DeepSeek V4 Pro: A Mathematical Reasoning Walkthrough

Step 1: Give DeepSeek V4 Pro precise mathematical constraints

A useful mathematical workflow separates three jobs: defining the problem, generating a candidate solution, and verifying the result. This walkthrough applies that structure to DeepSeek V4 Pro using a small equation whose misleading candidate makes verification essential.

Start with this prompt: Solve sqrt(x + 5) = x − 1 over the real numbers. State the domain restrictions, show a concise derivation, and verify every candidate in the original equation. Write the explanation in English.

The restrictions matter as much as the equation. The square root requires x ≥ −5, while its nonnegative value forces x − 1 ≥ 0. Together, these conditions require x ≥ 1. Establish that admissible domain before manipulating either side.

Step 2: Configure DeepSeek V4 Pro reasoning deliberately

DeepSeek documents the API identifier as deepseek-v4-pro. Use the documented identifier rather than assuming that a display label containing spaces is also a valid request value.

For Chat Completions, DeepSeek documents thinking as enabled by default, with high reasoning effort. Its supported effort levels include low, high, and max. As a workflow choice, begin with high for your evaluation set, then compare other settings against measured correctness and latency.

Do not use temperature adjustments as your main reasoning control: DeepSeek states that temperature has no effect in thinking mode. Change the reasoning effort or improve the problem specification instead.

This article is published by BestRelay, a quality-first API relay whose stated policy is to forward each request to the official model named in it without substitution. Its compatibility includes Claude Code, Codex, Gemini CLI, and any OpenAI SDK. Keep the requested model explicit when configuring your client. Create a BestRelay account to try this workflow.

Step 3: Make DeepSeek V4 Pro produce checkable mathematics

Ask for a compact solution organized around assumptions, transformations, candidates, and verification. The goal is an auditable mathematical argument, not an unnecessarily long explanation.

For the example, squaring both sides gives x + 5 = (x − 1)². Expanding and rearranging produces x² − 3x − 4 = 0, which factors as (x − 4)(x + 1) = 0. The algebra therefore produces two candidates: 4 and −1.

Require the response to distinguish candidates from solutions. Squaring preserves any original solution, but the resulting equation can admit additional values. Here, −1 already violates the admissible domain x ≥ 1.

Step 4: Verify DeepSeek V4 Pro answers independently

Substitute both candidates into the original equation. For x = 4, the left side is sqrt(9) = 3 and the right side is 3. For x = −1, the left side is sqrt(4) = 2 while the right side is −2. Consequently, the solution set is exactly {4}.

For larger problems, make verification a separate stage:

• Check substitutions using exact arithmetic where possible.
• Review denominators, logarithm arguments, and inequality directions.
• For proofs, inspect every implication and exceptional case.
• Treat numerical experiments as tests, not general proofs.

If you add executable verification through function tools, preserve the required conversation fields. DeepSeek requires previous reasoning_content to be passed back on subsequent requests carrying tools; omitting it can produce a request error.

Step 5: Measure DeepSeek V4 Pro reliability and cost

Build a small evaluation set containing equations, inequalities, counting problems, and proof tasks. Record the prompt, expected result, returned answer, verification outcome, reasoning setting, latency, and reported usage. Score mathematical correctness separately from explanation quality.

Allow enough output capacity for completion. In DeepSeek's Responses API, max_output_tokens covers both reasoning tokens and visible output tokens, so inspect completion status before accepting a response as finished.

BestRelay describes its billing as pay-as-you-go using provider-reported token counts, with cache reads and writes priced separately. Track those categories separately when comparing settings. Its member referral commission applies when invited users top up; keep that commercial incentive separate from technical evaluation. The acceptance criterion remains a verified mathematical result.