A detailed cost comparison of LLM inference pricing across DigitalOcean, Together AI, Fireworks AI, Modal, Nebius, Baseten, and OpenRouter as of July 30, prices verification. DigitalOcean has the lowest standard input price for gpt-oss-120b at $0.10/million tokens and the lowest cost for Llama 3.3 70B workloads, while Fireworks AI wins for batch and high-cache-hit scenarios due to its 50% batch discount. Together AI and Fireworks charge less for output-heavy workloads. The piece walks through worked examples (chatbot traffic, document classification, RAG embedding costs, dedicated GPU pricing) and argues that the real metric to optimize is cost per successful task rather than raw token price, since model quality, caching, batch eligibility, and deployment mode all shift the calculus.
Table of contents
Which provider is cheapest for what?How the full inference-platform competitor set comparesHow much does LLM inference cost per million tokens?Pricing table: Cost per one million tokens in July 2026What is the cheapest LLM API for batch processing millions of documents?Scenario 1: Classify one million documents overnightScenario 2: Run a chatbot at 10,000 requests per dayScenario 3: Embeddings and RAG over 100,000 documentsCheapest way to run DeepSeek or Llama 70B in productionWhat the per-token price does not tell youPractical strategies for reducing LLM inference costsConclusionFAQReferencesQuestions this post answers
How much does gpt-oss-120b inference cost on DigitalOcean versus Together AI and Fireworks AI?
DigitalOcean charges $0.10 per million input tokens and $0.70 per million output tokens for gpt-oss-120b, the lowest standard input price among compared providers. Together AI and Fireworks AI both charge $0.15 per million input tokens but a lower $0.60 per million output tokens. The cheaper option depends on the input-to-output token ratio, with the break-even point occurring when input volume is twice the output volume. Anyone sizing up inference providers for a production chatbot can track pricing shifts like these on daily.dev.
Does Fireworks AI offer a batch inference discount for gpt-oss-120b?
Yes, Fireworks AI applies a 50% discount on both input and output tokens for eligible serverless models processed through its batch API. In a worked example, classifying 1 million support tickets with gpt-oss-120b dropped from $138 at standard rates to $69 with batch pricing, making it the cheapest option for asynchronous, non-latency-sensitive workloads like classification or synthetic data generation. Teams weighing batch versus real-time inference costs can follow pricing comparisons like this on daily.dev.
How much can prompt caching reduce LLM chatbot inference costs?
A 70% cache-hit rate on Fireworks' published $0.015 cached-input rate can cut a chatbot's estimated monthly inference cost from $108 to $73.98 for a workload using gpt-oss-120b with 360 million input and 90 million output tokens monthly. This applies specifically to stable prompt prefixes like system instructions, safety policies, and tool definitions that repeat across requests. Developers optimizing inference spend can keep tabs on caching and pricing tactics via daily.dev.