DeepInfra raises $107M Series B to scale the inference cloud — read the announcement

Updated November 2026 for the DeepSeek-V4-Pro-0813 release.
DeepSeek V4 Pro reached general availability on August 13, 2026 as DeepSeek-V4-Pro-0813. It keeps the preview’s architecture: a Mixture-of-Experts model with 1.6T total and 49B active parameters, a 1M-token context window and up to 384K output tokens. On top of that it adds a built-in DSpark speculative decoding module and a three-level reasoning_effort control (low, high, max). The 0813 model card shows big vendor-reported gains on agentic and coding tasks. DeepSWE, for example, rose from 12.8 to 62.7. Because the weights are MIT-licensed, a lot of providers serve it, and they differ on price, speed and how long you wait for the first useful token.
This guide compares the providers that serve the 0813 build, based on pricing pulled on 2026-11-16. Speed, latency, blended price and error rate come from Artificial Analysis, which measures the model in its Max reasoning mode. Routed prices come from OpenRouter, and rates and features come from each provider’s own model or pricing page. V4 Pro reasons before it answers, so this guide tracks two latency figures. Time to first token (TTFT) measures when the stream starts. Time to first answer token measures when the model has finished thinking and starts writing the answer. For this model, every provider has a TTFT under 3 seconds, but the first answer token arrives anywhere from 32 to 74 seconds after the request, depending on the provider. The second figure is the one your users feel.
| Provider | Best For |
|---|---|
| DeepInfra | Lowest price overall: lowest output price on Artificial Analysis, a Flex tier at $2.08 output, 0% measured error rate |
| DeepSeek (official API) | First-party access, off-peak discounts and the cheapest cache reads |
| DigitalOcean | Fastest output and lowest latency to the first answer token |
| Baseten | Second-fastest output speed on a per-token Model API |
| Nebius Token Factory | Output speed on par with DeepSeek’s API, with dedicated endpoints |
| Novita AI | 0% measured error rate with a low cost per agentic task |
| SiliconFlow | Fast TTFT, plus on-demand and reserved GPU options |
| OpenRouter | One API key, provider fallback and access to 20 providers |
Prices are per 1M tokens. AA’s blended price weights cache hits, input and output 7:2:1. Speed, latency and error rate are Artificial Analysis headline values.
| Provider | Input | Output | Cache hit | Blended (7:2:1) | Output speed | TTFT | First answer token | Error rate |
|---|---|---|---|---|---|---|---|---|
| DeepInfra | $1.30 | $2.60 | $0.10 | $0.59 | 65.1 t/s | 2.17 s | 67.09 s | 0.0% |
| DeepSeek (peak) | $1.32 | $3.96 | $0.044 | $0.69 | 93.5 t/s | 2.57 s | 52.73 s | 0.8% |
| DigitalOcean | $1.32 | $3.96 | $0.044 | $0.69 | 140.0 t/s | 1.72 s | 31.67 s | 3.1% |
| Baseten | $1.32 | $3.96 | $0.132 | $0.75 | 114.2 t/s | 2.84 s | 43.87 s | 0.8% |
| Nebius | $1.32 | $3.96 | $0.04 | $0.69 | 93.9 t/s | 2.92 s | 45.40 s | 0.8% |
| Novita | $1.32 | $3.96 | $0.044 | $0.69 | 77.9 t/s | 2.49 s | 59.31 s | 0.0% |
| SiliconFlow | $1.32 | $3.96 | $0.044 | $0.69 | 41.4 t/s | 1.72 s | 73.93 s | 0.9% |
Pricing as of 2026-11-16. DeepSeek’s first-party prices are halved off-peak (see below).
DeepInfra is the best pick when cost per token is the deciding factor. Its output price is the lowest that Artificial Analysis measures for V4 Pro 0813, and its Flex tier goes lower still for workloads that can tolerate queueing.
deepseek-ai/DeepSeek-V4-Pro-0813, served in fp8"service_tier": "flex". A request waits up to 10 minutes for capacity, then returns HTTP 429, and a rejected Flex request is not charged (how Flex works)"service_tier": "priority")max_tokens is unset, the default is the lesser of the remaining context and 65,536reasoning_effort from none (thinking off) to max, prompt caching with prompt_cache_key, and up to 4 fallback models per requestimport os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["DEEPINFRA_API_KEY"],
base_url="https://api.deepinfra.com/v1/openai",
)
resp = client.chat.completions.create(
model="deepseek-ai/DeepSeek-V4-Pro-0813",
messages=[{"role": "user", "content": "Refactor this function to remove the N+1 query."}],
extra_body={"reasoning_effort": "high", "service_tier": "flex"},
)
print(resp.choices[0].message.content)
DeepInfra fits teams running V4 Pro at volume: agent backends, batch coding jobs, evaluation runs, and long-context document work where output tokens and reasoning tokens (billed as output) drive the bill. Flex is useful for overnight and asynchronous pipelines, and Priority covers the requests that can’t wait. DeepInfra does not have the cheapest cache reads. Several providers charge $0.044 or less, so prompts that are almost entirely cache hits can cost less elsewhere. And if response speed is your product, test DigitalOcean. The original deepseek-ai/DeepSeek-V4-Pro ID is still available at the same price, but new integrations should use the 0813 ID. Full cost scenarios are in the DeepSeek V4 Pro pricing guide.
The DeepSeek API is the best pick when you want first-party access, can schedule work into off-peak hours, or run cache-heavy prompts.
deepseek-v4-pro, which now serves DeepSeek-V4-Pro-0813deepseek-v4-proThe first-party API has the lowest cache-hit price in this guide, and off-peak its output rate ($1.98) is below DeepInfra’s standard $2.60. During peak hours it costs more than DeepInfra on output ($3.96 vs $2.60). It’s a strong fit for teams that can batch work into off-peak windows or that want features only DeepSeek offers, such as the Responses API. The trade-offs are time-of-day pricing that makes costs harder to forecast and a 500-request concurrency cap.
DigitalOcean Serverless Inference is the best pick when speed and latency are the product. It leads Artificial Analysis on output speed and on time to first answer token for V4 Pro 0813.
DigitalOcean is the provider to test for chat UIs, interactive coding assistants, and any agent loop where a user is waiting for the answer. Its first answer token arrives up to 42 seconds sooner than the slowest provider in this guide. It costs more per output token than DeepInfra ($3.96 vs $2.60), and its 3.1% error rate is the highest Artificial Analysis measured for this model, so build in retries.
Baseten’s Model API is the best pick when you want high output speed on per-token pricing, with the option to move to dedicated deployments later.
Baseten suits teams building coding agents who want fast generation and a single vendor for a Pro main agent plus Flash sub-agents. Its cache-hit price is 3× DeepSeek’s and above DeepInfra’s $0.10, so long cached contexts cost more here.
Nebius Token Factory is the best pick when you want output speed on par with DeepSeek’s own API, with dedicated endpoints as an option.
deepseek-ai/DeepSeek-V4-Pro ID and recommends deepseek-ai/DeepSeek-V4-Pro-0813 as its replacement. Requests aren’t rerouted automatically, and dedicated endpoints are unaffectedNebius’s speed and first-answer latency are close to DeepSeek’s own API. The catch is cost on agentic workloads. Artificial Analysis puts its cost per Index task at $13.43 with a 54.9% cache hit rate, far above every other provider here. If your workload depends on prompt caching, test your cache hit rate before committing.
Novita AI is the best pick when reliability and cost per agentic task matter more than raw speed.
deepseek/deepseek-v4-pro-0813, serverless, OpenAI-compatibleNovita and DeepInfra are the two providers with a 0% error rate in Artificial Analysis’s runs. Novita’s cheap cache reads give it a low cost per agentic task. Its output price is higher than DeepInfra’s at either rate ($2.97 via OpenRouter, $3.96 list vs $2.60).
SiliconFlow is the best pick when you want a fast stream start and the option to move from serverless to dedicated GPUs.
deepseek-ai/DeepSeek-V4-Pro-0813SiliconFlow’s low TTFT shows how misleading that metric is for a reasoning model. The stream starts as fast as DigitalOcean’s, but the answer arrives last of the measured providers. It makes sense if you are moving toward reserved capacity, or already use SiliconFlow for other DeepSeek models. It is not the pick for interactive use on the serverless tier.
OpenRouter is the best pick when you want one API key across providers, with automatic fallback when an endpoint degrades.
OpenRouter is a good fit for prototyping and for production apps that need uptime insurance across several backends. It lists a few routes with lower output prices than DeepInfra’s standard rate, such as Baidu at $1.976, plus DeepSeek’s off-peak price. Some of those routes show degraded status or an unstated quantization, and routing means you don’t always control which backend serves a given request. For the DeepInfra rates without a router in between, call DeepInfra directly.
This post originally covered both V4 variants. DeepSeek’s smaller models have moved on since May: the original V4 Flash was followed by a GA build (V4 Flash 0731) and then by V4.1 Flash. On Artificial Analysis’s Intelligence Index, V4.1 Flash 1017 scores 33, against 23 for V4 Pro 0813 (Max). On DeepSeek’s own API, the legacy deepseek-v4-flash name is now served by V4.1 Flash 1017. If your workload doesn’t need V4 Pro specifically, a Flash model is much cheaper. Current DeepInfra prices per 1M tokens:
| Model on DeepInfra | Input | Output | Cached input |
|---|---|---|---|
| DeepSeek-V4.1-Flash-1017 (51% off list) | $0.098 | $0.294 | $0.0029 |
| DeepSeek-V4.1-Flash (30% off list) | $0.14 | $0.42 | $0.0042 |
| DeepSeek-V4-Flash-0731 | $0.06 | $0.18 | $0.015 |
| DeepSeek-V4-Flash | $0.09 | $0.18 | $0.018 |
Both V4.1 Flash models list at $0.20 / $0.60 / $0.006 before the promotion. All four support Flex and Priority and have a 1M-token context.
The right DeepSeek V4 Pro provider depends on which constraint binds first:
For most production workloads, DeepInfra is the best overall choice for DeepSeek V4 Pro 0813. It has the lowest output price that Artificial Analysis measures, the cheapest Flex tier for anything that can wait, and Priority for requests that can’t. It also gives you an OpenAI-compatible API with the full 1M context and 384K output, and it posted zero errors in Artificial Analysis’s runs. Output tokens, including reasoning tokens, usually make up most of a V4 Pro bill, so that output price matters more than any other rate on the page.
Start with the model page or go straight to the API reference. For lighter workloads, DeepSeek V4.1 Flash costs a fraction as much. The DeepSeek V4 Pro pricing guide covers workload cost scenarios, and the API benchmarks guide has the full speed and latency comparison. Early-stage startups can apply for free credits through DeepStart.
Qwen3.5 9B API Benchmarks: Latency, Throughput & Cost<p>About Qwen3.5 9B Qwen3.5 9B is the flagship of Alibaba’s Qwen3.5 Small Model Series, released on March 2, 2026. It is a dense multimodal model combining Gated Delta Networks (a form of linear attention) with a sparse Mixture-of-Experts system, enabling higher throughput and lower latency during inference compared to traditional dense architectures. The architecture utilizes […]</p>
Hosted Agents: your own always-on AI agent, from $13/monthOne click gives you a dedicated, isolated AI agent, pre-wired to fast inference and ready to work the moment it boots. No VMs, no SSH hardening, no patching. From $13/month, and idle is free.
Kimi K3.1 Pro Pricing: API Costs, Providers & Flex (2026)<p>Kimi K3.1 Pro Executive Summary Pricing as of 2026-11-17. DeepInfra’s Standard rate is promotional (5% off a $2.00 / $10.00 list price, no published end date). Budget against list if you’re planning past this quarter. Best For Provider Recommendation Why Lowest price with a 0% error rate DeepInfra Standard $1.90 input / $9.50 output / […]</p>
© 2026 DeepInfra. All rights reserved.