AI API Pricing Comparison 2026: OpenAI, Claude, Gemini, and DeepSeek

2026 standard API price comparison

The table uses standard, on-demand text processing wherever possible. It does not mix consumer subscriptions with API billing. Cached-input prices are shown only when the selected provider publishes a directly comparable rate; DeepSeek uses its cache-hit rate, while the regular input column uses its cache-miss rate.

Output tokens are usually more expensive than input tokens. That difference makes response length, reasoning output, and retry behavior important parts of cost control. A model with a low input price can still be expensive for workflows that generate long reports or code files.

The cost formula behind every comparison

For a basic text request, estimated cost equals input tokens multiplied by the input rate, plus output tokens multiplied by the output rate, with both rates divided by one million. If 100,000 input tokens and 25,000 output tokens are processed on a model priced at $1 input and $6 output, the estimate is $0.10 plus $0.15, or $0.25.

Production billing can include more than the text visible in one prompt box. System instructions, message history, retrieved documents, image or audio tokens, tool schemas, internal reasoning tokens, cache writes, cache storage, web search calls, and regional processing can all change the invoice. Treat a simple calculator as a planning baseline, then compare it with provider usage reports after a representative test.

• Count the full request, not only the latest user message.
• Estimate output separately because it commonly has a higher rate.
• Model cache hits and cache writes as different billing events.
• Add a retry and failure allowance for real production traffic.
• Recalculate when prompts cross a provider's long-context threshold.

How the providers structure pricing differently

OpenAI separates standard input, cached input, and output rates, and also offers service modes such as Batch that may reduce asynchronous processing cost. The standard rate is the clearest starting point for interactive applications, while batch pricing is relevant only when a workload can wait and meets the provider's operational rules.

Anthropic publishes base input, cache-write durations, cache-hit, and output columns. A repeated prompt prefix can be economical after caching, but writing a cache is not free. Claude tool use also adds system-prompt tokens, so an agent request can consume more input than the visible conversation suggests.

Google includes thinking tokens in the output charge for Gemini 3.5 Flash and lists separate standard, batch, flex, priority, and cache-storage terms. Those tiers are not interchangeable: compare the mode your production endpoint actually uses. DeepSeek distinguishes cache-hit and cache-miss input prices, which makes prompt repetition and cache behavior especially visible in estimates.

Three workload shapes that change the winner

A short classifier has a large input-to-output ratio: perhaps 2,000 input tokens and 50 output tokens. Input price and latency dominate, and a smaller fast model may be the sensible choice. A support assistant may carry a growing conversation history and generate a few hundred tokens each turn, making history trimming and caching valuable. A research or coding agent may send tool definitions, retrieved files, command output, and long reasoning traces, so output price and tool overhead can dominate.

There is no universal cheapest model because equivalent quality is not guaranteed. A lower-priced model that needs extra retries, larger prompts, or a second validation call may cost more per successful task. Evaluate cost per accepted outcome: a correctly classified ticket, a useful answer, or a merged code change—not merely cost per request.

A safer way to turn token rates into a monthly budget

Start with a measured sample rather than an optimistic guess. Capture the input and output token counts for at least several normal, short, and unusually large tasks. Use percentiles instead of one average so a small group of long requests does not surprise the budget. Multiply the expected successful requests by cost per request, then add retry, evaluation, and development traffic.

Keep a model-specific price record with a source URL and review date. Do not hard-code marketing claims such as 'always under one cent' without stating the workload. Add provider billing alerts and application-side limits. If a product accepts arbitrary documents, constrain file size, extracted text, maximum output tokens, and the number of automatic retries before launch.

• Measure p50, p90, and maximum input and output tokens.
• Budget normal traffic, retries, evaluations, and internal testing separately.
• Set a monthly provider limit and a per-request application limit.
• Review model and price changes on a scheduled cadence.
• Compare the estimate with actual invoices before scaling traffic.

Choosing by value instead of price alone

Use a small benchmark drawn from your own workload. Score answer quality, structured-output reliability, latency, safety requirements, regional availability, rate limits, and total cost. Run the same representative inputs against candidate models and calculate cost per accepted result. This makes trade-offs visible and avoids selecting a model from a headline price that applies to a different service tier.

The RunAIToolkit estimator is useful for the first comparison because it keeps sample text in the browser and separates expected input and output. It is still an estimate: provider tokenizers and hidden request components can differ. Validate the final choice with the provider's token-count endpoint and billing dashboard when available.

Frequently asked questions