benchr
For high-volume work such as RAG ingestion, chat routing, and structured extraction, small rate differences add up. The table includes models in the benchr index priced at $1.00 or less per million input tokens, plus open-weight models intended for self-hosting.
| Model | Provider | Input / 1M | Output / 1M | License |
|---|---|---|---|---|
| Qwen3.6-27B | Alibaba | Self-hosted | Self-hosted | Open-weight |
| Llama 4 Maverick | Meta | Self-hosted | Self-hosted | Open-weight |
| Llama 4 Scout | Meta | Self-hosted | Self-hosted | Open-weight |
| Phi-4 | Microsoft | Self-hosted | Self-hosted | Open-weight |
| GPT-5 Mini | OpenAI | $0.250 | $2.00 | Proprietary |
| DeepSeek V4-Flash | DeepSeek | $0.440 | $1.32 | Open-weight |
| Mistral Large 3 | Mistral | $0.500 | $1.50 | Open-weight |
| Kimi K2.6 | Moonshot AI | $0.950 | $4.00 | Open-weight |
| Claude Haiku 4.5 | Anthropic | $1.00 | $5.00 | Proprietary |
| DeepSeek V4-Pro | DeepSeek | $1.32 | $3.96 | Open-weight |
One caveat on the DeepSeek rows: those are peak-hour rates. Anything sent outside 01:00–04:00 and 06:00–10:00 UTC on a weekday costs half as much, which puts V4-Flash at $0.22/$0.66 for about four fifths of the week.
Key Highlights in the Low-Cost Space
- DeepSeek V4-Flash ($0.44/$1.32 peak, $0.22/$0.66 off-peak): It spent most of 2026 as the cheapest million-token context on the market, then tripled its input rate on August 16 and made the discount conditional on sending traffic outside European and Asian business hours. The coding and reasoning scores did not change; the budget case did.
- GPT-5 Mini ($0.25/$2.00): OpenAI's low-cost entry with Structured Outputs, Batch API support, a 400K context window, and 128K max output. Measure latency in your deployment; OpenAI does not publish a universal 160 tok/s guarantee.
- Claude Haiku 4.5 ($1.00/$5.00): While pricier than DeepSeek and GPT-5 Mini, it features high prompt caching capabilities and Anthropic's signature safety and formatting alignment.
Want to compare these with the flagship models? Check out the full AI Model Rankings or compute exact volume pricing on our Cost Calculator.
When cheap stops being cheap
The cheapest API is not always the lowest-cost system. A model with a lower token price can still lose if it requires more retries, produces longer-than-needed answers, misses formatting constraints, or needs more human review. For production, the useful metric is cost per accepted answer.
For high-volume routing, extraction, and classification, budget models can be excellent because the task is narrow and failure is easy to detect. For legal review, complex code repair, research synthesis, or agent loops with side effects, pay attention to failure cost before optimizing the token bill.
A practical shortlist strategy
For most teams, the cheapest-model search should produce a shortlist, not a single winner. Pick one model below $0.50/1M input, one between $0.50 and $1.50, and one stronger fallback above that. Route easy work to the cheapest tier and escalate only when confidence, formatting, or safety checks fail.
This keeps the system cheap without betting the whole product on the lowest line in the table. It also gives you a migration path: if the budget model improves, increase its traffic; if it fails on an edge case, the fallback path is already defined.
Keep provider concentration in mind too. Saving a few dollars per million tokens is not worth losing redundancy if the workload is business-critical. For important systems, pair the cheapest default with a second provider that can take over when latency, rate limits, or availability degrade.