DeepSeek's new free model tops Opus 5 on one coding test and loses another by 14 points
DeepSeek shipped V4.1-Flash-1017 on October 17 under an MIT license. By its own table it beats Anthropic's Opus 5 on DeepSWE. By the same table it trails by more than 13 points on Humanity's Last Exam.
DeepSeek’s announcement page calls its new model “a Pareto improvement in performance and cost.” That’s a fancy way of saying it’s better and cheaper at the same time, and for coding work the company’s own numbers mostly back it up.
DeepSeek-V4.1-Flash-1017 went live late on Saturday, October 17, on its website, its app and its API. The weights are on Hugging Face under the MIT license, which has no revenue threshold or usage catch. The model card says it’s a “minor update” to V4.1-Flash, then in the next breath claims “substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks.” Both can’t be fully true, so I went to the table.
The scorecard, with the caveat up front
Every number below comes from DeepSeek’s model card. Nobody independent has reproduced them yet. DeepSeek ran Terminal-Bench 4.0 through Claude Code and DeepSWE v1.1 through the Mini-SWE-V2 test setup, and says so.
| Benchmark | New Flash | Old Flash | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|
| DeepSWE v1.1 | 77.0 | 74.2 | 73.0 | 74.0 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 39.9 | 51.8 |
| ProgramBench | 29.1 | 20.3 | 23.0 | 37.0 |
| HLE, no tools | 42.6 | 36.8 | 44.5 | 56.3 |
The first row is the headline. On DeepSWE, an open model with MIT-licensed weights sits three points above Opus 5. The jump on Terminal-Bench is bigger, 31.2 to 42.3, and it now edges past GPT-5.6 Sol there too.
The last row is the part the announcement skips. On Humanity’s Last Exam without tools, Opus 5 is at 56.3 and DeepSeek is at 42.6. That’s a 13.7-point hole, and it barely moved from the old Flash’s 36.8 compared with the coding gains. Even on Terminal-Bench, the benchmark it improved most, it’s still about 9 points behind Opus 5. To DeepSeek’s credit, the table includes those rows instead of quietly dropping them.
So the accurate summary is narrow. This is a strong, cheap coding and agent model. It isn’t a general-purpose Opus replacement, and DeepSeek’s own chart says that.
The search API has a date on it
The release also opens up DeepSeek’s in-house search engine as two API endpoints, Search and Fetch, at $0.75 per 1,000 calls each. DeepSeek says the engine already serves hundreds of millions of requests a day.
The part that matters for anyone building agents: search and fetch calls made through server-side tools in DeepSeek’s own agent app, Claude Code and Codex are free for DeepSeek API users for now. Billing starts on October 31, 2026, at 16:00 UTC. If you wire it into a coding agent this week and forget about it, you’ll get a bill in two weeks.
Why the price part is a claim, not a fact
DeepSeek says the new model finishes tasks faster and writes shorter outputs than the old one, which is where the “cost” half of the Pareto claim comes from. Its announcement includes a chart of DeepSWE score against cost per task. I couldn’t find per-token prices in the announcement itself, and the chart is DeepSeek’s own. Bloomberg’s report repeats the company’s description of the update and doesn’t add independent testing.
That’s not unusual for a Saturday release. It’s also why I wouldn’t switch a production pipeline because of a table on a model card.
What I think
I’d treat this as a good reason to run your own tests, not a result. A 3-point lead on one benchmark, measured by the vendor, in a test setup the vendor picked, is the kind of gap that can vanish when someone else runs it. The Terminal-Bench jump is more interesting, because it’s large enough to survive some noise.
What I do find notable is the pattern. Meta’s Muse Spark Open arrived this week with a license that has a $20 million catch, and Reflection’s Beam is the American open-weight bet. The one with the plainest license is the Chinese one. If you’re a US startup choosing weights you can legally ship in a product, that’s a real consideration, whatever you think about where the model came from.
Independent numbers on DeepSWE and Terminal-Bench should show up within days. That’s what I’m waiting for.
Sources
- DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real (DeepSeek API Docs, October 17, 2026)
- DeepSeek-V4.1-Flash-1017 model card (Hugging Face, updated October 17, 2026)
- DeepSeek Pushes Out Updated AI Model It Says Is Faster, Leaner (Bloomberg, via Bloomberg Law, October 17, 2026)