
  • Asynchronous batch execution: Batch inference endpoints delivering standard list-price discounts for non-interactive data processing workloads.

By matching request complexity to specific open-weight tiers and using automated routing, Lyceum Serverless Inference lets organisations bring down total token spend while keeping production workloads on one endpoint.

Frequently asked questions

What does 1 million tokens mean in AI?

Tokens are the foundational text units processed by an LLM. In English, 1 million tokens equals roughly 750,000 words. That is enough volume to cover several thousand workplace emails or a medium-sized corporate knowledge base.

Why are output tokens more expensive than input tokens?

Reading a prompt is highly parallelised. Generating new text requires the model to sequentially predict one word at a time, making output tokens 3 to 6 times more computationally expensive to produce than input tokens.

How much is 1 million tokens in ChatGPT?

Pricing for the OpenAI API depends entirely on the model selected. The flagship GPT-5.6 Sol costs $5.00 per million input tokens and $30.00 per million output tokens. Lighter models like GPT-5.4 Nano cost $0.20 per million input tokens.

How do open-weight models reduce API token costs?

Open-weight models remove the premium markup of proprietary APIs. Qwen3.5 9B delivers strong performance for $0.15 per million input tokens, up to 94% below closed-source alternatives.

How does context caching lower the LLM bill?

When processing long, repeated prompts, API providers often discount the cached input. For example, cached Claude API queries cost just 0.1x of the base input price, dramatically lowering the bill for persistent context windows.

Sources

  1. What are tokens and how to count them? help.openai.com
  2. OpenAI API pricing in 2026: every model after the July price cuts cloudzero.com
  3. Pricing platform.claude.com
  4. Claude pricing in 2026: every plan, API rate, and what it actually costs cloudzero.com
  5. GPT-5.5 272K Context Pricing Cliff: Costs Double tokencost.app

Rankings and scores on Optinest cannot be bought — no commercial relationship changes how a provider is assessed. How we score

Questions this piece answers

What is 1 million tokens in AI?

In natural language processing, a token represents the basic unit of text that a machine learning model reads and generates. For the English language, standard subword tokenizers (such as Byte-Pair Encoding) divide written material into character clusters, where one token corresponds to approximately 0.75 words, or roughly four characters of text[1]. Consequently, a metric volume of 1 million tokens equals roughly 750,000 English words.

How much is 1 million tokens in ChatGPT and Claude?

API pricing across proprietary frontier models spans a wide cost range depending on model parameter size and reasoning capability. As documented in verified LLM API pricing statistics, commercial providers bill per million tokens (MTok) and charge distinct rates for prompt ingestion (input) versus text generation (output).

Related reading

All Inference →