DeepSeek released the V4 Pro general-availability (GA) model. Early measurements show it cuts reasoning-token usage by 18 % to 62 % compared with the preview build. That matters for anyone paying per token: the same prompts now cost noticeably less while still producing comparable output.

What prompted the comparison

The GA release arrived without a blog post or changelog, so developers had to discover the differences themselves. A community test ran identical tasks on both versions and found the biggest change—a steep drop in the tokens the model spends “thinking” before answering. On trivial look-ups the GA model used 62 % fewer reasoning tokens; on more complex queries the reduction was 18 %.

Token efficiency and its impact

In a typical extraction workflow the preview build burned 159 reasoning tokens to pull a handful of fields from a prompt. Switching to the GA build with the “thinking” feature disabled collapsed that number to 40 tokens. For users on metered plans the savings turn directly into lower bills, especially at scale.

JSON extraction: the hidden snag

Both builds stumble when the “thinking” mode is on: they pass a JSON schema check but insert wrong numeric values. Only the GA version returns correct JSON when “thinking” is turned off. Teams that need structured output should disable the thinking flag for extraction tasks, or they will receive syntactically valid but numerically incorrect data.

Refusal handling flips the script

The preview model could refuse a question it deemed unanswerable, replying “I do not know.” The GA model no longer does this. Instead it either runs out of its token budget without answering or fabricates a response. This change improves token efficiency but removes a safety net that kept the model from hallucinating on unknown topics.

Reliability boost

A dangerous loop in the preview build—triggered by a modest “thinking” budget—could make the model repeat the same text until the 8,192-token window filled up. The GA release patches this bug, ending the runaway repetition that previously risked exhausting the request window and inflating costs.

What users should watch

  • Enable the GA build for lower token counts. Measured reductions hold across task complexity.
  • Turn “thinking” off for any JSON or structured-data extraction. This yields correct values and keeps token use minimal.
  • Don’t count on built-in refusals. If a prompt asks for unverifiable data, the GA model may still produce an answer, so downstream validation remains essential.
  • Watch peak-hour billing. New pricing rules make token consumption during high-traffic periods affect overall spend more sharply than before.

Bottom line

DeepSeek’s V4 Pro GA model delivers a clear efficiency win—up to 62 % fewer reasoning tokens—and fixes a critical repetition bug. However, the loss of explicit refusal responses and the need to disable thinking for accurate JSON output add new considerations. Teams that adapt their pipelines can lower costs without sacrificing functionality.

Source: https://dev.to/synthorai/deepseek-v4-pro-ga-vs-preview-measured-18-62-less-thinking-539l