AI Model Pricing Comparison 2026

What You'll Learn

• How API billing separates input, output, cached and batch tokens
• What current official pricing pages show for OpenAI, Anthropic, Google, xAI and DeepSeek
• How to compare a monthly subscription with metered API use
• How routing, caching and batch jobs can change a monthly AI bill

AI model pricing comparison 2026: what changed

AI pricing is not one number. See our AI model launch comparison for the product context behind fast-moving model names. Providers may publish separate rates for input tokens, output tokens, cached input, batch jobs, priority capacity and long-running context. A model can look cheap on input while costing much more on output. A team that sends long repeated instructions may benefit from caching, while an overnight workload may qualify for batch pricing. Comparing only the headline input rate can produce the wrong decision.

This update uses official provider pages fetched for this article on August 22, 2026. OpenAI's public page redirects to its business pricing area and describes credit-based and token-based enterprise arrangements. Anthropic, Google, xAI and DeepSeek publish model or API details on their own documentation pages. Rates below are in US dollars per 1 million tokens unless stated otherwise. Taxes, exchange rates, committed-use terms, account limits and regional contracts can change the final invoice.

How API token billing works

Input tokens are the text, images or other content sent to a model. Output tokens are the response generated by the model. Many reasoning systems count hidden thinking tokens in output billing, so a short visible answer can still use more billable tokens. Cached input is a separate rate when a provider recognizes repeated prompt content. Batch or flex processing may reduce the rate in exchange for slower or less immediate execution.

For a simple estimate, use this formula: monthly cost equals input tokens divided by 1,000,000 multiplied by the input rate, plus output tokens divided by 1,000,000 multiplied by the output rate. Add cache storage, tool calls, search grounding, image or audio charges and taxes where applicable. This estimate is useful for comparison, but the provider invoice remains the final source.

Official pricing snapshot by provider

The table deliberately separates verified facts from missing facts. The previous version of this article presented several future model names and exact prices without a primary record in the source set. Those figures are removed here rather than repeated as if confirmed.

Google Gemini pricing: a worked example

The fetched Google page shows Gemini 3.5 Flash-Lite Standard at $0.30 per 1 million input tokens and $2.50 per 1 million output tokens. It also lists a Batch tier with its own rates and separate handling for context caching. Google lists paid tiers such as Standard, Flex and Priority for some Gemini 3 models. These are not interchangeable. A priority request can cost more because it buys faster capacity, while flex processing may trade speed for a lower rate.

Suppose an application sends 20 million input tokens and generates 4 million output tokens in a month using the Standard Flash-Lite rates. Input would cost 20 multiplied by $0.30, or $6. Output would cost 4 multiplied by $2.50, or $10. The token subtotal would therefore be $16 before any storage, grounding, tax or other service charge. This is a calculation from the retrieved rates, not a quote for a particular account.

xAI Grok pricing and model identity

The xAI fetched model page lists Grok 4.6 at $2 per 1 million input tokens and $6 per 1 million output tokens. The same page says a dated model name refers directly to a specific release and is not automatically updated. That distinction matters in production. A moving alias may point to a later model with different behavior, while a dated ID is intended for repeatable workflows when the provider supports it.

The xAI page also states that Grok has no knowledge of current events beyond its training data. Applications that require live information should not treat a model response as a current database. Add retrieval, search or a verified data source, and budget those extra calls separately from the model token charge.

DeepSeek model aliases and price verification

DeepSeek's official documentation lists the model aliases deepseek-v4-flash, deepseek-v4-pro and deepseek-v4-flash-vision-exp. It says the Flash and Pro aliases have been updated to dated releases while the calling method remains unchanged. This is useful for developers who need a stable API name, but it is not enough to infer a price. The fetched page did not expose a complete price table, so this article does not assign DeepSeek a dollar rate.

That restraint is important in a price comparison. A secondary table can be useful for discovery, but it should not turn an unverified number into a fact. Check the provider billing page immediately before deployment and record the date in the application documentation.

Anthropic and OpenAI: separate plans from API usage

Anthropic's pricing page presents Claude consumer plans and business or enterprise options separately from metered API use. OpenAI's fetched business page lists a $20 per user per month business price when billed annually and $25 when billed monthly, with enterprise arrangements that can use token-based or credit-based pricing. Neither plan price should be treated as an unlimited API allowance.

A subscription is usually easier for a person who works inside a chat product. An API is usually easier to forecast when software sends requests on behalf of customers. Teams should compare actual usage, rate limits, included features, data controls and user seats rather than comparing $20 with an API headline rate.

Subscription versus API: a practical break-even check

Break-even depends on the product's actual limits. A fixed subscription may be good value for a human user with frequent interactive work, but it does not automatically grant permission to automate requests at scale. Conversely, an API can be cheaper for a small request volume if a subscription includes features the developer does not need. Measure requests, input tokens, output tokens, latency and failure retries for at least one representative period.

Why output tokens can dominate cost

Output tokens often cost more than input tokens because generation uses more compute than reading a prompt. A short prompt that requests a long report can therefore cost more than a large document summarized into a few lines. Set output limits, request the desired format, and avoid asking for repeated explanations that the user does not need.

Reasoning models add another layer. Some providers include reasoning tokens in output billing or expose separate reasoning settings. A cost test should record visible output and total billed output where the provider makes both available. Do not compare a visible word count from one model with an invoice token count from another without checking the billing definition.

Routing, caching and batch controls

Model routing sends each task to a model that meets its quality, latency and privacy requirements. A classifier may use a lower-cost model, while a difficult code review may use a higher-capability model. The saving is not guaranteed. Routing adds engineering work, evaluation cost, fallback calls and monitoring. It should be tested on the site's real prompts rather than assumed from a percentage claim.

Caching can help when the same long instructions or reference material are sent repeatedly. Batch processing can help when results are not needed immediately. Both features have eligibility rules and separate price rows. Keep cache storage, grounding, image, audio and tool charges outside the basic text-token calculation.

Regional, tax and currency effects

Most provider tables are displayed in US dollars, but a final invoice can include tax, currency conversion, payment fees and regional terms. A team in India should model the rupee amount at the expected payment date and add applicable tax treatment after confirming it with its accountant. A global company should also check data residency, cross-border transfer rules and contract terms. A lower token rate may not be the right choice if it conflicts with the application's data requirements.

How to audit an AI model bill

Start with the provider invoice and compare the result with our AI infrastructure cost analysis. Then group charges and group charges by model, input, output, cache, batch, grounding and media. Compare each group with application logs. Look for retries, duplicated prompts, unexpected context growth, tool loops and a change from a dated model ID to a moving alias. Then calculate cost per successful user task, not only cost per million tokens. A cheaper model that causes more retries can produce a higher completed-task cost.

AI model pricing comparison 2026: decision guide

Choose a subscription when a person needs a predictable chat workspace and the product's usage limits fit the work. Choose an API when software needs metered, auditable requests. Choose batch or flex processing when latency is flexible and the provider offers a lower rate. Choose caching when repeated context is large and the cache terms are clear. For DeepSeek and any other provider whose price row was not verified in this research pass, check the official billing page before publishing a number.

The main finding from this August 22, 2026 snapshot is that price comparison is a measurement problem, not a leaderboard. Official pages differ in what they show, model aliases change, and tiers may include different limits. Keep the source URL, retrieval date, model ID, token definition and account tier beside every cost estimate.