Cheapest AI APIs in 2026: Price vs Capability

Summary

The cheapest AI API is the one that completes your work at an acceptable quality and total cost. A low input-token rate is only the starting point: output length, repeated context, retries and model calls per task can change the result.

This guide compares six paid API options and works through three editable budgets: customer support, document summaries and coding agents. Prices were checked against first-party documentation on September 11, 2026. The calculations are illustrative planning examples, not measured performance or provider invoices.

Quick Take

• Gemini 2.5 Flash-Lite has the lowest standard text-token rates in this six-model shortlist.
• Count every model call, including retries, before estimating a monthly bill.
• Gemini 3.8 Flash's quoted introductory prices end on December 31, 2026.
• DeepSeek V4 Pro has different peak and off-peak rates; budget for the hours you actually use.
• Compare cost per accepted result alongside latency and quality.

Start with the work, not one million tokens

Our API cost calculator now accepts monthly tasks, average calls per task, and average input and billed output tokens per call. All example inputs can be changed. If you already have token totals from logs, switch to Monthly tokens.

The arithmetic is deliberately transparent:

Monthly input tokens = tasks × calls per task × average input tokens per call.

Calculate output tokens the same way. Then multiply each total, in millions, by its corresponding rate and add the two costs. A single task can involve several calls, so counting tasks as API requests can substantially understate consumption.

The calculator uses the verified rates available in our model catalog. The catalog includes the six models in this guide, alongside other options. A calculator winner means the lowest estimated token cost among its eligible catalog entries, not the cheapest service across the entire market.

Six prices to put on your shortlist

All prices below are USD per one million text tokens, using uncached input and standard synchronous output. DeepSeek uses its peak rates for the main comparison. Free quotas, signup credits, negotiated contracts, third-party routing and specialized embedding, image or audio APIs are outside this shortlist.

Google lists Flash-Lite's text input at $0.10 and output at $0.40. Gemini 3.8 Flash is $0.75/$3.75 through December 31, then $1.50/$7.50 from January 1, 2027. Output pricing includes thinking tokens. Google pricing.

OpenAI lists Luna at $0.20/$1.20 and Terra at $2/$12. Both apply 2× input and 1.5× output pricing to the full request above 272K input tokens. Luna documentation, Terra documentation.

Mistral lists Small 4 at $0.15 input and $0.60 output per million tokens. This is a hosted API comparison, not an estimate of self-hosting expenses. Mistral API pricing.

DeepSeek lists V4 Pro peak cache-miss input/output at $1.32/$3.96 and off-peak at $0.66/$1.98. Peak hours are Monday–Friday, 01:00–04:00 and 06:00–10:00 UTC; other hours are off-peak. Its current Flash endpoint serves V4.1 Flash, including requests using retired V4 Flash names. Do not carry forward an old Flash price without checking the endpoint. DeepSeek pricing.

Three monthly budgets you can reproduce

These assumptions are examples, not averages for real businesses. They hold calls and tokens constant across models to isolate price. Actual models can produce different output lengths, need different retries and complete different proportions of the work.

Applying the listed rates gives the following monthly token-only estimates, rounded to cents:

These numbers demonstrate token arithmetic only. A low figure in the coding column does not establish that a model can perform the coding task, use the required tools or meet the same success rate.

Customer support: repeated history matters

Open the support example. Three calls per conversation means 3,000 calls, not 1,000. The input allowance must include instructions, relevant policy text and the conversation history sent on each call.

At Luna's listed rate, input costs $0.90 and output costs $1.08, totaling $1.98. Doubling the number of calls while keeping their sizes constant doubles that estimate. A retrieval service, human escalation or a paid tool invocation adds costs outside this token calculation.

Document summaries: count the whole pipeline

Open the document example. Twelve thousand input tokens per document produces 12 million input tokens across 1,000 documents, even if the summaries are short.

The one-call assumption is appropriate only if the document fits and one pass is enough. If your pipeline first extracts text, summarizes sections and then combines them, count each model call and the repeated text. OCR, storage and retrieval expenses need separate lines in the budget. Check the largest document as well as the average.

Coding agents: retries belong in the estimate

Open the coding example. Two hundred tasks at 12 calls each gives 2,400 calls. Those calls can repeatedly contain repository excerpts, instructions and previous tool output.

Use billed output from provider usage records, including billed reasoning where applicable. Terminal, sandbox and other tool costs remain separate. For a relevant editorial comparison, see our curated model comparisons; a rate card alone cannot decide which model writes acceptable code.

Turn token price into cost per accepted result

Choose representative tasks and define acceptance before comparing providers. For support, that could mean a correct answer grounded in the supplied policy. For summaries, it could require preserving specified facts. For coding, use relevant tests plus review of the requested behavior.

Record the total cost of all attempts, including rejected outputs and retries, then divide by accepted results. For example, a hypothetical $10 run with 80 accepted results costs $0.125 per accepted result. A $15 run with 140 accepted results costs about $0.107. These invented examples illustrate the formula; they are not results from any named model.

Also record completion time and human review effort. A token bill can be lower while the overall workflow becomes more expensive. Keep prompts, tool access and evaluation criteria comparable, and report uncertainty instead of declaring a universal winner from a small sample.

Conditions that can change the bill

Cache behavior: cache-hit rates are not the price for every input token. Measure which tokens actually qualify, plus any cache-write or storage charges.

Batch or scheduled work: discounts can change the economics when latency is flexible. Keep discounted traffic separate from the standard-rate baseline. For the V4 Pro examples above, entirely off-peak traffic would halve the token cost; mixed traffic needs a weighted calculation.

Long context: the calculator excludes a workload estimate when its average call reaches a recorded higher pricing tier or exceeds a recorded context/output limit. Monthly token totals cannot reveal per-request limits. Missing catalog limits also do not prove that an unusually large request is supported.

Price changes: a verified rate is a dated observation. The calculator excludes records after a stored pricing period ends until they are reverified. Recheck provider documentation before committing a budget, especially for introductory rates or region-specific deployments.

AI World Scope take

Start with a small, measurable workload and a clear acceptance rule. Use the lowest-priced candidates that plausibly meet your requirements, then compare their actual cost per accepted result. Move to a more expensive option when the improvement in reliability, capability or review effort justifies it.

This update replaces the August price snapshot with checked September rates and reproducible workload examples. We have not run an original six-model performance benchmark, so we do not assign a universal price-performance winner. The useful next step is to enter your own volumes and validate the assumptions against a real sample of your work.