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

Updated November 2026 for the DeepSeek-V4-Pro-0813 release.
DeepSeek’s GA release of V4 Pro moves its DeepSWE score from 12.8 to 62.7 and Terminal Bench 2.1 from 72.1 to 87.9, according to DeepSeek’s model card (vendor-reported). DeepSeek-V4-Pro-0813 shipped on August 13, 2026 and supersedes the April preview. It keeps the same 1.6-trillion-parameter Mixture of Experts design with 49B active parameters and a 1M-token context window, and adds new post-training, a built-in speculative decoding module and a three-level reasoning_effort control.
The DeepSeek V4 Pro 0813 release is built for agentic work: coding agents that edit whole repositories, terminal and tool-use loops, security tasks and long multi-step automations. DeepSeek-V4-Pro-0813 is now available on DeepInfra, live since August 18, 2026, at $1.30 per 1M input tokens and $2.60 per 1M output tokens. That is the lowest output price among the eight providers Artificial Analysis measures for this model, and the Flex tier takes output down to $2.08.
The weights are on Hugging Face under an MIT license, so self-hosting stays an option. DeepInfra runs the model for you on an OpenAI-compatible endpoint, so you can skip provisioning a 1.6T-parameter deployment.
| Benchmark | V4-Pro-0813 | V4-Pro (Preview) | Δ vs preview | GLM-5.2 | Kimi K3 | Opus-4.8 | Fable-5 (w/ fallback) |
|---|---|---|---|---|---|---|---|
| HLE (no tools) | 42.7 | 37.7 | +5.0 | 40.5 | 43.5 | 49.8 | 53.3 |
| HLE (with tools) | 60.0 | 48.2 | +11.8 | 54.7 | 56.0 | 57.9 | 63.0 |
| Terminal Bench 2.1 | 87.9 | 72.1 | +15.8 | 81.0 | 88.3 | 85.0 | 88.0 |
| NL2Repo | 61.5 | 38.5 | +23.0 | 48.9 | – | 69.7 | – |
| Cybergym | 83.3 | 52.7 | +30.6 | – | 80.0 | 78.3 | 83.1 |
| DeepSWE | 62.7 | 12.8 | +49.9 | 46.2 | 67.5 | 58.0 | 70.0 |
| Toolathlon-Verified | 74.1 | 55.9 | +18.2 | 59.9 | 76.5 | 76.2 | 77.9 |
| Agents’ Last Exam | 25.7 | 16.5 | +9.2 | 23.8 | 27.6 | 25.7 | – |
| AutomationBench (Public) | 31.8 | 12.8 | +19.0 | 12.9 | 30.8 | 27.2 | 29.1 |
| DSBench-FullStack* | 71.1 | 41.8 | +29.3 | 61.8 | 73.7 | 71.6 | 77.2 |
| DSBench-Hard* | 67.2 | 31.1 | +36.1 | 54.5 | 63.0 | 71.7 | 68.3 |
Internal DeepSeek test sets. All scores are vendor-reported from the DeepSeek-V4-Pro-0813 model card, run by DeepSeek at max effort; peer columns are as DeepSeek reported them.
0813 improves on the preview in every row. The biggest jumps are on agentic coding: DeepSWE (+49.9 points), Cybergym (+30.6) and repository-level generation on NL2Repo (+23.0). It leads the whole comparison set on Cybergym (83.3, just ahead of Fable-5 at 83.1) and AutomationBench (31.8). Among the open-weights models in the table, it also leads HLE with tools, NL2Repo and DSBench-Hard.
It doesn’t win everywhere. Closed models lead most rows: Fable-5 on HLE, DeepSWE, Toolathlon and DSBench-FullStack, and Opus-4.8 on NL2Repo and DSBench-Hard. The open-weights Kimi K3 edges 0813 on Terminal Bench 2.1 (88.3 vs 87.9), DeepSWE (67.5 vs 62.7), Toolathlon, Agents’ Last Exam and HLE without tools. DeepSeek ran every number above, and the two DSBench rows use internal test sets that nobody else can reproduce.
For an independent view, Artificial Analysis scores DeepSeek V4 Pro 0813 at 23 on its Intelligence Index v5.0.1 at Max effort. That equals the median for open-weights models of its size. On the same index, DeepSeek’s newer V4.1 Flash scores higher. Read the table above as evidence for agentic coding and tool use, not as a general intelligence ranking.
DeepSeek V4 Pro is a 1.6-trillion-parameter Mixture of Experts model that activates 49B parameters per token. Each token is routed to 6 of 384 experts plus a shared expert, so you get the capacity of a very large model at the per-token compute of a much smaller one.
The main architectural change from DeepSeek V3.2 is hybrid attention, which combines Compressed Sparse Attention (CSA) with Heavily Compressed Attention (HCA). DeepSeek says that at 1M tokens of context, V4 Pro needs 27% of the single-token inference FLOPs and 10% of the KV cache that V3.2 needs. Manifold-Constrained Hyper-Connections (mHC) stabilize signal flow across the 61 layers, and pre-training used the Muon optimizer on more than 32 trillion tokens. 0813 keeps all of this. The base structure is unchanged from the preview.
DeepSeek describes 0813 as the official release of V4 Pro, built on the preview’s model structure with new post-training and a speculative decoding module attached. In practice:
reasoning_effort replaces reasoning modes. The preview had three modes (Non-think, Think High, and Think Max set through a special system prompt). 0813 uses one parameter with three levels: low for simple tasks, high for everyday agent workflows and max for the hardest problems.high and max effort, DeepSeek recommends allowing up to 384K output tokens.If you built on the preview, check four things:
reasoning_effort instead. On DeepInfra, "none" turns thinking off entirely, and the reasoning docs list "high" as the default.max_tokens for long reasoning. If you leave max_tokens unset, DeepInfra uses the smaller of the remaining context and 65,536 tokens. At max effort, set it explicitly. The ceiling on DeepInfra is 393,216 (384K) output tokens.deepseek-ai/DeepSeek-V4-Pro-0813.The model is text-only, so use DeepInfra’s multimodal model catalog if your pipeline needs image input. Deep reasoning also takes time: at Max effort, Artificial Analysis measures tens of seconds before the first answer token on every provider. Faster endpoints exist (DigitalOcean leads output speed on Artificial Analysis); if response speed is your product, test them. For batch, agents and cost-sensitive workloads, the per-token gap compounds.
DeepSeek-V4-Pro-0813 is available now in DeepInfra’s text generation catalog. Pricing is per token, with no minimums, and there’s no promotional discount on this model, so the standard rate is the list price to budget against.
| Tier | Input | Output | Cached input |
|---|---|---|---|
| Standard | $1.30 | $2.60 | $0.10 |
| Flex | $1.04 | $2.08 | $0.08 |
| Priority | $1.95 | $3.90 | $0.15 |
Per 1M tokens. Pricing as of 2026-11-16. Source: model page and pricing page.
Output is where the price gap shows. On Artificial Analysis’s providers page, six of the other seven providers list $3.96 per 1M output tokens, and the seventh lists $3.17. DeepInfra’s $2.60 is 34% lower than $3.96, and Flex at $2.08 is 47% lower. Artificial Analysis shows DeepSeek’s own API at its peak rate; DeepSeek charges half that off-peak. Artificial Analysis also recorded a 0% error rate on DeepInfra’s endpoint. Long agent runs produce a lot of output, including reasoning tokens, so the output rate tends to drive the bill.
Choosing a tier:
"service_tier": "flex". A Flex request waits up to 10 minutes for capacity and then returns HTTP 429, and rejected Flex requests aren’t charged. It suits batch jobs, evals and background agents. See how the Flex tier works."service_tier": "priority" for latency-sensitive traffic that shouldn’t queue.Model specs on DeepInfra:
deepseek-ai/DeepSeek-V4-Pro-0813max_tokens is 65,536 if unsetreasoning_effort: low, high or max (plus "none" to disable thinking)The endpoint is OpenAI-compatible. Point your client at https://api.deepinfra.com/v1/openai, call /chat/completions and authenticate with Authorization: Bearer $DEEPINFRA_API_KEY. DeepInfra runs a zero-retention data policy and is SOC 2 and ISO 27001 certified. If you need dedicated compute, GPU instances are available for private deployments.
Get an API key from the DeepInfra dashboard and export it as DEEPINFRA_API_KEY.
curl
curl "https://api.deepinfra.com/v1/openai/chat/completions" \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $DEEPINFRA_API_KEY" \
-d '{
"model": "deepseek-ai/DeepSeek-V4-Pro-0813",
"reasoning_effort": "high",
"messages": [
{
"role": "user",
"content": "Explain the difference between CSA and HCA attention in one paragraph."
}
]
}'
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["DEEPINFRA_API_KEY"],
base_url="https://api.deepinfra.com/v1/openai",
)
response = client.chat.completions.create(
model="deepseek-ai/DeepSeek-V4-Pro-0813",
messages=[
{
"role": "user",
"content": "Review this function for race conditions and propose a fix:\n\n<paste code here>",
}
],
max_tokens=131072, # unset defaults to 65,536; the ceiling is 393,216
top_p=0.95, # DeepSeek's recommendation for agentic use
extra_body={
"reasoning_effort": "max", # low | high | max ("none" disables thinking)
"service_tier": "flex", # 0.8x price; drop this line for standard
},
)
print(response.choices[0].message.content)
JavaScript
import OpenAI from "openai";
const openai = new OpenAI({
apiKey: process.env.DEEPINFRA_API_KEY,
baseURL: "https://api.deepinfra.com/v1/openai",
});
const response = await openai.chat.completions.create({
model: "deepseek-ai/DeepSeek-V4-Pro-0813",
reasoning_effort: "high",
messages: [{ role: "user", content: "Explain the difference between CSA and HCA attention in one paragraph." }],
});
console.log(response.choices[0].message.content);
To try prompts before writing code, open the interactive demo on the DeepSeek-V4-Pro-0813 model page. The API reference lists every supported parameter, including reasoning_effort, service_tier and prompt-caching options. For tool use, see function calling on DeepInfra.
DeepSeek V4 Pro 0813 is a large step up from the preview where it counts for agent builders: repository-scale coding, terminal and tool-use loops, and long automations. Teams running coding agents, security tooling or multi-step workflows on open weights should evaluate it, ideally on their own tasks, since the benchmark gains are vendor-reported. The 1M context, 384K output ceiling and three-level reasoning_effort give you room to trade depth for cost per request.
It isn’t the default answer for every workload. On Artificial Analysis’s general index it sits at its class median, and DeepSeek’s newer DeepSeek-V4.1-Flash scores higher there at a fraction of the per-token price. Route classification, extraction and other easy steps to V4.1 Flash, and keep V4 Pro 0813 for the agentic coding work where its results are strongest. Mixture of Experts economics explains why per-token pricing on models this size has dropped so far.
For deeper numbers, see the DeepSeek V4 Pro pricing guide, the provider speed and latency benchmarks, the model overview and our list of the best API providers for DeepSeek V4.
Building a startup? DeepStart gives eligible startups about 50 billion free tokens (counted at DeepSeek V4 Flash prices) to spend on any model, including V4 Pro 0813. Or head to the model page, grab an API key and make your first call.
Best API Providers for NVIDIA Nemotron 3 Super 120B<p>Nemotron 3 Super 120B is available across a growing number of hosted APIs and deployment platforms. At 120B total parameters with 12B active per inference pass, the right provider matters: latency, throughput, and cost vary significantly depending on where you run it. This guide covers the top options by use case — from fully managed […]</p>
Langchain improvements: async and streamingStarting from langchain
v0.0.322 you
can make efficient async generation and streaming tokens with deepinfra.
Async generation
The deepinfra wrapper now supports native async calls, so you can expect more
performance (no more t...
DeepSeek V4.1 Flash 1017 on DeepInfra: Pricing & API<p>Updated November 2026 for DeepSeek-V4.1-Flash-1017. DeepSeek’s October update to V4.1-Flash lifts Terminal-Bench 4.0 from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0, according to DeepSeek’s model card (vendor-reported). DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026 as a minor update with “substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks”, plus shorter task […]</p>
© 2026 DeepInfra. All rights reserved.