Pricing comparison
The cheapest LLM API, with the receipts
Per-token prices are public, so this page just lines them up: OpenRelay's open-model rates, the cheapest tiers from OpenAI, Google, and Anthropic, and the two levers (batch and caching) that move real bills more than rate cards do.
- From $0.05 per 1M tokens
- Batch at 50% off
- Cached input at 10x off
- Deposit $5, get $10
The OpenRelay price ladder.
Every model in the catalog, cheapest first, with the batch rate you get by trading latency for a 50% discount.
Full pricing page| Model | Realtime in / out per 1M | Batch in / out per 1M | Sweet spot |
|---|---|---|---|
| DeepSeek-OCR 2openrelay/deepseek-ocr-2 | $0.039 / $0.039 | $0.019 / $0.019 | Document to markdown, OCR, data extraction |
| GPT-OSS 20Bopenrelay/gpt-oss-20b | $0.05 / $0.20 | $0.025 / $0.10 | High-throughput, low-latency tasks |
| Gemma 4 31B NVFP4 32Kopenrelay/gemma-4-31b-nvfp4-32k | $0.12 / $0.35 | $0.06 / $0.175 | Low-latency, high-volume chat |
| GPT-OSS 120Bopenrelay/gpt-oss-120b | $0.15 / $0.60 | $0.075 / $0.30 | Agents, tool use, complex reasoning |
| DeepSeek V3.1 Terminusopenrelay/deepseek-v3.1-terminus | $0.27 / $1.00 | $0.135 / $0.50 | Agents, reasoning, code |
| Gemma 4 31Bopenrelay/gemma-4-31b | $0.99 / $1.49 | $0.495 / $0.745 | Vision-grounded chat |
| GLM 5.2openrelay/glm-5.2 | $1.82 / $5.72 | $0.91 / $2.86 | Reasoning, coding, bilingual agents |
| BGE-M3openrelay/bge-m3 | $0.013 (input only) | Realtime endpoint | Embeddings, 100+ languages |
Against the cheapest tier everywhere else.
The budget model from each major proprietary provider, next to our two cheapest text models. Same units, same discount structure.
| Provider | Cheapest model | In / 1M | Out / 1M | Batch |
|---|---|---|---|---|
| OpenRelayYou are here | GPT-OSS 20B (open weights) | $0.05 | $0.20 | $0.025 / $0.10 |
| OpenRelay | DeepSeek-OCR 2 (documents) | $0.039 | $0.039 | $0.02 / $0.02 |
| OpenAI | GPT-5 Nano | $0.05 | $0.40 | 50% off |
| OpenAI | GPT-5 Mini | $0.13 | $1.00 | 50% off |
| Gemini 2.5 Flash-Lite | $0.10 | $0.40 | 50% off | |
| Anthropic | Claude Haiku 4.5 | $1.00 | $5.00 | 50% off |
Reference rates as publicly documented, August 2026; providers change prices, so verify against their current rate cards. Batch discounts are 50% across the industry, which makes the base rate the whole comparison.
Rate cards are the smallest lever.
Three decisions move an LLM bill more than switching providers for a cheaper sticker price.
Right-size the model
Most pipeline volume is mechanical work a 20B model handles: classify, score, flag, extract. Screen everything cheap and escalate only the rows the small model marks uncertain with a two-tier routing pipeline, which typically cuts spend several-fold with no quality loss on the easy majority.
Batch everything offline
Anything nobody is waiting on (backlogs, evals, OCR archives, synthetic data, agent backfills) runs through the Batch Inference API at a flat 50% off: JSONL in, results within 24h, failures isolated per row.
Cache the shared prefix
Pipelines repeat the same system prompt thousands of times. Cached input bills at one tenth of the fresh rate on GPT-OSS models ($0.005 per 1M on the 20B), automatically, with no configuration. The best-practices guide covers the prompt structure that keeps the cache hitting.
What real jobs cost at these rates.
Computed from the catalog rates on this page, at batch pricing. The workload pages show the runnable JSONL for each.
Cheap LLM APIs, answered.
What is the cheapest LLM API right now?
For text generation, the cheapest per-token rates cluster around small open models. On OpenRelay, GPT-OSS 20B runs at $0.05 per 1M input and $0.20 per 1M output tokens realtime, and $0.025/$0.10 in batch, which undercuts the cheapest tiers OpenAI, Google, and Anthropic publish. For OCR-shaped work DeepSeek-OCR 2 is cheaper still at $0.039 per 1M each way, and BGE-M3 embeddings run at $0.013 per 1M input tokens.
Is a cheap LLM API good enough for production?
For closed-set classification, sentiment, moderation, short extraction, and routing, small open models track frontier models closely; the gap opens on deep reasoning and gnarly generation. The honest way to decide is a benchmark batch: run a few hundred rows of your real task on the cheap model and a bigger one, compare against a hand-labeled sample, and let the numbers pick. That experiment costs cents.
How do I cut LLM API costs without switching providers every month?
Three levers matter more than provider-shopping: right-size the model per task (screen with a small model, escalate the hard 10%), move offline work to a batch API for the flat 50% discount, and structure prompts so shared prefixes hit cached-input rates (10x cheaper on GPT-OSS models here). Together those routinely cut spend 5-20x more than a rate-card difference does.
Is there a free LLM API?
Free tiers exist but are rate-limited and revocable, which makes them fine for exploration and wrong for production. At real open-model rates, free stops mattering: a million classification calls on GPT-OSS 20B in batch costs about a dollar. Here the $5 starting deposit is doubled to $10 of credit, which covers tens of millions of tokens of evaluation.
What is the cheapest LLM API for high-volume workloads?
High volume is where batch pricing dominates the comparison: any offline pipeline (classification backlogs, OCR archives, evals, synthetic data) runs at half the realtime rate through the Batch Inference API, with 50,000 records per job and results within 24h. GPT-OSS 20B in batch at $0.025/$0.10 per 1M tokens is the floor for general text work on this platform.
Why are open-model APIs cheaper than GPT, Claude, and Gemini?
Open weights remove the model-development premium from the per-token price: providers compete on serving efficiency alone, and you can verify the price is honest because anyone can serve the same weights. You also keep an exit: the day self-hosting beats the API for your volume, the model goes with you.
Run the benchmark, not the debate.
500 rows of your real workload settles which model is cheap enough and good enough. Deposit $5 to get $10; at these rates that is the whole experiment, many times over.