Tested DeepSeek V4 Pro and Flash Against Claude Opus 4.7 and Kimi K2.6

DeepSeek V4 Pro and DeepSeek V4 Flash were released under the MIT license on April 24, 2026. As DeepSeek's first new architecture and a 2-tier open-weight lineup of Pro/Flash, they were directly compared against Claude Opus 4.7 and Kimi K2.6 using the same FlowGraph spec, the same prompts, and the same scoring criteria.

The evaluation target was a workflow orchestration backend with 20 endpoints, persistent state, lease management, retries, and event streaming. DeepSeek V4 Pro passed its own tests but scored only 77/100, with an execution cost of $2.25. Its performance fell between Opus 4.7 (91) and Kimi K2.6 (68), but core flaws remained.

• The original worker could still record step completion even after its lease had expired.
• When one run became saturated, it couldn't see the next candidate, blocking work from other runs as well.
• npm run build failed, and npm start also didn't actually work.

DeepSeek V4 Flash scored 60/100 with a total execution cost of $0.02, and its output token pricing was 1/14 that of Kimi K2.6 and 1/89 that of Claude Opus 4.7. However, the setup script failed first while forcibly resetting the DB, so even the first test couldn't run, and the actual HTTP API was mounted at /runs/key/:key/runs instead of /workflows/key/:key/runs, producing a 404. The tests called internal functions directly, missing this path error.

• Even after a failed run, some steps returned to waiting_retry instead of blocked.
• The core entry point for clients to start a run was itself missing.

DeepSeek is also running a 75% discount on V4 Pro through May 31, 2026. With the discount applied, input pricing drops to about $0.036/M and output to $0.87/M, and input cache pricing has been permanently cut to one-tenth across the entire lineup.