LLM pricing comparison, 2026

Comparing LLM pricing across providers is harder than comparing almost any other API-metered product, because the unit economics aren't uniform: input and output tokens are priced separately, cached and batch tokens carry their own discounts, and providers ship multiple tiers at once with overlapping capability. A screenshot of a pricing page is stale within weeks. This page is a live table pulled from the same dataset that powers the rest of Meterlark, so the numbers below are whatever is currently verified, not whatever was true when this was published.

Why you can't compare on the input rate alone

Output tokens are priced higher than input on every model in this dataset, usually by a fixed multiple specific to that model. A provider with a cheap input rate and an expensive output multiple can end up costlier than a competitor with a higher input rate but a tighter output multiple, once your actual completion length is factored in. The only honest comparison is a worked example against your real input and output token counts — see how to estimate LLM costs before you build for the method, or run it directly on the calculator.

What actually differs between providers, beyond price

• Context window and max output tokens vary widely even within a provider's own lineup — check the model page, not just the price.
• Caching support and discount depth differ by provider (and sometimes by model within a provider); see prompt caching cost savings explained.
• Batch API availability and discount percentage differ by provider; see batch API discounts, when to use them.
• Capability flags — vision, tool use, extended thinking — are listed per model and matter as much as price for whether a cheaper tier can actually do the job.

The cheapest model is rarely the right default

Every provider in this comparison ships a low-cost tier built for high-volume, low-complexity work — classification, short extraction, simple chat turns — and a premium tier for harder reasoning. Defaulting your whole application to one model, in either direction, usually means either overpaying on easy requests or under-delivering on hard ones. See cheapest LLM API for production workloads for a dataset-ranked view of the low end specifically, and consider routing by task rather than picking one model for everything.

Every number above links through to its model page, which shows the full field set, the official source, and the exact verification date — click through before you build a cost model on it.

Reading the table by workload type, not just by name

It helps to mentally group providers by what their tiers are built for rather than comparing every model against every other model at once. Every provider in this comparison ships a fast, low-cost tier for high-volume simple tasks, a balanced mid-tier for general application traffic, and a premium reasoning tier for hard multi-step problems — the naming differs, but the three-tier shape repeats across the market. Once you know which tier your task actually needs, the comparison narrows from a dozen-plus models down to two or three genuinely comparable options.

How to use this page

Treat it as a starting shortlist, not a final decision. Narrow to two or three candidate models whose capability and context window fit your task, then run your actual workload shape through the calculator or a direct side-by-side comparison page for each pair. Pricing changes often enough across this market that a decision made from memory is a decision made on stale data — see why AI API prices keep changing for how often that actually happens.

If you're evaluating a switch away from a model you're already using in production, don't stop at the price comparison — run the same prompts through both candidates and compare output quality before committing. A cheaper model that needs a longer prompt or a retry to match your current output quality can erase the savings this table suggested, which is exactly why the shortlist step matters more than the ranking itself.