DeepSeek has done it again. The Chinese lab’s new DeepSeek-V4.1-Flash-1017 is a “minor update” by name, but its own benchmarks show it beating much bigger models like Claude Opus 5 and GPT-5.6 Sol on several coding and agent tests, at a fraction of the cost per task. The weights are open under the MIT license, and DeepSeek is opening its in-house search engine to developers too.
What shipped
DeepSeek announced the model on X on October 17:
- Where: Live on DeepSeek’s website, app and API, with open weights on Hugging Face and ModelScope under the MIT license.
- API: According to DeepSeek’s API docs, the “deepseek-flash” model name now points to V4.1-Flash-1017, and the older V4 Flash names are routed to the new model at the Flash price.
- Context: 1 million tokens, with DeepSeek recommending a maximum output of at least 256K tokens.
- The pitch: stronger benchmark scores, faster task completion and shorter outputs. DeepSeek calls it “a Pareto improvement in performance and cost.”
The numbers
Here’s how DeepSeek’s model card compares the new model with the previous V4.1-Flash and the bigger rivals (all scores from DeepSeek’s own testing):
- DeepSWE v1.1 (software engineering): 77.0, up from 74.2. Claude Opus 5 scores 74.0 and GPT-5.6 Sol 73.0.
- Terminal-Bench 4.0: 42.3, up from 31.2. GPT-5.6 Sol scores 39.9, while Opus 5 still leads at 51.8.
- ProgramBench: 29.1, up from 20.3 (Opus 5: 37.0).
- Automation-Bench v1.0.6: 60.5, ahead of Opus 5 at 50.3.
- Humanity’s Last Exam (no tools): 42.6, up from 36.8, but well behind Opus 5’s 56.3.
So it’s not a clean sweep. On harder reasoning and some coding tests, Anthropic’s flagship is still ahead. The real story is cost. The Decoder points to DeepSeek’s DeepSWE cost chart (below): the new model hits 77.0% at just under $0.20 per task, and its low setting still clears 72% for about $0.10. The previous version scored 74.2% at around $0.50. Rival models in the chart, including GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5, cost between $2 and more than $10 per task without matching that accuracy.
The Decoder also says DeepSeek trained the update with reinforcement learning from self-feedback (RLSF), where the model uses its own confidence scores or self-critiques as the reward signal instead of human raters or a separate reward model. DeepSeek says that helps with building whole projects from scratch, from games to 3D modeling. One important caveat: independent benchmarks haven’t come out yet, and real-world code quality can differ a lot from leaderboard numbers.
Search for everyone
The other launch might matter just as much for developers. DeepSeek opened its Search and Fetch APIs, which run on an in-house search engine it says handles hundreds of millions of requests a day. Each endpoint costs $0.75 per 1,000 calls, and calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users until billing starts on October 31 at 16:00 UTC (November 1 at 1:00 a.m. JST).
The fine print
It hasn’t been a perfect week. Nikkei Asia reports DeepSeek suffered a roughly five-hour outage on Thursday that took its V4-Pro and V4.1-Flash models and chat service offline from 8 a.m. to 1 p.m. local time. It gave subscribers partial refunds without naming a cause, and Nikkei Asia suggests a shortage of computing power under U.S. chip export curbs could be a factor. The Decoder also notes reports that Beijing may tighten rules on publishing open-weight models, possibly depending on whether a model ranks among the top 10 most capable. That would hit DeepSeek’s open-source strategy hard.
The EVX News take
Every DeepSeek release raises the same question for the U.S. labs: how do you justify $5 to $10 per task when a free-to-download model gets close for 20 cents? That pressure is good for anyone building agents, including Grok, which competes on speed and price as much as raw intelligence. The cheaper capable AI gets, the faster it spreads into robots, cars and everything else we cover. Whether the benchmark gains hold up in independent testing is the next thing to watch.
Sources: @deepseek_ai on X, DeepSeek model card on Hugging Face, DeepSeek API docs, The Decoder, Nikkei Asia.
— The EVX News team
Featured image: DeepSeek logo. Image: DeepSeek (from DeepSeek’s GitHub), via Wikimedia Commons, MIT License. Chart: DeepSWE v1.1 accuracy vs. cost per task. Image: DeepSeek, via DeepSeek on X.
