DeepSeek calls it a minor update: V4.1-Flash-1017 beats Opus 5 on DeepSWE at a fifth of a dollar per task, MIT weights, 1M context
DeepSeek shipped on a Saturday again. V4.1-Flash-1017 went up on Hugging Face at 02:40 UTC on October 17 under MIT, for weights and code, with a 1M-token context and the same Flash price as before: $0.15 in and $0.60 out per million tokens off-peak, $0.30 and $1.20 at peak. The model card calls it a minor update to V4.1-Flash that delivers substantial improvements, and the HN thread spent a good part of its 534 comments laughing at that sentence. The launch post from DeepSeek's account passed 3.5 million impressions and 2,100 quote posts in a day. The two HN threads took 611 and 511 points.
The table is the story, and this time DeepSeek named the harness on each row. Terminal-Bench 4.0 went from 31.2 to 42.3, run inside the Claude Code harness, against 39.9 for GPT-5.6 Sol, 37.9 for GLM-5.3 and 51.8 for Opus 5. DeepSWE v1.1 went from 74.2 to 77.0 in the Mini-SWE-V2 harness, above Opus 5 at 74.0, Sol at 73.0 and Kimi K3 at 67.5. Automation-Bench is 60.5 against 50.3 for Opus 5. Agents' Last Exam is 35.4 against 28.6. Apex Bench jumped from 64.6 to 72.1. ProgramBench went from 20.3 to 29.1 but Opus 5 still leads at 37.0, and the same is true of NL2Repo at 68.2 against 75.3 and HLE at 42.6 against 56.3. So the pattern is clear: agent loops and tool use are where it wins, raw knowledge is where it still loses. DeepSeek's own cost chart puts the top setting at 77 percent on DeepSWE for just under 20 cents per task, the low setting above 72 percent at about a dime, and the frontier closed models at two to ten dollars per task on the same benchmark. Cline put it as beating Sol on Terminal-Bench at 17 times lower price. All numbers are DeepSeek's; nobody independent had run it by Saturday night.
How they got there matters more than the numbers. The release note says the model uses reinforcement learning with self-feedback for end-to-end project work, with the demos being a Forbidden City model in 3D, a browser racing game and a Portal 2 clone in Godot. The DeepSeek Harness account said the model was specifically optimized for DeepSeek Harness during training, and shipped Harness 0.2.1 the same hour; the card also lists Tencent CodeBuddy and WorkBuddy as targets. Cline's read, quoting the team, is that the 35 percent Terminal-Bench jump came with no algorithmic innovation, just RL across multiple coding agent harnesses. That is the harness-not-the-model thesis stated by a lab in its own release: the training target is the loop, and the model card now reports which loop each score came from. Also in the note: the Search and Fetch APIs behind DeepSeek's in-house search engine are open to everyone at $0.75 per thousand calls, and are free inside DeepSeek Harness, Claude Code and Codex until October 31.
The reactions sorted into two camps. Teortaxes said this puts DeepSeek in above-Sol territory faster than expected and is a foundation for a V4.1 Pro. Nathan Lambert asked if the king of open weights is back. Susan Zhang noted it took one month to move the cost-performance frontier on this benchmark and five months to go from V4 to V4.1. The top HN comment was more useful than any of them: DeepSeek is not better than Sol or Gemini on the benchmarks, but it feels better in practice because the tokens per second and time to first token from providers are high, you can see the chain of thought, and you never get silently blocked for a cyber review. Another commenter ran four agents over the same dirty C++ codebase and found GPT-6.1 Sol surfaced two P1 issues in four and a half minutes while DeepSeek took twenty minutes for one, which is the kind of apples-to-bananas test you actually want.
The same day, PromptArmor published a 9,800-trial indirect prompt injection study that reads as the other side of the open-weights argument. Agents built on open-weight models (DeepSeek, Kimi, GLM, Qwen) fell for end-to-end attacks 26.7 percent of the time against 14.2 percent for GPT, Claude and Gemini, which PromptArmor rounds to 88 percent more susceptible. The three sample attacks were lopsided: summarize-a-URL-then-run-malware 49 against 9 percent, phishing login overlay 36 against 5, email theft via a markdown image 23 against 9. The twist is in the cheap tiers. Claude Haiku 5.5 was 38 points worse than Fable 5.1 and GPT-6 Luna 21 points worse than Astra, while DeepSeek V4.1-Flash was 31 points better than V4-Pro. PromptArmor sells agent security and gave no publication date, so treat it as a vendor benchmark. But if you are about to swap a $25-per-million model for a 60-cent one inside an agent that reads email, the injection rate is the number to check before the DeepSWE score.
Links: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017 , https://api-docs.deepseek.com/news/news261017 , https://x.com/deepseek_ai/status/2111375640717033777 , https://x.com/DeepSeekHarness/status/2111376365497077958 , https://news.ycombinator.com/item?id=50129045 , https://www.promptarmor.com/resources/open-weight-susceptibility
← Back to all articles
The table is the story, and this time DeepSeek named the harness on each row. Terminal-Bench 4.0 went from 31.2 to 42.3, run inside the Claude Code harness, against 39.9 for GPT-5.6 Sol, 37.9 for GLM-5.3 and 51.8 for Opus 5. DeepSWE v1.1 went from 74.2 to 77.0 in the Mini-SWE-V2 harness, above Opus 5 at 74.0, Sol at 73.0 and Kimi K3 at 67.5. Automation-Bench is 60.5 against 50.3 for Opus 5. Agents' Last Exam is 35.4 against 28.6. Apex Bench jumped from 64.6 to 72.1. ProgramBench went from 20.3 to 29.1 but Opus 5 still leads at 37.0, and the same is true of NL2Repo at 68.2 against 75.3 and HLE at 42.6 against 56.3. So the pattern is clear: agent loops and tool use are where it wins, raw knowledge is where it still loses. DeepSeek's own cost chart puts the top setting at 77 percent on DeepSWE for just under 20 cents per task, the low setting above 72 percent at about a dime, and the frontier closed models at two to ten dollars per task on the same benchmark. Cline put it as beating Sol on Terminal-Bench at 17 times lower price. All numbers are DeepSeek's; nobody independent had run it by Saturday night.
How they got there matters more than the numbers. The release note says the model uses reinforcement learning with self-feedback for end-to-end project work, with the demos being a Forbidden City model in 3D, a browser racing game and a Portal 2 clone in Godot. The DeepSeek Harness account said the model was specifically optimized for DeepSeek Harness during training, and shipped Harness 0.2.1 the same hour; the card also lists Tencent CodeBuddy and WorkBuddy as targets. Cline's read, quoting the team, is that the 35 percent Terminal-Bench jump came with no algorithmic innovation, just RL across multiple coding agent harnesses. That is the harness-not-the-model thesis stated by a lab in its own release: the training target is the loop, and the model card now reports which loop each score came from. Also in the note: the Search and Fetch APIs behind DeepSeek's in-house search engine are open to everyone at $0.75 per thousand calls, and are free inside DeepSeek Harness, Claude Code and Codex until October 31.
The reactions sorted into two camps. Teortaxes said this puts DeepSeek in above-Sol territory faster than expected and is a foundation for a V4.1 Pro. Nathan Lambert asked if the king of open weights is back. Susan Zhang noted it took one month to move the cost-performance frontier on this benchmark and five months to go from V4 to V4.1. The top HN comment was more useful than any of them: DeepSeek is not better than Sol or Gemini on the benchmarks, but it feels better in practice because the tokens per second and time to first token from providers are high, you can see the chain of thought, and you never get silently blocked for a cyber review. Another commenter ran four agents over the same dirty C++ codebase and found GPT-6.1 Sol surfaced two P1 issues in four and a half minutes while DeepSeek took twenty minutes for one, which is the kind of apples-to-bananas test you actually want.
The same day, PromptArmor published a 9,800-trial indirect prompt injection study that reads as the other side of the open-weights argument. Agents built on open-weight models (DeepSeek, Kimi, GLM, Qwen) fell for end-to-end attacks 26.7 percent of the time against 14.2 percent for GPT, Claude and Gemini, which PromptArmor rounds to 88 percent more susceptible. The three sample attacks were lopsided: summarize-a-URL-then-run-malware 49 against 9 percent, phishing login overlay 36 against 5, email theft via a markdown image 23 against 9. The twist is in the cheap tiers. Claude Haiku 5.5 was 38 points worse than Fable 5.1 and GPT-6 Luna 21 points worse than Astra, while DeepSeek V4.1-Flash was 31 points better than V4-Pro. PromptArmor sells agent security and gave no publication date, so treat it as a vendor benchmark. But if you are about to swap a $25-per-million model for a 60-cent one inside an agent that reads email, the injection rate is the number to check before the DeepSWE score.
Links: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017 , https://api-docs.deepseek.com/news/news261017 , https://x.com/deepseek_ai/status/2111375640717033777 , https://x.com/DeepSeekHarness/status/2111376365497077958 , https://news.ycombinator.com/item?id=50129045 , https://www.promptarmor.com/resources/open-weight-susceptibility
Comments