How to Use the DeepSeek V4 Pro API on EvoLink: First Call to Claude Code

What you will build

1. a first successful V4 Pro request in Anthropic Messages format;
2. a Claude Code setup that runs on V4 Pro through EvoLink;
3. correct thinking-mode control (and why budget_tokens silently does nothing);
4. handling for the three parameter mappings that break Claude migrations;
5. a 429/concurrency strategy and a fallback route for production.

Prerequisites

• An EvoLink account and an API key from the dashboard.
• Any HTTP client. Examples below use cURL and plain Python (requests) so the request shape is explicit.
• The full parameter contract lives in the DeepSeek V4 Messages API documentation; this guide focuses on flow and pitfalls, not on duplicating the reference.

Step 1 — Your first V4 Pro request

curl https://direct.evolink.ai/v1/messages \
  -H "Authorization: Bearer $EVOLINK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-pro",
    "max_tokens": 1024,
    "messages": [
      {"role": "user", "content": "Refactor this function to be iterative: def f(n): return n*f(n-1) if n else 1"}
    ]
  }'

The same call in Python, dependency-light:

import requests, os

resp = requests.post(
    "https://direct.evolink.ai/v1/messages",
    headers={"Authorization": f"Bearer {os.environ['EVOLINK_API_KEY']}"},
    json={
        "model": "deepseek-v4-pro",
        "max_tokens": 1024,
        "messages": [{"role": "user", "content": "Summarize the tradeoffs of MoE routing in two sentences."}],
    },
    timeout=120,
)
resp.raise_for_status()
blocks = resp.json()["content"]
print(next(b["text"] for b in blocks if b["type"] == "text"))

Step 2 — Switch Claude Code to DeepSeek V4 Pro

Because EvoLink exposes V4 Pro on an Anthropic-compatible Messages endpoint, Claude Code can run on it by overriding its endpoint variables:

export ANTHROPIC_BASE_URL="https://direct.evolink.ai"
export ANTHROPIC_AUTH_TOKEN="your-evolink-api-key"
export ANTHROPIC_MODEL="deepseek-v4-pro"
claude

That is the entire switch: your agent workflow, tools, and prompts stay the same. Community reports consistently describe V4 Pro as strongest on long, multi-step coding tasks — the 0813 build roughly doubled its scores on terminal-agent benchmarks — so an agentic harness like Claude Code is exactly where it earns its price difference against closed models.

Two practical notes for this setup:

• Tool calling works through the standard Anthropic tool_use / tool_result flow, so Claude Code's file edits and shell tools function normally.
• V4 Pro has no vision input. Claude Code features that attach screenshots or images will not work on this route; keep a vision-capable model configured for those tasks.

Step 3 — The three migration pitfalls

These are the mappings that silently differ from Anthropic's native API. All three come from the current EvoLink contract, verified August 13, 2026.

{
  "thinking": {"type": "enabled"},
  "output_config": {"effort": "high"}
}

{
  "model": "deepseek-v4-pro",
  "system": "You are a terse senior reviewer.",
  "messages": [{"role": "user", "content": "Review this diff..."}]
}

Step 4 — Thinking effort and what it does to your bill

DeepSeek bills thinking tokens as output tokens, and V4 Pro is a heavy thinker: community measurements have shown it consuming several times more reasoning tokens than closed-model peers on the same task. Practical guidance:

• The default is effort: "high" — a heavy setting for routine work. Set low explicitly for bulk steps and keep high for tasks where a failed attempt costs more than the extra tokens; max is the escalation tier.
• Cache-hit input is billed at roughly 1/120 of the cache-miss rate until August 16, 2026 16:00 UTC; DeepSeek's published repricing then moves the Pro ratio to about 1/30 with peak/off-peak dual rates. Check live per-token rates on the DeepSeek model pricing on EvoLink rather than trusting any blog's numbers, including this one's.
• For high-volume, low-difficulty steps (classification, summaries), route to deepseek-v4-flash instead and keep Pro for the hard steps.

Step 5 — Concurrency, 429s, and fallback

1. Treat 429 as backpressure: exponential backoff with jitter, and cap in-flight requests below your measured ceiling.
2. Set client timeouts generously for high effort tasks — thinking time precedes the first token.
3. Configure a fallback: because the EvoLink route speaks the same Messages format for multiple models, a router-level fallback from deepseek-v4-pro to another available model is a config change, not a rewrite. Community threads are full of exactly this pattern — Flash for bulk steps, Pro for hard steps, a closed model as final fallback.

FAQ

Where to go next

• DeepSeek models on EvoLink — live pricing and model access.
• DeepSeek V4 Pro 0813: what changed — the GA build's agent and Codex changes.
• Pro vs Flash decision guide — which tier fits which workload.