← sources

Has anyone actually replaced Claude with DeepSeek V4.1 Flash/Pro for tool-heavy daily work?

source: reddit (reddit:ClaudeAI)published 2026-09-20 16:05ingested 2026-09-21 02:36

retained text

I am considering dropping Claude because the token burn and effective cost are hard to justify for my actual use: mostly administrative documents, light web research, and occasional tool-assisted work. I care about intelligence, reliable tool use, and cost per completed task—not benchmark theater.



I did some homework first:



\- DeepSeek's V4.1 Flash model card reports 90.6 vs 89.1 for Opus 5 on Terminal-Bench 2.1, and 74.2 vs 74.0 on DeepSWE v1.1.

\- Opus 5 still leads clearly on harder/newer tasks: Terminal-Bench 4.0 is 51.8 vs 31.2, NL2Repo is 75.3 vs 64.0, and HLE without tools is 56.3 vs 36.8.

\- Current API pricing per million tokens is dramatically different. DeepSeek Flash is $0.15–$0.30 uncached input and $0.60–$1.20 output; V4 Pro is $0.66–$1.32 input and $1.98–$3.96 output. Claude Sonnet 5 is $2/$10 and Opus 5 is $5/$25.

\- Both DeepSeek models support a 1M context, tool calls, Responses API and Anthropic API formats. DeepSeek documents direct Claude Code integration.

\- Important caveat: DeepSeek's Responses compatibility ignores built-in web search, file search, code interpreter, computer use and MCP; only function tools and apply\_patch are supported there. The harness therefore matters a lot.



Sources:

[https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash)

[https://api-docs.deepseek.com/quick\_start/pricing/](https://api-docs.deepseek.com/quick_start/pricing/)

[https://api-docs.deepseek.com/guides/responses\_api/](https://api-docs.deepseek.com/guides/responses_api/)

[https://platform.claude.com/docs/en/about-claude/pricing](https://platform.claude.com/docs/en/about-claude/pricing)

[https://swescience.github.io/](https://swescience.github.io/)



I cannot justify paying for overlapping AI subscriptions just to run a long trial. For people who have genuinely moved real daily work from Claude to DeepSeek:



1. Which model and harness do you use—Flash or Pro, Claude Code, Deep Code, OpenCode, Codex, something else?

2. For documents and web research, what actually became worse?

3. How often do tool calls fail, loop, lose context, or require manual recovery?

4. What is your measured end-to-end cost per accepted task, including retries—not just price per token?

5. Do you still keep Claude as a fallback? If so, for which tasks?



I am looking for a more economically sustainable workflow, not merely the cheapest model on paper.

document fetched 2026-09-21 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1wlkzoa/has_anyone_actually_replaced_claude_with_deepseek