DeepSeek-V4-Flash tops Haiku in evals

This is less about a flashy benchmark and more about a practical shift in the “cheap assistant” tier: if Flash really does reliable tool calling better than Haiku, it’s immediately interesting for production chat systems.

• –DeepSeek’s own docs list `deepseek-v4-flash` as the API model name, so this is not just a community alias but a shipping preview target.
• –The key signal tool-use behavior, not raw reasoning; for agentic chat systems, that matters more than abstract benchmark bragging.
• –If Flash is genuinely cheaper than Haiku while matching or beating it on structured tool calls, teams will have a strong incentive to re-run cost/perf evals.
• –The post is still anecdotal, so the right takeaway is “promising candidate,” not “new default” until independent evals confirm the result.