HomeAI models
DeepSeek: DeepSeek V4 Pro 0423
•
DeepSeek V4 Pro 0813 (Reasoning, Max Effort)
•
Released 
April 2026


DeepSeek V4 Pro 0813 (Reasoning, Max Effort)

Open-weight 1.6T-parameter MoE with 1M-token context and three reasoning modes for coding and long-horizon agents.

Modality:
Text
Video
model ID
deepseek/deepseek-v4-pro
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

80.25
tok/s

Intelligence Index *

36.3
/ 100

Context Window *

1048576
tokens

Input price

9.6
Anytoken

Output price

19.2
Anytoken
DeepSeek V4 Pro 0423: Open-Weight Frontier Reasoning at Million-Token Scale DeepSeek V4 Pro 0423 is the preview build of DeepSeek's V4 flagship, a Mixture-of-Experts model with 1.6T total and 49B active parameters, released under the MIT license. It is the most capable model in DeepSeek's V4 lineup, folding general and reasoning behavior into one endpoint with selectable effort modes. The model defaults to a 1M-token context window and targets coding, complex reasoning, and long-horizon agentic workflows. Its strongest draw is near-frontier intelligence with open weights and a substantially lower cost profile than comparable proprietary flagships. Start building with DeepSeek V4 Pro 0423 via the AnyAPI.ai API.

Performance

Where V4 Pro 0423 Earns Its Place: Reasoning and Coding at Scale

V4 Pro 0423 targets hard reasoning, coding, and agentic tasks. Its maximum reasoning-effort mode scores 45 on the Artificial Analysis Intelligence Index, well above the ~27 median for comparable open-weight models, and the tech report reports 80.6% on SWE-bench Verified for V4-Pro-Max. That combination places it near proprietary frontier models while shipping under an MIT license. For production, this means teams can run genuinely difficult reasoning and repository-scale coding workloads with open weights and predictable economics. The trade-off is verbosity: Max mode consumes large output token volumes, so effort selection directly drives cost and latency.

Benchmarks

V4 Pro 0423 Independent Benchmarks: Intelligence, Long Context, Hallucination

On Artificial Analysis, the 0423 preview build scores 45 on the Intelligence Index at Max reasoning effort, ranking among leading open-weight models while remaining very verbose (roughly 180M output tokens across the index). On long-context retrieval, independent testing reports around 83.5 on MRCR 1M, indicating solid recall across the full window, though Claude Opus leads that specific test. A critical caveat: independent measurement puts V4 Pro's hallucination rate near 94% on the AA-Omniscience unknown-answer test, meaning it almost always answers rather than abstaining. Confidence-sensitive applications must add external verification.

Output Speed

*
80.25
tok/s

Intelligence Index

*
36.3
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
93
%

HLE *

Adherence to multi-step structured instructions.
41
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

V4 Pro 0423 is a text-in, text-out Mixture-of-Experts model with a 1M-token default context and up to roughly 384K output tokens. It exposes three reasoning effort modes (Non-Think, High, Max) in a single endpoint, and tool calls function inside thinking mode. The two production-critical facts: the 1M context enables full-repository and large-document workloads without chunking, and reasoning effort is the main cost/latency lever since Max mode is token-intensive. Note there is no image, audio, or video input despite the family's later multimodal siblings.
Verified Specifications — 
DeepSeek V4 Pro 0813 (Reasoning, Max Effort)
*
Input modalities
Text
Video
output modalities
Text
Context window
1048576
 tokens
Maximum output tokens
384000
Reasoning
Yes
Knowledge cutoff
April 2026
Pricing (standard)
9.6
 AnyTokens in
 / 
19.2
 AnyTokens out

Comparison

V4 Pro 0423 vs V4 Flash: When You Actually Need the 1.6T Model

V4 Pro 0423 and V4 Flash ship in the same V4 preview release, share the 1M-token context, the same three reasoning modes, and MIT open weights, so switching between them is a model-name change rather than a re-integration. The difference is scale: Pro activates 49B parameters per token from a 1.6T backbone, while Flash activates roughly 13B from 284B total. The practical decision is how much per-token capability a workload genuinely needs versus how much throughput and cost efficiency you want.

Dimension
DeepSeek V4 Pro 0813 (Reasoning, Max Effort)
DeepSeek V4.1 Flash
Input pricing
9.6
AnyToken
0.9
AnyToken
Output pricing
19.2
AnyToken
3.6
AnyToken
Knowledge cutoff *
April 2026
September 2026

Choose V4 Pro 0423 when the work is genuinely hard: repository-scale debugging, complex reasoning, long-horizon agentic investigations, and knowledge-heavy tasks where the larger active-parameter count matters. Choose V4 Flash when responsiveness and cost dominate — routine transformations, summarization, test generation, chat, and high-volume pipelines. Flash reportedly approaches Pro on simple agent tasks and reasoning with a larger thinking budget, so reserve Pro for cases where Flash measurably falls short.

Limitations & Trade-offs

Where DeepSeek V4 Pro 0813 (Reasoning, Max Effort) falls short

1
High hallucination on unknown answers. Independent measurement reports a hallucination rate near 94% on the AA-Omniscience benchmark, meaning V4 Pro almost never abstains when it doesn't know an answer. This is a specific, measured limitation rather than generic caution. It matters most in factuality-critical or confidence-calibrated applications — customer-facing knowledge answers, medical or legal support, and any workflow where a wrong-but-confident response is costly. For these, pair the model with retrieval grounding and external verification, or prefer a model with better calibration.
2
Verbose, token-intensive reasoning. At Max effort the model generated roughly 180M output tokens across the Artificial Analysis Intelligence Index, well above the median for comparable models. Because output tokens drive both cost and latency, unconstrained Max reasoning can make high-volume use surprisingly expensive and slow. Workloads with tight latency budgets or large request volumes should default to lower effort modes and reserve Max for the hardest tasks where the accuracy gain justifies the token cost.
3
Text-only, no multimodality. V4 Pro 0423 accepts text input and produces text output only — no image, audio, or video. Later family members added native vision, but that capability does not transfer to this build. Any pipeline needing document image parsing, screenshots for computer-use agents, or chart understanding must route those steps to a separate vision model. If multimodal input is central to your product, this model cannot serve as the single backbone.
4
Preview build and lifecycle churn. The 0423 build is a preview; a later GA build (0813) and V4.1 Flash followed, and DeepSeek at one point scheduled — then reversed — routing of the deepseek-v4-pro name to a newer model. Teams pinning behavior to this exact build should verify which underlying version their provider serves and monitor for silent redirection, since the callable name has been reused across builds with materially different benchmark results.

Best-Fit Workloads

Where this model earns its place

01

Repository-scale coding and refactoring

‍
The 1M-token context lets V4 Pro 0423 ingest full codebases without chunking, and V4-Pro-Max reports 80.6% on SWE-bench Verified. This suits long-horizon debugging, architecture changes, and multi-file refactors where whole-repo awareness matters. Note that independent code testing shows stronger performance on syntax-similarity than on functional correctness for some categories, so keep human review in the loop for production commits.

02

Long-horizon agentic workflows
‍

V4 Pro supports tool calling inside thinking mode and leads open-weight peers on agentic benchmarks like GDPval-AA. Combined with a large context, this fits tool-heavy investigations, terminal and computer-use agents, and multi-step task automation. Use High effort for routine agent loops and Max only for complex planning, since Max's token volume compounds cost across long agent trajectories.

03

Long-document analysis and RAG

‍
With a 1M-token window and roughly 83.5 on MRCR 1M retrieval, the model can reason over large document sets in a single pass, reducing chunking complexity in RAG pipelines. It fits legal, financial, and technical document synthesis. Given the high hallucination rate on unknown answers, ground responses in retrieved context and add citation or verification steps for factuality-sensitive outputs.

04

Complex reasoning and STEM problem-solving

‍
Max reasoning effort scores 45 on the Artificial Analysis Intelligence Index, well above the open-weight median, making V4 Pro 0423 suitable for math, science, and multi-step analytical tasks. It fits research assistants and technical decision-support where depth outweighs latency. Reserve Max for genuinely hard problems and drop to High or Non-Think for simpler queries to control verbosity and response time.

Pricing in anytokens via AnyAPI
Input
9.6
₳
Output
19.2
₳
Cache write
—
₳
Cache read
0.03
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsComparisonLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access DeepSeek V4 Pro 0813 (Reasoning, Max Effort) via AnyAPI.ai

Access DeepSeek V4 Pro 0813 (Reasoning, Max Effort) through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate DeepSeek V4 Pro 0813 (Reasoning, Max Effort) without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access DeepSeek V4 Pro 0813 (Reasoning, Max Effort) and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test DeepSeek V4 Pro 0813 (Reasoning, Max Effort) against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use DeepSeek V4 Pro 0813 (Reasoning, Max Effort) from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use DeepSeek V4 Pro 0813 (Reasoning, Max Effort) for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

DeepSeek V4 Pro 0423 has a 1,048,576-token (1M) context window, the default across DeepSeek's V4 services. It supports up to roughly 393,216 (about 384K) completion tokens per response. This lets it process full codebases and large document sets in a single request without chunking.

Yes. V4 Pro 0423 folds reasoning into a single endpoint with three effort modes — Non-Think, High, and Max — selectable per request. It supports function calling via tools and tool_choice, and unlike earlier DeepSeek reasoning models, tool calls work inside thinking mode. It also supports structured JSON-schema outputs.

No. DeepSeek V4 Pro 0423 is text-in, text-out only, with no image, audio, or video input. Later models in the family added native vision, but that build does not include multimodal support. For image or document-image tasks, pair it with a separate vision-capable model.

The 0423 preview build scores 45 on the Artificial Analysis Intelligence Index at Max reasoning effort, above the open-weight median, and the tech report reports 80.6% on SWE-bench Verified for V4-Pro-Max. However, independent testing reports a very high hallucination rate near 94% on unknown-answer tasks, so factuality-sensitive uses need external verification.

DeepSeek V4 Pro 0423 is released as open weights under the MIT license, which permits commercial use and self-hosting. It is also available via API with OpenAI ChatCompletions, Responses, and Anthropic-compatible interfaces. You can access it through the AnyAPI.ai API alongside other models in one integration.

* Benchmark data source: Artificial Analysis artificialanalysis.ai