The workhorse tier: best cheap models in mid-2026
OpenKey team ·
Below $0.50 per million input tokens, you’re not choosing between “good” and “cheap” anymore — you’re choosing between different flavors of good. Of the 338 models in our catalog, 162 price under that threshold on provider input cost, and that group includes several models that were flagship-tier twelve months ago. This post pulls the ones worth actually building on, with the numbers pulled straight from our live catalog.
The four worth knowing
DeepSeek V3.2 — $0.2288/M input, $0.3432/M provider price ($0.2357/M in, $0.3535/M out via OpenKey). 131,072-token context window. Carries benchmark data in our catalog (Design Arena coding/dataviz/gamedev categories), and at this price point it’s one of the few sub-$0.25 models with a real benchmark trail rather than just a rate card.
MiniMax M2.5 — $0.12/M input, $0.48/M output provider price ($0.1236/M in, $0.4944/M out via OpenKey). 204,800-token context window. Also benchmarked in our catalog, and the cheapest input price of the four here — less than a tenth of a cent per thousand input tokens.
Qwen3 235B A22B Instruct 2507 — $0.09/M input, $0.10/M output provider price ($0.0927/M in, $0.103/M out via OpenKey). 262,144-token context window. This is a 235B-parameter model (22B active, mixture-of-experts) priced like a small dense model — the lowest combined input+output rate of any model in this piece, and it carries benchmark data in our catalog.
Gemini 2.5 Flash Lite — $0.10/M input, $0.40/M output provider price ($0.103/M in, $0.412/M out via OpenKey). 1,048,576-token context window — over a million tokens, by a wide margin the largest context of the four, at a price nearly tied with Qwen3-235B. Our catalog doesn’t carry benchmark data for this one; judge it on context window and Google’s own model card rather than a third-party score here.
The rest of the sub-$0.50 tier, for context
These four aren’t the only options. A few other data points from the same pull, sorted by provider input price, show how crowded this tier has gotten:
- Llama 3.1 8B Instruct (Meta) — $0.02/M in, $0.03/M out, 131,072 context, benchmarked
- gpt-oss-20b (OpenAI) — $0.029/M in, $0.14/M out, 131,072 context, benchmarked
- Qwen3 30B A3B Instruct 2507 — $0.048/M in, $0.193/M out, 131,072 context
- DeepSeek V4 Flash — $0.09/M in, $0.18/M out, 1,048,576 context, benchmarked
- Llama 4 Scout (Meta) — $0.10/M in, $0.30/M out, 10,000,000 context, benchmarked
- Kimi K2.5 (Moonshot AI) — $0.375/M in, $2.025/M out, 262,144 context, benchmarked
Llama 4 Scout is worth calling out on its own: a 10-million-token context window at $0.10/M input is the widest window-to-price ratio in the entire catalog, not just this tier.
Worked math: 10M input + 2M output tokens
Take a workload that’s realistic for a mid-volume RAG pipeline or batch-summarization job — 10 million input tokens and 2 million output tokens in a billing period — and run it through the four headline models at OpenKey’s provider-price-plus-3% rate:
| Model | Provider cost | OpenKey cost | Fee |
|---|---|---|---|
| Qwen3 235B A22B Instruct 2507 | $1.1000 | $1.1330 | $0.0330 |
| Gemini 2.5 Flash Lite | $1.8000 | $1.8540 | $0.0540 |
| MiniMax M2.5 | $2.1600 | $2.2248 | $0.0648 |
| DeepSeek V3.2 | $2.9744 | $3.0636 | $0.0892 |
The spread across four “cheap” models on the exact same workload is 2.7x — from $1.13 to $3.06. “Sub-$0.50/M input” is not a single price point; it’s a range spanning almost 3x depending on your output-to-input ratio and which model you pick within it, since output pricing varies more than input pricing at this tier.
How to pick among them
If your workload is output-heavy (long generations, verbose completions), input price matters less than output price — Qwen3-235B’s $0.103/M output rate wins by a wide margin over DeepSeek V3.2’s $0.3535/M. If you need a huge context window for long documents or codebases, Gemini 2.5 Flash Lite’s 1,048,576-token window (or Llama 4 Scout’s 10M) matters more than shaving another tenth of a cent off input price. If you want catalog benchmark data to justify the choice to a team lead, DeepSeek V3.2, MiniMax M2.5, and Qwen3-235B all carry it; Gemini 2.5 Flash Lite currently doesn’t in our data.
None of these four require picking blind — every model here, and the other 158 under $0.50/M input, has its live price and context window on its own model page, filterable at /models/cheapest. Full price table: /pricing. If your workload can tolerate zero cost instead of low cost, check /models/free first — some of what’s there (Qwen3 Coder 480B A35B, Llama 3.3 70B) would otherwise sit in this exact tier.