DeepSeek V4.1-Flash-1017 vs. Claude Haiku 5.5: More accurate at a fraction of the price

DeepSeek rolled out V4.1-Flash-1017 on October 17 as an update to its Flash model, and the changelog says it trims both runtime and output length. The model card lists 11 benchmark scores, including 42.3 on Terminal-Bench 4.0. It’s open-weight, so developers can hit DeepSeek through its API, use a service like OpenRouter, or download it from Hugging Face and run it themselves.

Anthropic answered with Claude Haiku 5.5 on October 7, positioning it as its fastest model at standard speed for high-volume work like classification, summarization, and database queries. Anthropic says its Opus models can run faster in Fast Mode. On Terminal-Bench 4.0, Anthropic reports 39.2, which trails DeepSeek’s reported score by 3.1 points. None of that is a direct lab comparison, though. These are vendor numbers.

So The New Stack built three tests around the tasks each company says its model handles well: an agentic bug fix, a database-query task, and a triage job with 1,600 scored fields. Both models were run once per test through OpenRouter, pinned to DeepSeek’s and Anthropic’s own servers, with default settings. DeepSeek was locked to model ID 1017 and ran off-peak. The prompts and answers were graded with scripts and a second check before either model saw the keys.

The clearest win for DeepSeek came on the triage job, where it got all 1,600 fields right and Haiku missed four. Two of Haiku’s misses were plain severity mistakes: it rated a 49-customer report and a 0.8% error-rate report as sev3 when both should’ve been sev4. The other two were routing errors. DeepSeek also edged out Haiku on the bug-fix test, while the database-query task ended in a tie at 20 out of 20.

The cost gap was the real eye-opener. Across the three runs, DeepSeek came to $0.07 and Haiku to $0.30. Haiku was faster on the two tool-based tests, but DeepSeek was faster on the big triage prompt and cheaper in the end, especially because Haiku crossed its higher long-prompt rate on that job. With this setup, DeepSeek looked like the better buy.