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

DeepSeek V4 Pro 0813 API Benchmarks: Speed, Latency & Cost

Published on 2026.04.30 by DeepInfra
DeepSeek V4 Pro 0813 API Benchmarks: Speed, Latency & Cost

Updated November 2026 for the DeepSeek-V4-Pro-0813 release.

About DeepSeek V4 Pro 0813

DeepSeek-V4-Pro-0813 is the general-availability release of DeepSeek V4 Pro, published by DeepSeek on August 13, 2026. It supersedes the April preview and keeps the same Mixture-of-Experts base: 1.6 trillion total parameters, 49 billion active per token, and a 1M-token context window. What’s new is the post-training, a built-in DSpark speculative decoding module, and a three-level reasoning_effort control (low, high, max).

The architecture still pairs Compressed Sparse Attention (CSA) with Heavily Compressed Attention (HCA). DeepSeek reports that at 1M tokens of context the V4 Pro design needs 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2 (vendor-reported). Weights are MIT-licensed.

This page covers the Max reasoning setting, which is the configuration Artificial Analysis measures across providers. At Max effort the model reasons at length before answering, so two latency numbers matter: time to first token (when the stream starts) and time to first answer token (when the reasoning ends and the answer begins). For a reasoning model at this effort level, the second one is what your users wait for.

Below is the November 2026 comparison of every provider Artificial Analysis currently benchmarks for DeepSeek V4 Pro 0813, ordered by what most teams pay attention to first: the bill.

DeepSeek V4 Pro 0813 (Max) API Review Summary

  • Providers benchmarked: 8: Baseten, DeepInfra, DeepSeek, DigitalOcean, GMI, Nebius, Novita and SiliconFlow (Artificial Analysis providers page). All 8 support JSON mode and function calling.
  • Lowest output price among the 8 providers Artificial Analysis benchmarks: DeepInfra at $2.60 per 1M tokens. Most providers charge $3.96; GMI charges $3.168.
  • Lowest blended price: GMI at $0.55 per 1M tokens, then DeepInfra at $0.59. Five providers sit at $0.69 and Baseten at $0.75, a 1.4× spread.
  • Fastest output and lowest time to first answer token: DigitalOcean at 140.0 tokens per second and 31.67 seconds.
  • Lowest time to first token: DigitalOcean and SiliconFlow, tied at 1.72 seconds.
  • Intelligence: 23 on the Artificial Analysis Intelligence Index v5.0.1 (Max), equal to the median of 23 for open-weights models of similar size.
  • Verbosity: 119M output tokens to run the Intelligence Index, against a class median of 104M. Output price carries more weight than usual for this model.
  • Cost per Index task (DeepSeek’s own API): $1.77; 17.6 minutes per task against a class median of 13.5.
  • Context and output on DeepInfra: 1M-token context, up to 384K output tokens, served in fp8.

DeepSeek V4 Pro 0813 (Max): Provider Comparison Data

Prices are per 1M tokens. Blended price uses Artificial Analysis’s current 7:2:1 weighting of cache-hit, input and output tokens. Speed and latency are Artificial Analysis headline values.

ProviderInputOutputCache hitBlended (7:2:1)Output speed (t/s)Time to first answer token (s)Best for
DeepInfra$1.30$2.60$0.10$0.5965.167.09Cost: lowest output price, Flex tier, 0% error rate
GMI$1.056$3.168$0.035$0.5573.253.39Lowest blended price
DeepSeek (first-party)$1.32$3.96$0.044$0.6993.552.73Baseline; cheapest per Index task
Novita$1.32$3.96$0.044$0.6977.959.31Cache-heavy agent work, 0% errors
Nebius$1.32$3.96$0.04$0.6993.945.40Mid-pack speed and latency
DigitalOcean$1.32$3.96$0.044$0.69140.031.67Speed and latency
SiliconFlow$1.32$3.96$0.044$0.6941.473.93Low time to first token (1.72 s)
Baseten$1.32$3.96$0.132$0.75114.243.87Second-fastest output

Methodology. Data from the Artificial Analysis providers page for DeepSeek V4 Pro 0813 (Max), pulled November 16, 2026. Artificial Analysis now runs its “Performance v3” workloads: the headline figures average single-request and multi-turn agentic runs over the past five days, benchmarked every six hours. The blended price is now a 7:2:1 cache-hit:input:output blend, so it can’t be compared with the 3:1 input:output blend ($2.17) this page used in April. DeepSeek’s first-party prices are shown at its peak rate; off-peak is half. Figures change over time. Pricing as of 2026-11-16.

Quick Verdict: Which DeepSeek V4 Pro API Provider Is Best?

On cost, DeepInfra is the recommended DeepSeek V4 Pro API. It has the lowest output price of the eight benchmarked providers ($2.60 per 1M, 34% below the $3.96 most providers charge), the second-lowest blended price ($0.59, behind GMI’s $0.55), a Flex tier that cuts every rate by 20%, and a 0% error rate over Artificial Analysis’s measurement window.

DigitalOcean is the speed and latency leader: 140.0 tokens per second and a 31.67-second time to first answer token, the best on both measures. If response speed is your product, test DigitalOcean; for cost-sensitive, batch and agent workloads, DeepInfra’s per-token rates compound.

Overall Recommendation for Cost: DeepInfra

DeepSeek V4 Pro 0813 on DeepInfra costs $1.30 per 1M input tokens, $2.60 per 1M output tokens and $0.10 per 1M cached input tokens. There’s no promotional discount; that’s the list price. It is lower than in April, so if you priced the model here then, re-run your numbers.

TierInputOutputCached input
Standard$1.30$2.60$0.10
Flex$1.04$2.08$0.08

Per 1M tokens. Pricing as of 2026-11-16.

Why the output rate matters here. DeepSeek V4 Pro 0813 at Max effort writes a lot: 119M output tokens to complete the Artificial Analysis Intelligence Index, against a 104M median for its class. Reasoning tokens are billed as output on DeepInfra, so the output rate is where most of the bill accumulates. A batch job that produces 10M output tokens costs $26.00 on DeepInfra Standard, $31.68 on GMI, and $39.60 at the $3.96 rate most other providers charge.

Flex for batch and agents. The Flex service tier prices the model at $1.04 / $2.08 / $0.08. Send "service_tier": "flex" and the request waits up to 10 minutes for capacity before returning HTTP 429; a rejected Flex request isn’t charged. At $2.08, Flex output is about 47% below the $3.96 standard rate elsewhere and below every other provider’s measured output price. At the same 7:2:1 weighting Artificial Analysis uses, Flex works out to about $0.47 blended (our calculation, not an Artificial Analysis figure).

Reliability. DeepInfra recorded a 0.0% error rate and a 98.3% cache hit rate over Artificial Analysis’s measurement window.

  • Blended Price: $0.59 per 1M tokens (Standard); input $1.30, output $2.60, cached $0.10
  • Output Speed: 65.1 t/s
  • Latency (TTFT): 2.17 s to first token; 67.09 s to first answer token
  • Optimal Workload: output-heavy agent loops, batch reasoning jobs on Flex, long-context work that needs the full 1M window

DeepInfra serves the model in fp8 with the full 1M-token context and up to 384K output tokens; if you don’t set max_tokens, it defaults to the smaller of the remaining context and 65,536. JSON mode, structured output and tool calling are supported, and reasoning_effort accepts none through max (none turns thinking off). The original deepseek-ai/DeepSeek-V4-Pro model ID also remains available on DeepInfra; new integrations should use the 0813 ID.

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 diff and list the risky changes."}],
    max_tokens=32768,
    extra_body={
        "reasoning_effort": "max",
        "service_tier": "flex",  # remove for standard tier
    },
)
print(response.choices[0].message.content)
copy

The endpoint is /chat/completions on that base URL, authenticated with Authorization: Bearer $DEEPINFRA_API_KEY. Full parameters are on the DeepSeek-V4-Pro-0813 API page.

Start using DeepSeek V4 Pro 0813 on DeepInfra →

Provider Analyses

1. DigitalOcean: Speed and Latency Leader

DigitalOcean leads both speed measures that matter for a reasoning model. Its 140.0 t/s output speed is the highest in the set, and its 31.67-second time to first answer token is the shortest. It also ties SiliconFlow for the lowest time to first token at 1.72 seconds and completes an Index task in 11.8 minutes, the fastest of the eight.

The trade-offs are price and error rate. DigitalOcean charges the common $1.32 / $3.96 rate ($0.69 blended), and its 3.1% error rate is the highest Artificial Analysis recorded for this model. For interactive products where users watch the model think, it’s the first provider to test.

  • Blended Price: $0.69 per 1M tokens
  • Output Speed: 140.0 t/s
  • Latency (TTFT): 1.72 s; 31.67 s to first answer token
  • Optimal Workload: user-facing chat and coding assistants where the wait for the answer is the product metric

2. Baseten: High-Throughput Alternative

Baseten is second on output speed at 114.2 t/s and second on time to first answer token at 43.87 seconds, with a 0.8% error rate. It is also the most expensive option in the set at $0.75 blended, because its cache-hit rate ($0.132) is the highest listed.

  • Blended Price: $0.75 per 1M tokens
  • Output Speed: 114.2 t/s
  • Latency (TTFT): 2.84 s; 43.87 s to first answer token
  • Optimal Workload: fast generation when DigitalOcean’s error rate is a concern

3. GMI: Lowest Blended Price

GMI has the lowest blended price at $0.55, thanks to the lowest input rate ($1.056) and a $0.035 cache-hit rate. Its output rate ($3.168) is higher than DeepInfra’s $2.60, so the gap between the two narrows or reverses as your workload shifts toward output tokens. GMI recorded a 0% error rate, 73.2 t/s and a 53.39-second time to first answer token.

  • Blended Price: $0.55 per 1M tokens
  • Output Speed: 73.2 t/s
  • Latency (TTFT): 3.11 s; 53.39 s to first answer token
  • Optimal Workload: input-heavy and cache-heavy traffic with short outputs

4. DeepSeek (First-Party API): The Baseline

DeepSeek’s own API serves 0813 under the deepseek-v4-pro model name. Artificial Analysis lists it at $1.32 / $3.96 with a $0.044 cache-hit rate, which is DeepSeek’s peak price. DeepSeek now prices by time of day: off-peak rates are half ($0.66 input, $1.98 output, $0.022 cache hit), with peak hours set at 01:00–04:00 and 06:00–10:00 UTC on weekdays. DeepInfra’s $2.60 output sits below DeepSeek’s peak output price and above its off-peak one.

Because its cache reads are cheap and its cache hit rate high (99.1%), DeepSeek’s API has the lowest cost per Intelligence Index task at $1.77. It measured 93.5 t/s and a 52.73-second time to first answer token, with a 0.8% error rate.

  • Blended Price: $0.69 per 1M tokens (peak pricing)
  • Output Speed: 93.5 t/s
  • Latency (TTFT): 2.57 s; 52.73 s to first answer token
  • Optimal Workload: teams that want first-party access, can schedule work off-peak, or use DeepSeek-only features such as its Anthropic-format endpoint

5. Novita, Nebius and SiliconFlow

The other three providers all charge $1.32 / $3.96 ($0.69 blended):

  • Novita: 77.9 t/s, 59.31 s to first answer token, 0% error rate, and the second-lowest cost per Index task ($1.86).
  • Nebius: 93.9 t/s and the third-lowest time to first answer token (45.40 s), but a 54.9% cache hit rate pushed its cost per Index task to $13.43, the highest in the set.
  • SiliconFlow: ties for the lowest time to first token (1.72 s), but at 41.4 t/s it has the slowest output and the longest time to first answer token (73.93 s).

Technical Deep-Dive: What Developers Need to Know

1. Time to First Token vs Time to First Answer Token

  • Time to first token (TTFT): how long until the first streamed token arrives. For DeepSeek V4 Pro at Max effort, that’s usually the start of reasoning.
  • Time to first answer token: how long until reasoning finishes and the answer starts.

Across the eight providers, TTFT ranges only from 1.72 to 3.11 seconds, while time to first answer token ranges from 31.67 to 73.93 seconds. A provider with a fast first token can still keep users waiting a minute for the answer. Judge interactive deployments on time to first answer token.

2. Reasoning Effort Changes Both Cost and Latency

The figures above are for Max. On DeepSeek’s own API, Artificial Analysis measured High effort at 42.81 seconds to first answer token and Non-reasoning mode at 2.48 seconds, compared with 52.73 seconds at Max. If a task doesn’t need Max, setting reasoning_effort to high or low cuts both wait time and output tokens, and output tokens are the larger cost line for this model.

3. Blended Price Depends on Your Cache Mix

Artificial Analysis’s 7:2:1 blend assumes 70% of tokens are cache hits. That suits long agent sessions that resend the same context. If your traffic is output-heavy (long reasoning, long code generation), output price dominates and DeepInfra’s $2.60 rate matters most. If it’s mostly cache hits with short answers, providers with $0.035–$0.044 cache reads close the gap. Run your own token mix through each provider’s rates before committing.

4. Intelligence Context

DeepSeek V4 Pro 0813 (Max) scores 23 on the Artificial Analysis Intelligence Index v5.0.1, equal to the class median of 23 for open-weights models of similar size. The April preview has no score on this index version, and Artificial Analysis cautions that scores from different index versions aren’t comparable. DeepSeek’s own benchmark table shows large agentic gains over the preview (for example DeepSWE from 12.8 to 62.7 and Terminal Bench 2.1 from 72.1 to 87.9, vendor-reported). Artificial Analysis currently scores DeepSeek V4.1 Flash higher than V4 Pro 0813. The DeepSeek V4 Pro model overview covers the full benchmark table.

5. API Feature Parity

All 8 providers support JSON mode and function calling, so the model can sit behind a router with one provider for cost and another for latency-sensitive traffic without code changes beyond the base URL and model name.

FAQ

What is the cheapest DeepSeek V4 Pro API?

It depends on the token mix. DeepInfra has the lowest output price ($2.60 per 1M, or $2.08 on Flex). GMI has the lowest blended price at $0.55 per 1M under Artificial Analysis’s cache-heavy 7:2:1 weighting, with DeepInfra second at $0.59. DeepSeek’s own API is cheaper off-peak ($0.66 / $1.98) but costs $1.32 / $3.96 at peak.

Which provider is fastest for DeepSeek V4 Pro 0813?

DigitalOcean, at 140.0 tokens per second and a 31.67-second time to first answer token. Baseten is second on both (114.2 t/s, 43.87 s).

What context window does DeepInfra offer for DeepSeek V4 Pro 0813?

The full 1M tokens, served in fp8, with up to 384K output tokens. The 66K limit this page reported in April no longer applies.

What is the DeepSeek V4 Pro 0813 model ID on DeepInfra?

deepseek-ai/DeepSeek-V4-Pro-0813. It’s OpenAI-compatible at https://api.deepinfra.com/v1/openai.

How does DeepSeek V4 Pro 0813 score on independent benchmarks?

23 on the Artificial Analysis Intelligence Index v5.0.1 at Max effort, equal to the class median. Benchmark scores in DeepSeek’s release table are vendor-reported.

What is the difference between time to first token and time to first answer token?

Time to first token measures when the stream starts, which for a reasoning model is usually the start of its thinking. Time to first answer token measures when the answer itself starts. For DeepSeek V4 Pro 0813 at Max effort, the first is about 2–3 seconds on every provider, while the second ranges from about 32 to 74 seconds.

Conclusion

  • Lowest output price among the providers in this guide: DeepInfra ($2.60 output, $0.59 blended; Flex at $2.08 output)
  • Lowest blended price: GMI ($0.55)
  • Speed and latency: DigitalOcean (140.0 t/s, 31.67 s to first answer token); Baseten as the alternative
  • Baseline and off-peak pricing: DeepSeek’s first-party API

For pricing scenarios, see the DeepSeek V4 Pro pricing guide; for a wider provider list, see best API providers for DeepSeek V4. To run the model on dedicated GPUs with your own scaling, deploy a private endpoint, or start with the DeepSeek V4 Pro 0813 API on DeepInfra and run your real prompt mix.

Related articles
MiMo-V2.6-Pro: Model Documentation & Integration GuideMiMo-V2.6-Pro: Model Documentation & Integration Guide<p>Updated November 2026. MiMo-V2.6-Pro is a flagship sparse Mixture-of-Experts (MoE) model built for omnimodal intelligence and autonomous agentic workflows. It has 1.02 trillion total parameters, with 42 billion activated during inference, and is natively built to process text, image, video, and audio within a single, unified architecture. With a 1,048,576-token (1M) context window and optimization [&hellip;]</p>
Kimi K3.1 Pro API Benchmarks: Providers, Speed & CostKimi K3.1 Pro API Benchmarks: Providers, Speed & Cost<p>Kimi K3.1 Pro is Moonshot AI&#8217;s new open-weights flagship, and it tops the Artificial Analysis Intelligence Index among open-weights models of its size. It is also a large, talkative reasoning model, so where you run it changes both the bill and the wait. This guide compares the Artificial Analysis data for the 16 endpoints from [&hellip;]</p>
Qwen3.8-27B Pricing: DeepInfra vs Alibaba APIQwen3.8-27B Pricing: DeepInfra vs Alibaba API<p>Qwen3.8-27B is Alibaba&#8217;s 27B open-weight reasoning model, released August 14, 2026 under Apache 2.0. It ships with a large context window, multimodal input, and API availability across eight providers. Pricing varies sharply by host: Alibaba&#8217;s first-party rate runs over 3x DeepInfra&#8217;s for the same weights. Context window figures differ slightly by source. Artificial Analysis reports [&hellip;]</p>