The cheapest AI inference API in 2026 depends on the model you want to run, because the market has split: frontier models cost dollars per million output tokens on their makers' endpoints, while open-weight models are served by cents-per-token across a dozen competing hosts. The table below is the current cheapest verified endpoint per tier, from our weekly cost check across 88 models.
The cheapest inference right now, by tier
Per 1M output tokens, cheapest verified endpoint from our latest check:
| Tier | Model | From | Where |
|---|---|---|---|
| Tiny | Mistral Nemo | 4 cents | DeepInfra |
| Small | Llama 3.1 8B | 5 cents | DeepInfra, Novita |
| Mid | gpt-oss-20b | 14 cents | DeepInfra |
| Large | gpt-oss-120b | 17 cents | DeepInfra |
| Large | Llama 4 Scout | 30 cents | DeepInfra |
| Very large | Llama 3.3 70B | 32 cents | OpenRouter |
| Flagship-class open | DeepSeek-V3 | 89 cents | DeepInfra |
| Reasoning | DeepSeek-R1 | 2.15 dollars | DeepInfra |
All prices are verified against our live rankings, which re-check 88 models across every known provider weekly. Treat that live table as the source of truth, since prices here can be weeks stale by the time you read this. For context, frontier flagships list at 10 to 30 dollars per 1M output tokens. If your workload does not need frontier reasoning, the cheapest inference API is effectively always an open-weight model on a discount host, at 1 to 3 percent of flagship cost. The full comparison lives in LLM API pricing in 2026.
Who actually wins on price
Six names keep appearing at the top of our tables:
- DeepInfra: the most frequent cheapest-endpoint winner across LLMs and Whisper transcription. See DeepInfra pricing in 2026.
- Novita: consistently within cents of the leader across a large open-model catalog. See Novita AI pricing in 2026.
- OpenRouter: an aggregator rather than a host; sometimes routes below every direct provider, and its free routes are unbeatable for development. See OpenRouter free models.
- Groq and Cerebras: rarely the absolute cheapest, but the fastest tokens per second, which matters when latency is the real cost. Groq used to run a $10K startup credit program, but it is no longer published; see what happened to the Groq startup program.
- Together AI and Fireworks: mid-priced but production-grade, with wide model coverage. Together AI's startup credits are worth checking once you have an entity.
Cheapest by cost vs cheapest by benchmark
Cost per token and speed per second are different rankings, and the cheapest-ai-inference-benchmark question is really asking about the second one. In every third-party throughput benchmark we track, Groq and Cerebras post the fastest tokens-per-second by a wide margin, often several times a standard GPU-backed endpoint, but they are rarely the lowest price per token. DeepInfra and Novita win on raw cost but run on more conventional GPU serving, so latency is closer to average. The decision rule: if your bottleneck is a user waiting on a response, benchmark tokens-per-second before you pick a host; if your bottleneck is a monthly bill for offline or async work, optimize for the price table above and ignore speed entirely.
Price is not the whole bill
Three things make nominally cheap inference expensive:
- Rate limits: a cheap host that throttles you into retries is not cheap. Check advertised limits at your tier before committing volume.
- Quantization and quality drift: some discount hosts serve more aggressively quantized weights. For most product tasks it does not matter; for evals it can. Test one day of real traffic before switching.
- Egress and hidden per-request fees: rare among the majors we track, but read the pricing page. Our tables normalize to the posted per-token price.
The practical playbook: develop on free keys, pick your model, then take the cheapest host for that exact model from the rankings. Switching hosts later is usually a one-line base URL change, so the cost of a wrong first pick is nearly zero.
Cut it further
Cheap inference stacks with everything else: batch discounts halve frontier bills, caching discounts repeated prefixes, and startup credits cover months of usage before you pay cash. The 80 percent playbook is in how to cut your OpenAI API bill, and most of it applies to any provider.
Bottom line
Cheapest inference in 2026: an open-weight model on DeepInfra, Novita, or OpenRouter, at 4 to 90 cents per 1M output tokens depending on size, versus 10 to 30 dollars for frontier flagships. Match the model to the task, take the cheapest verified host from the live rankings, and claim the credits in the catalog before spending real money.
Related reading: LLM API pricing comparison, the cheapest way to run LLMs, every free LLM API key.