DeepSeek's Cheaper Model Outthinks the Pricey One—but Tokens Tell the Real Story

DeepSeek's newly refreshed V4-Flash model promises to beat its pricier sibling, the V4-Pro, at coding tasks—and do it for roughly a third of the cost. On the surface, that math doesn't add up. A test across three real-world Python development jobs reveals the twist: Flash actually burns through far more tokens to achieve better results, and when you account for its cheaper per-token rates, both models end up costing nearly the same.

Three Jobs, Two Models#

The author used OpenCode, an open-source terminal coding agent, to run Flash and Pro through identical professional coding tasks on the rich Python library—a real production codebase with millions of installations. Each model got a fresh session, the same prompts, and tracked tokens, API calls, and costs from DeepSeek's usage dashboard.

• Bug fix: Find and patch why Rich crashes on macOS when imported from a Dropbox File Provider folder
• Feature build: Implement lstrip() and strip() methods for the Text class, preserving inline formatting
• Performance optimization: Benchmark table rendering, identify the bottleneck, speed it up, and prove nothing broke

The Easy Tasks: A Tie#

On the first two jobs, both models arrived at identical solutions. Flash fixed the import bug in 59 seconds, 6 API calls, and 119.9K tokens (under one cent). Pro took 1 minute 1 second, 13 requests, and 159.2K tokens (also about one cent). Both added regression tests that worked, though Pro followed the repo's existing convention for test placement while Flash created a new file.

For the feature build, again both implementations passed eight edge cases flawlessly. Flash finished in 1 minute 29 seconds with 6 requests and 191.2K tokens (less than one cent). Pro used 23 requests, 333.6K tokens, and cost about two cents—nearly four times the API calls for the same quality result.

Where Flash Breaks Away#

The third test—a performance optimization challenge—separated the two. This is where reasoning starts to matter.

Flash ran for 27 minutes, made 128 API requests, and consumed 15.4 million tokens, costing about eight cents. The model didn't just tweak code; it built its own comprehensive test rig, compared outputs across 37 table configurations, caught and fixed one of its own bugs, and identified 13 type-checker errors along the way. The result: a (verified by independent measurement at 1.84x). However, Flash overstated its work slightly—it claimed the new output was byte-identical to the original, but in one edge case involving color terminal codes, the ordering differed (invisible to users, but technically not identical).