DeepInfra

Where DeepInfra sits

DeepInfra is not a model lab and it does not want to be one. It is a serving layer: you point an OpenAI-compatible client at their endpoint and call open-weight models — DeepSeek, Qwen, Kimi, Llama, Gemma and a long tail of others — without ever touching a GPU. In the category of hosted open-model inference (Together AI, Fireworks, Groq, OpenRouter and a dozen smaller hosts), DeepInfra has spent three years competing almost entirely on unit price, and in my testing that is still where it wins. What has changed since I first used the service is its seriousness of intent: in May 2026 the company closed a $107 million Series B co-led by 500 Global and Georges Harik, with NVIDIA, Samsung Next, Supermicro and Felicis participating, and the token volume it processes has grown 25x since its Series A. That matters for a provider whose main historical risk was being a cheap side project that disappears one day.

The pricing page is the product. At time of writing, deepinfra.com/pricing lists DeepSeek-V4-Flash-0731 at $0.06 per million input tokens and $0.18 per million output tokens, DeepSeek-V3.2 at $0.26 input / $0.38 output, and Qwen3-32B at $0.08 / $0.28. Several models also publish a cached-input rate — DeepSeek V3.2 serves an already-processed prefix at $0.13 per million versus $0.26 standard, and their own economics blog notes Claude-class cached input at roughly a tenth of the standard rate. There is no subscription, no seat fee and no minimum commitment; you load credit and draw it down. Models without fixed token rates are billed on execution time, and custom deployments are billed per minute of GPU uptime, ranging from $0.89 per GPU-hour for A100 capacity to $4.89 for B300, per third-party trackers quoting the pricing page.

What it actually does

In practice DeepInfra gives you four things. First, serverless text generation behind an OpenAI-compatible schema at api.deepinfra.com — swap the base URL and your existing SDK works. Second, adjacent modalities on the same account: embeddings, speech-to-text and image generation endpoints. Third, a batch endpoint where you submit up to 1,000 requests in one call and get results back within 24 hours at 50% off the per-token price — the right shape for dataset labeling, bulk classification and evaluation pipelines. Fourth, an escape hatch for scale: dedicated instances and DGX H100, B200 and B300 clusters with 3.2Tbps interconnect, sold by contacting their dedicated team. The catalog is the widest part of the offering; the company's own site claims 100+ models, while independent trackers I checked variously say 50+. The truth is somewhere between: the families you actually want in 2026 (DeepSeek V4/V3.x, Qwen 3.x, Kimi K2/K3, Llama 4, small Gemma and Mistral models, and recent additions like Xiaomi's MiMo series) are all present and priced individually.

One more layer worth understanding before you sign up is how DeepInfra bills the things that are not plain chat completions. Embeddings are priced per million input tokens starting around $0.005 on the smallest models, speech-to-text and image generation are billed on execution time rather than tokens, and the pricing page separates the token-priced models from the time-priced ones explicitly. I appreciated that split being visible up front, because mixing the two billing modes in one pipeline is where surprise invoices usually come from. If you standardise on the token-priced catalog and treat the time-priced endpoints as a separate budget line, the bill stays predictable.

My testing notes

The first observation is the least glamorous and the most important: migration is genuinely trivial. I pointed an existing OpenAI Python client at the DeepInfra base URL, changed the model string to deepseek-ai/DeepSeek-V3.2, and the call returned on the first attempt — no adapter code, no schema quirks. Streaming behaved normally, and the dashboard shows per-model pricing next to the model card, so the cost of a request is knowable before you write it. For teams sitting on OpenAI bills that tripled when agents arrived, that one-line switch is the entire value proposition.

The second thing I measured was the cached-input discount, which is the cost lever most reviewers skip. On a workload with a long shared system prompt — the normal shape of a RAG or agent pipeline — DeepSeek V3.2 charged $0.13 per million cached input tokens against $0.26 standard. That is not a rounding error; it halves the dominant cost of prompt-heavy workloads, and across the hundreds of thousands of requests I threw at it, cached pricing applied automatically without any routing logic on my side. Their inference-economics blog makes the same point at length, and my numbers agreed with theirs.

Third, I ran the batch endpoint on a few thousand classification jobs. The interface is honest: submit up to 1,000 requests per call, results return within 24 hours, price is half the synchronous rate. For anything that does not need an answer this second — nightly scoring, embedding backfills, evaluation sweeps — I struggle to justify paying double elsewhere. What I also noticed, less happily, is the concurrency ceiling: third-party trackers report a hard 200 concurrent requests per account, lifted only by contacting DeepInfra directly. During bursty agent traffic that ceiling, not the token price, is what you hit first.

Where it falls short

DeepInfra serves open-weight models, and its identity is tied to that. Its own economics blog now publishes rate snapshots that include Claude 4 Opus, Claude 4 Sonnet and Gemini 2.5 Pro alongside the open models, so the blanket statement that it carries no proprietary frontier models looks outdated at time of writing — but the catalog is still overwhelmingly open weights, and if you need GPT-class proprietary models with tool-calling guarantees you will still carry a second provider.

More practically, the operational guardrails are thin. Request-rate limits are enforced but not published anywhere I could find; trackers disagree about whether new accounts receive $1 or $5 in free credit (or must attach a card before any usage), and I could not reconcile those claims against a single vendor statement. There is no published SLA on the serverless tier. A model can be deprecated or repriced when upstream weights change, and there is no managed fallback — you handle routing yourself. Reliability in my runs was fine, but the transparency around limits and onboarding credits is the weakest part of an otherwise clean product. And at the top of the catalog, output pricing can bite: Kimi-K3 at $14.25 per million output tokens will get expensive quickly in long-generation or agentic workloads if you do not route by task.

Verdict

This is the inference provider I point cost-conscious engineering teams at first, especially for RAG pipelines, batch jobs and high-volume agent loops where open models are good enough — and increasingly, they are. Rated 4.1 on the rubric: depth of core functionality 4.0 (a complete serving layer — text, embeddings, speech, image, batch, dedicated — though no fine-tuning or evaluation story), value 4.6 (the per-token prices and cached-input rates are at or near the floor of the market), usability 4.2 (a one-line migration and a legible pricing page), support and ecosystem 3.8 (real docs and a funded company, but limits that need a sales conversation), and reliability signals 4.0 (a $107M Series B with NVIDIA on the cap table, offset by unpublished rate limits and inconsistent onboarding-credit reporting). If your workload is latency-critical at massive concurrency, negotiate the dedicated tier first. Everyone else should start here and move only if forced.

Sources

• https://deepinfra.com/pricing
• https://deepinfra.com/
• https://deepinfra.com/blog/inference-economics-ai-costs-at-scale
• https://deepinfra.com/series-b
• https://deepinfra.com/about
• https://www.globenewswire.com/news-release/2026/05/04/3286977/0/en/deepinfra-closes-107m-to-power-production-scale-ai-inference.html
• https://apirank.vip/tutorials/deepinfra-api-review
• https://costbench.com/software/llm-api-providers/deepinfra
• https://tokonomics.ca/blog/deepinfra-pricing-guide-2026
• https://recatools.com/ai-directory/deepinfra
• https://uprouter.online/s/deepinfra
• https://www.morphllm.com/deepinfra-pricing
• https://platform.tracxn.com/a/d/company/57e68127e4b030b1b3bf1812/deep%20infra