DeepSeek V4 Pro Continues After September 14: Retained Billing and API Verification

Key Takeaways from Current Documentation as of September 17, 2026
In the current Quick Start and pricing sections, DeepSeek confirms that the API service for DeepSeek V4 Pro continues to operate after September 14, 2026, with the previous billing method retained. While the dated September 10 announcement referenced an initial plan to route deepseek-v4-pro calls to V4.1 Flash at Flash rates, the active reference sections of the platform remove the need for any forced migration.

When maintaining integrations, rely on the active service and pricing documentation. We base this guidance on published official specifications and make no claim of independent verification regarding server-side model weights or runtime billing audits. In this situation, there is no need to hastily migrate stable working configurations from V4 Pro—it is sufficient to verify your client configuration, the gateway in use, and regular billing deductions.

First, Identify Which API Receives Your Requests

Different integration layers react differently to platform changes. Before modifying your code or configuration, determine exactly where your requests are routed:

Connection	What to Check
Direct DeepSeek API	Account status, current documentation sections, and actual billing deductions in your console
Third-party API provider (including BetterToken)	The provider's own model catalog, active Model IDs, billing terms, and stated routing rules; official DeepSeek rules do not automatically apply to third-party gateways
Coding agent (Cursor, Cline, Aider, etc.)	Actual Base URL and Model ID in profile parameters, checked separately for interactive context and background tasks

Because model parameters can be configured across multiple layers, inspect the actual Base URL and Model ID dispatched by your client.

To quickly locate references to deepseek-v4-pro across your local codebase, run a search using ripgrep:

rg -l --hidden -g '!.git' -g '!node_modules' -g '!.venv' \
  'deepseek-v4-pro' .

Run this command from the project root. It makes no file changes and requires no network access. A lack of matches does not guarantee the absence of dependencies: the model name may have been defined in CI/CD variables, cloud gateway configurations, or local client aliases. Do not include API keys in verification checklists.

Model IDs and Routing Rules

Understanding how the service resolves model identifiers prevents false conclusions during verification.

As of September 17, 2026, in DeepSeek's official pricing list, the identifier deepseek-v4-pro is assigned to the DeepSeek-V4-Pro-0813 version. At the same time, the legacy aliases deepseek-v4-flash and deepseek-v4-flash-vision-exp have been transitioned to DeepSeek-V4.1-Flash (current Model ID: deepseek-flash) and are billed at Flash rates. These are two independent rules: the migration of legacy Flash aliases does not imply an automatic status change for V4 Pro.

Compatible interfaces require particular attention. According to the DeepSeek Anthropic API guide, model mapping rules on the official endpoint operate as follows:

• Model names starting with claude-opus map to deepseek-v4-pro and are billed at V4 Pro rates;
• Model names starting with claude-haiku or claude-sonnet map to deepseek-flash;
• Any unsupported model name is automatically routed by DeepSeek's official backend to deepseek-flash.

These routing rules apply strictly to DeepSeek's official gateway (https://api.deepseek.com/anthropic) and are not transferred automatically to third-party platforms.

Furthermore, the model field returned in an API response or self-identification text generated by the model does not serve as independent proof of which server-side weights were actually executed.

Verifying Workload Stability Under Conflicting Documentation

When documentation contains contradictions, scenario-specific testing complements official guidance and validates client behavior, though it cannot prove which server-side weights are executed:

Scenario	What to Monitor
Code editing	The test suite fails prior to changes and passes after; syntax and adjacent modules remain intact
Tool calling	The tool name and arguments match the expected schema; the client cleanly consumes the returned payload and returns control to the conversation
Streaming	The client cleanly initiates and terminates streams, does not crash on chunks, and accurately captures final usage metadata
Long context	The agent consistently retains system instructions and critical facts from earlier stages of the conversation
Rate limits and retries	Retry logic caps total attempts and avoids duplicating external operations (database writes, webhooks)

If you invoke reasoning models through the Anthropic-compatible protocol, verify your reasoning parameters. In DeepSeek's adapter, thinking.budget_tokens is ignored, while effort inside output_config is respected. Full mapping details are available in the Thinking Mode guide.

Expense Tracking and Billing Control

Because current official documentation keeps separate billing in effect for V4 Pro, calls to deepseek-v4-pro (including requests with a claude-opus prefix via the Anthropic endpoint) cannot be calculated at reduced Flash rates. Financial parameters of your integration must be tracked separately from operational health.

For test verification requests, record:

1. The exact request timestamp and endpoint used;
2. The submitted Model ID;
3. Input, output, and cached token totals;
4. The final deduction from your balance.

Compare recorded charges against the current DeepSeek pricing table. If you connect through BetterToken, token consumption metrics can be viewed in the Workspace and active rates in the BetterToken catalog. BetterToken data reflects parameters specific to your integration with that provider, but does not serve as proof of internal routing within the direct DeepSeek API.

Factor retry frequency into cost projections. Even if unit token rates are lower, total spend per completed task can increase if an agent requires more iteration cycles.

Fallback Strategy: Why Implicit Aliases Cannot Be Trusted

A resilient fallback plan must not rely on undocumented aliases or assumptions of automated routing.

If your fallback profile simply duplicates deepseek-v4-pro, any gateway policy change will affect both primary and secondary configurations identically. For genuine redundancy, configure an explicitly documented Model ID or an alternative vetted provider.

Avoid preemptively changing working production identifiers until your fallback candidate passes basic validation for tool calling, stream handling, and billing predictability. If an alternative model fails to meet operational criteria, temporarily throttling automated execution of critical workloads is safer than migrating to an unverified setup.

Next Steps

Inspect the actual Model ID and Base URL in your client configuration or outgoing request logs. If your current integration with the official DeepSeek API is configured for deepseek-v4-pro and consistently handles your workloads, there is no need to forcefully switch it: the service continues to operate under its existing billing terms.

If you connect through BetterToken, verify parameters in the active model catalog to confirm the availability of your desired identifier and current billing rules. To verify billing accuracy, run a small test request and compare token consumption metrics against balance deductions in the Workspace.

---

Originally published on the BetterToken blog.

BetterToken provides pay-as-you-go access to AI model APIs through
OpenAI-compatible and Anthropic-compatible endpoints — useful if you are wiring
Claude Code, Codex, or your own tooling to a custom base URL.
See the docs to get started.