Back to models
NewChat Models

DeepSeek

DeepSeek V4 Pro

Best for Research

V4 Pro 0423 on OpenRouter for long-context reasoning, with provider-specific pricing and direct API version caveats.

Deep reasoning1M context

At a glance

Know the model before you prompt.

OpenRouter API specifications
Context window
1,048,576 tokens
Maximum output
393,216 tokens
Inputs → output
Text → Text
Knowledge cutoff
Not verified

API model ID: deepseek/deepseek-v4-pro

Capabilities & boundaries

What it supports. Where the limits are.

Tool support requires the appropriate API integration; a supported tool is not automatically active in every chat.

Supported API features and tools

  • The OpenRouter V4 Pro 0423 listing, with 1,048,576 context tokens and up to 393,216 completion tokens at the model-listing level
  • A mixture-of-experts architecture described as 1.6T total parameters with 49B active per token
  • Reasoning-oriented text generation for code, analysis and multi-step tasks
  • Tool-call workflows through compatible hosting endpoints and a caller-managed integration

Before you choose

  • Model-level token ceilings are not guarantees for every host. Check the serving endpoint's input/output limits and parameter support.
  • The direct DeepSeek alias, OpenRouter's 0423 listing and the separately named 0813 version are not equivalent identifiers.
  • The reviewed listing does not establish a training cutoff or native audio/video output. Do not borrow V4.1 Flash vision features.
  • Long responses and additional reasoning can increase latency and cost. Ask for bounded deliverables and validate important conclusions.

Choose the reasoning effort

highxhigh

OpenRouter's listing describes high and xhigh reasoning, with xhigh corresponding to the model's max effort. No universal default is asserted here. Check endpoint support and do not confuse these settings with the newer direct-API effort mappings.

  • The OpenRouter slug deepseek/deepseek-v4-pro currently labels the 0423 version in the reviewed source. Keep the source date alongside benchmarks and saved prompts.
  • The table below is a named hosting route, not a lowest-price claim or the direct DeepSeek V4 Pro 0813 tariff.
  • Confirm tool-call schemas and permissions in the integration. Ask for evidence and verification steps rather than assuming more reasoning eliminates mistakes.

Put it to work

Start with a more useful prompt.

Original examples from EZ Ai Assist. Adapt these to your task and the features available in your workspace.

Workflow 01

Stress-test an architecture proposal

Look for concrete failure conditions before approving a large change.

Evaluate the architecture proposal below against our stated scale, reliability and migration constraints. Identify the three strongest failure scenarios, the evidence for each and the smallest experiment that would test it. Distinguish a demonstrated defect from a speculative risk. Finish with a staged recommendation and explicit rollback conditions.

Workflow 02

Reconcile conflicting evidence

Make contradictions visible instead of smoothing them away.

Read these dated source documents and build an evidence matrix for the decision we need to make. Identify agreements, contradictions and facts that may have changed over time. Cite the relevant source and date for each claim. Recommend what to verify next, and do not choose an unsupported fact merely to make the story consistent.

Workflow 03

Design an evaluation for a difficult task

Translate a complex problem into a measurable acceptance test.

Using this task specification and sample failures, design an evaluation for a reasoning model. Define success criteria, adversarial edge cases, an evidence-checking procedure and a scoring sheet. Separate model mistakes from missing inputs or tool failures. Suggest an efficient test order and a clear stop condition without inventing benchmark results.

Developer reference

OpenRouter API pricing

These are named-provider OpenRouter reference prices, not EZ Ai Assist subscription prices.

View EZ Ai Assist plans
Relace route reference on OpenRouter · USD per 1,000,000 tokens
Token typePrice
Input$0.2067
Output$4.20
  • These are the Relace route's displayed input/output rates for the 0423 listing checked October 6, 2026, not a tariff shared by all OpenRouter providers.
  • Other serving routes can charge different rates or offer different discounts. Check the actual endpoint and billed usage; no universal cache rate is assumed.
  • The current direct DeepSeek table and September routing notice conflict about the older Pro alias. Do not use this OpenRouter table to estimate direct-API charges.

Common questions

A few things worth knowing.

Is this the V4 Pro 0813 model?

No. The supplied OpenRouter page labels deepseek/deepseek-v4-pro as V4 Pro 0423. DeepSeek's direct pricing table names 0813 separately; this guide does not silently substitute it.

What happens to the direct deepseek-v4-pro alias?

The September release notice says it redirects to V4.1 Flash from September 14 until V4.1 Pro launches, but the current pricing table still names V4 Pro 0813. Verify the actual service before relying on either interpretation.

What are the listed token limits?

OpenRouter lists a 1,048,576-token context and up to 393,216 completion tokens for this model. The selected host can impose narrower limits, so these are not a promise for every endpoint or app workspace.

What does xhigh mean here?

The OpenRouter description maps xhigh to the model's max effort. It can require more computation and time; test it against high on representative tasks instead of assuming it is always the better choice.

Is the price a DeepSeek subscription price?

No. It is the named Relace hosting route's token pricing on OpenRouter. Direct DeepSeek API rates and EZ Ai Assist subscription prices are separate.

When is an older pinned version useful?

A fixed version can help repeat an evaluation or investigate a regression. Pin the host and settings too, and maintain a replacement plan because hosting availability and alias behavior can change.

Will it verify every claim automatically?

No. Supply the evidence, ask for source references and check the conclusions. External verification needs an available tool or a human review; the model name alone does not provide either.

Check the source

Official documentation

Specifications and API prices checked on . Example prompts and workflow advice are editorial guidance from EZ Ai Assist.

Same provider