Cheapest AI model API leaderboard

Models without a per-token API rate

These records may be available for self-hosting, but the shared index has no input/output API pair to rank. License terms and infrastructure cost are separate questions.

How the budget tier works in 2026

The budget tier changed hands this month. DeepSeek moved its V4 models to peak and off-peak rates on August 16, 2026, and V4-Flash input went from a flat $0.14/1M to $0.44 during peak hours (01:00–04:00 and 06:00–10:00 UTC, Monday to Friday) and $0.22 the rest of the week. On the blended index above, that hands first place to MiniMax M3 at $0.75 while DeepSeek V4-Flash sits at $0.88 — though at its off-peak rate the same model blends to $0.44 and is still the cheapest row in this set by a distance. The table shows the peak rate, because that is the rate you pay unless you deliberately schedule around it.

One newer option is not in this index yet: Z.ai published GLM-5.3-Flash on August 26, 2026 at $0.15/1M input and $0.50/1M output, halved to $0.075 and $0.25 through September 9. It is in benchr's verified ledger but has not been added to the tool set that feeds this ranking, so treat it as a lead to check rather than a row you can compare here.

The blended ranking also includes output price, so the order can differ from a list sorted on input alone.

Lower price does not imply that a model is limited to trivial work. DeepSeek reports 79.0% on SWE-bench for V4-Flash, while OpenAI reports 48.0% for GPT-5 Mini. Provider figures and token prices still need to be tested against the same workload before they are treated as a quality-per-dollar result.

Self-hosted vs managed API

An open-weight model can have no license fee and still be expensive to serve. Hardware, utilization, concurrency, context length, electricity or cloud rent, engineering, monitoring, and support all enter the total. Models without a per-token API pair stay outside the hosted ranking.

There is no universal traffic level at which self-hosting becomes cheaper. Get a hardware or cloud quote for the exact checkpoint and serving target, measure sustained throughput, and compare that total with the managed API bill for the same accepted workload.

Methodology

Blended price = (input price + output price) / 2. Only records with both rates are ranked. The equal-weight result is a comparison index, not a bill estimate. All listed prices are sourced from provider documentation; use the Cost Calculator to model your token mix.

When the cheapest model is not the cheapest choice

A low token price only wins when the model completes the task with a similar retry rate and similar human-review cost. If a cheaper model produces two extra failed generations for every successful answer, the apparent savings disappear quickly. Measure cost per accepted result, not cost per generated token.

For production routing, use this leaderboard as the first filter, then run a narrow evaluation on your own prompts. The practical test is simple: compare one budget model, one mid-tier default, and one frontier model on the same 100 to 300 examples, then price only the outputs your team would actually accept.

Update cadence and verification

Prices in this table are reviewed against provider documentation when a major model ships, when a provider announces a pricing change, or when benchr updates the shared model index. Because API prices can change quietly, the table should be treated as a decision aid, not a procurement contract. Before signing a long-term vendor agreement, re-check the provider's live pricing page and any enterprise discounts available to your account.

The table also separates token price from platform fit. A model can be cheapest and still be wrong for a team that needs existing SDK support, SOC review, cloud-region guarantees, or internal approval for a specific vendor. Those constraints belong in your evaluation sheet next to the token cost.