What enterprise LLM APIs really cost in 2026: tokens, caching, batch, and routing

The short answer: an enterprise LLM bill is not “unit price × tokens”. It is driven by input and output tokens, cache hits, batch discounts, reasoning tokens, tool calls, and retries. At 100 million input and 20 million output tokens per month, DeepSeek Flash costs about RMB 360 at peak, while OpenAI GPT-6.1 Sol and Claude Sonnet 5.5 cost about USD 400 at standard list prices. Caching, batch, and model routing can change the total materially — but only after quality and latency pass the gate.

This guide is for technical and procurement teams building AI applications, agents, knowledge bases, or content workflows. Data cut-off: 25 October 2026. Prices come from public DeepSeek, OpenAI, Anthropic, and Alibaba Cloud Model Studio pages; RMB and USD are not converted and no ranking is implied. The model is an editorial estimate, not a quote, and does not promise savings, accuracy, or a go-live date.

How are LLM APIs actually priced?

Most text models price input and output tokens separately; caching, batch, long context, and reasoning tokens change the effective rate. Do not compare a single headline price.

Input tokens include prompts, documents, and context. Output tokens include the answer, reasoning content, and tool arguments. Output is often more expensive because generation is sequential. Long prompts may enter a higher price band; cache hits only reduce repeated input; batch usually discounts non-real-time work but adds latency.

List price is also not the bill. It may exclude tax, tool calls, retries, search/fetch, embeddings, vector storage, and observability. Region, peak/off-peak time, and deployment scope can change the rate. Confirm currency, unit, region, and validity before comparing vendors.

What do 100M input and 20M output tokens cost per month?

At standard list prices, DeepSeek Flash is about RMB 360 at peak; OpenAI GPT-6.1 Sol and Claude Sonnet 5.5 are about USD 400. Model choice matters more than the last decimal.

Assumptions: 100 million input tokens and 20 million output tokens per month; standard text inference only; no cache, batch, tools, retries, tax, or observability; RMB and USD listed separately. At this scale, the workload could be tens of thousands of medium-length Q&A sessions or a batch of document/content tasks, depending on prompt and output length.

Editorial model: 100M × input price + 20M × output price. Excludes cache, batch, tools/search, retries, tax, and observability. RMB and USD are not converted. DeepSeek peak/off-peak follows the official page; OpenAI/Anthropic/Qwen are public standard prices and may change by region/version.

The most common misreading is “cheapest”. A low Haiku or Qwen Flash price does not mean it can handle every task. If a cheaper model causes rework, human review, or a wrong commitment, the saving disappears. The next section looks at cache, batch, and routing.

How much can caching, batch, and model routing save?

Cache only cuts repeated input; batch usually halves the rate; routing sends simple tasks to cheaper models. They can be combined, but not at the expense of correctness.

They work on different levers. Caching suits long, repeated system prompts or fixed knowledge; batch suits non-real-time classification, summarisation, and offline evaluation; routing sends tasks by difficulty, so cheap models handle most simple requests and expensive models handle complex reasoning.

Same 100M input + 20M output model. “50% cache” means half the input hits; cache writes, minimum cache length, and vendor rules differ. Batch and cache may not stack; columns are separate, not additive. Tax, tools, retries, and observability excluded.

Caching is not an automatic switch. It needs a stable request prefix; otherwise hit rate stays low. Batch is for offline work, not interactive support or agents. Cache writes may cost extra, and long context can move the whole request into a higher band. The sustainable optimisation is routing + evaluation + budget caps, not simply buying the cheapest model.

Why is the real bill higher than the token estimate?

Reasoning tokens, tool calls, retries, long context, cache writes, and observability all add usage. A simple input/output estimate misses those lines; calculate per task.

An agent request may include system prompts, retrieved passages, conversation history, tool definitions, tool returns, reasoning, and the final answer. Output tokens are not only the customer-facing paragraph. Retrieval, reranking, embeddings, vector storage, search/fetch, logs, and monitoring may be billed separately. Failed retries and timeout replays can charge the same request multiple times.

• Reasoning/thinking tokens: billed as output or separately by some models; complex tasks use far more than the visible answer.
• Tool calls: arguments, results, and loops consume tokens; search/fetch may also have a per-request fee.
• Long context: once a threshold is crossed, the whole request may be billed at a higher rate.
• Cache writes: creating a cache may cost 125% of standard input; low hit rates can make it uneconomic.
• Retries and concurrency: timeouts, rate limits, and replays add duplicate charges; peak pricing may apply.
• Observability and governance: logs, tracing, evaluation, gateways, and audit are not model fees but are part of the AI application cost.

When can a cheaper model not be used?

Classification, extraction, and summaries can start cheap; complex reasoning, code, high-stakes commitments, and long agent chains need a same-task evaluation before switching.

Editorial guidance, not a model ranking. Use your own task set to compare accuracy, latency, and cost per successful task; model versions, prompts, and tools change results.

What should procurement and acceptance measure?

Measure cost per successful task, P95 latency, accuracy, retry rate, cache hit rate, tool cost, and compliance — not just the price per million tokens.

Replace “token price” with “cost per successful task”: (model + tools + retrieval + retries + human review + observability) ÷ successful tasks. The denominator must be defined by the business: a correct answer, a closed ticket, usable copy, or a merged code change. Otherwise a cheaper model can create more rework at a lower unit price.

There is no universal “cheap” threshold. Compare the same test set and time window before deciding that a model or pricing model actually lowers cost per successful task.

API or private deployment?

Use APIs for low-to-medium volume and fast launches; calculate private TCO when data cannot leave the intranet, permissions are complex, or long-term load is high.

APIs avoid GPU procurement, scale by usage, and update quickly; the trade-offs are data residency/retention, vendor lock-in, rate limits, and price changes. Private deployment improves data boundaries, permissions, and customisation; it adds hardware, model serving, backup, security, and operations. The question is which constraint matters, not which option is universally cheaper.

For a knowledge-base, content, or lightweight agent POC, start with APIs. When volume is stable, data is sensitive, and multiple internal systems need integration, evaluate private deployment. See private AI deployment cost and RAG knowledge-base cost for those layers.

Is Lingyi a fit?

Lingyi fits private gateways, model routing, permissions, and system integration. If you only need a public API for a light app, a cloud provider is simpler.

Lingyi does not resell model API discounts and does not claim any model is “the cheapest”. LingHarness is for enterprises that need a private gateway, model routing, permissions, audit, and multi-system integration. If you only call public APIs for a lightweight app, use the cloud provider directly. For platform selection, see agent base comparison; to scope a budget and routing policy, talk to us.

FAQ

Are LLM APIs billed only by tokens?

Text generation is usually billed by input and output tokens, but tool calls, search/fetch, embeddings, vector storage, retries, long context, and batch/cache rules may be separate. Confirm unit, region, tax, minimum spend, and price validity before signing.

What do 100M input and 20M output tokens cost per month?

Under this model and public standard prices: DeepSeek Flash about RMB 360 at peak, OpenAI GPT-6.1 Sol about USD 400, Claude Sonnet 5.5 about USD 400, and Qwen3.8-Flash about USD 24.4. Excludes cache, batch, tools, retries, and tax; RMB and USD are not directly comparable.

How much can caching and batch save?

In the same model, 50% input cache cuts GPT-6.1 Sol from about USD 400 to USD 305 and Claude Sonnet 5.5 from USD 400 to USD 310; batch cuts Sonnet 5.5 to USD 200. Actual savings depend on hit rate, cache-write cost, latency, and whether the vendor allows stacking.

Can a cheaper model replace a premium one?

It depends on the task. Classification, extraction, and simple summaries can start cheap; complex reasoning, code, long agent chains, and quotes/compensation need the same test set comparing accuracy, latency, and rework. Do not choose on unit price alone.

Is API or private deployment cheaper?

There is no universal answer. APIs are faster and cheaper to start at low-to-medium volume. Private deployment adds hardware, model serving, integration, operations, and depreciation; it can be justified by data boundaries, permissions, or stable high load, but it is not automatically cheaper.

How do we prevent an AI bill from running away?

Set model/task budget caps, record usage by task, monitor retries and cache hits, limit long context and tool loops, and add human confirmation for high-stakes work. Review cost per successful task monthly, not the token headline.

Sources and caveats

This is public-source research and an editorial model, not a vendor quote, interviews, or testing. RMB and USD are not converted; prices, regions, versions, and billing rules can change. The model excludes tax, tools/search, retries, observability, human review, and private infrastructure. Cache and batch are calculated separately; no stacking is assumed. Model capability and price are different questions; this article does not rank “cheapest” or “best”.

1. DeepSeek, Models & Pricing (accessed 25 Oct 2026): deepseek-flash / v4-pro input, cache-hit, output, and peak/off-peak prices.
2. OpenAI, Introducing GPT-6.1 Sol (29 Sep 2026): USD 2/M input, USD 0.10/M cached input, USD 10/M output.
3. Anthropic, Claude model pricing (effective 15 Oct 2026): Sonnet 5.5 USD 2/10, Haiku 5.5 USD 0.10/0.50, cache and batch pricing.
4. Alibaba Cloud Model Studio, model pricing (updated 22 Oct 2026): Qwen3.8-Max/Flash Singapore prices, 50% batch, context-cache rules.

The Chinese version is at 企业大模型 API 成本与路由.