The Cheapest AI APIs That Are Still Actually Good

The cheapest AI API is easy to find. The cheapest one that still does the job is the harder search, because plenty of models are cheap for a reason: they miss instructions, hallucinate structure, and cost you more in retries than they ever saved on tokens.

So we set a floor. Every model below clears Arena 1300 on the LMArena leaderboard, and we rank what remains by blended price. That filter kills the false economy and leaves only options that are both cheap and competent. Here is the current board.

Data from our live pipeline, updated September 7, 2026. Prices sync every 6 hours.

How the blended-price math works

List prices split input and output tokens, and the two rarely cost the same. Output usually runs several times higher. A model with cheap input and expensive output can look great in a headline and hurt you in production, especially for chatty, generation-heavy work.

We collapse both numbers into one figure with a simple weighting:

blended = 0.75 × input + 0.25 × output

That 75/25 split assumes most real workloads read more than they write: long prompts, retrieved context, system instructions, then a comparatively short answer. If your traffic is the opposite (short prompts, long generations), weight it 50/50 or flip it, and the ranking shifts. Keep that in mind before you commit.

Combine blended price with an Arena score and you get value, which is exactly what our value leaderboard tracks. The goal is not the absolute lowest number. It is the lowest number that still clears the bar.

The cheapest capable models right now

All ten below sit under $0.15 blended per 1M tokens and above Arena 1300. Prices come from our live market sync; scores from LMArena.

Arena Scores © LMArena, licensed CC-BY-4.0, as of 2026-09-02.

Reading the board

gpt-oss-120b is the value anchor. At $0.07 blended it is the cheapest thing on the list, and its Arena 1366 is not a token-tier score. That $0.037 input price is the lowest of the ten, which matters if your workloads are input-heavy: long context, RAG, document processing. For read-mostly pipelines, nothing here beats it on cost per useful token.

The Gemma small models win on output economy. Gemma 3n 4B ($0.08 blended, $0.12 output) and Gemma 3 12B ($0.08, $0.15 output) have the two lowest output prices on the board. If your app generates a lot of text per call, that flips the math in their favor even though gpt-oss-120b has cheaper input. Gemma 3n 4B sits closest to the floor at Arena 1305, so it is the one to validate hardest before you ship.

Qwen3 30B A3B is the quiet all-rounder. Same $0.08 blended as the two Gemma small models, but Arena 1384, higher than both. If you want the cheapest tier without dropping near the 1300 line, this is the pick.

The 1400-club shows what one or two cents buys. Step up to $0.11 and DeepSeek V4 Flash 0423 lands at Arena 1432. At $0.14 and $0.15, Gemma 4 26B A4B (1435) and Gemma 4 31B (1442) sit at the top of this whole group on capability. Going from gpt-oss-120b to Gemma 4 31B roughly doubles blended cost, from $0.07 to $0.15, and buys you 76 Arena points. Whether that trade is worth it depends entirely on your workload.

When paying more stops being worth it

Here is the uncomfortable part for cost-driven builders: above a certain point, extra spend buys headroom you may never use.

The spread across this entire list is $0.07 to $0.15 blended. That is an 8-cent range per million tokens, and inside it you get Arena scores from 1305 to 1442. If your task is classification, extraction, routing, summarization, or templated generation, most of these models will clear it, and the difference between 1334 and 1442 will not show up in your output quality. It will only show up on your invoice.

The smart move is to match the model to the task floor, not to the top of the board:

• Simple, high-volume, structured tasks: start at the cheapest tier (gpt-oss-120b, Gemma 3 12B, Qwen3 30B A3B). If eval pass rates hold, stop there.
• Reasoning, nuance, or messy inputs: step to the 1400-plus group (DeepSeek V4 Flash, Gemma 4 26B A4B, Gemma 4 31B). The extra few cents earns its keep here.
• Output-heavy generation: weight output price, not blended. Watch GPT-5 Nano, whose $0.4 output is the highest on this list despite a low $0.05 input; a wrong weighting will flatter it.

The rule: pay more only when a real eval on your own data shows the cheaper model failing. Not before.

Five of these ten are open-source (the Qwen and DeepSeek entries), which adds a second lever. If a hosted price moves against you, you can self-host and reset the economics. That optionality is worth factoring in even if you never exercise it. You can filter the full field on our models directory or watch what is climbing on trending.

A quick workflow

1. Define your task and write 20 to 50 real eval cases.
2. Start with the cheapest model that clears the floor (top of the table).
3. Run the eval. If it passes, you are done.
4. If it fails, move up one tier and re-run.
5. Re-check prices before you lock in; our sync updates every 6 hours, and a mid-week cut can reorder this list.

For coding-specific choices the tradeoffs differ, so check our coding models view and see how full stacks combine a cheap workhorse with a stronger model for the hard 10%.

FAQ

What is the single cheapest AI API that is still good right now?

gpt-oss-120b at $0.07 blended per 1M tokens ($0.037 input, $0.17 output), with an Arena score of 1366. It is the lowest-cost model on our board that clears the 1300 capability floor.

Why use blended price instead of just the input price?

Because output tokens usually cost more than input, and a low input price can hide an expensive output price. Our blend (0.75 × input + 0.25 × output) approximates a read-heavy workload. If yours generates more than it reads, reweight it, and the ranking can change.

Is a higher Arena score always worth the extra cost?

No. Across this list the whole price range is $0.07 to $0.15 blended. For simple, structured, high-volume tasks the cheaper models pass just fine, and the extra points from a 1400-plus model show up only on your bill. Pay more when your own eval shows the cheaper model failing, not before.

Ezra, Scout AI Team