{"slug": "deepseek-quietly-ships-v4-1-flash-1017-beating-rivals-on-agent-tasks", "title": "DeepSeek Quietly Ships V4.1-Flash-1017, Beating Rivals on Agent Tasks", "summary": "DeepSeek released V4.1-Flash-1017 on October 17, an update to the V4.1-Flash model it launched five weeks earlier, posting new weights on Hugging Face and ModelScope under an MIT license. According to MarkTechPost, the update raised the model's DeepSWE software engineering benchmark score from 74.2 to 77.0, and DeepSeek says it beats the original V4.1-Flash on all 11 benchmarks it reported by margins of 2.0 to 11.1 points. The model keeps a 552-billion-parameter backbone with 8 billion parameters active during prompt reading and 16 billion during generation, supports context windows up to 1 million tokens, and is priced at $0.30 per million input tokens and $1.20 per million output tokens at peak, halving to $0.15 and $0.60 off-peak.", "body_md": "*DeepSeek didn't hold a launch event or send a press release. It just dropped a smarter, cheaper model on Hugging Face two days ago, and the gap with pricier Western rivals keeps shrinking.*\n\nOn October 17, DeepSeek released V4.1-Flash-1017, an update to the V4.1-Flash model it launched barely five weeks earlier. No keynote, no countdown clock, just new weights on Hugging Face and ModelScope under an MIT license, the kind of release you'd miss entirely if you weren't already watching. According to MarkTechPost, the update pushed the model's score on DeepSWE, a software engineering benchmark that tests whether a model can actually resolve real coding tasks, from 74.2 to 77.0. DeepSeek says the new version beats the original V4.1-Flash on all 11 benchmarks it reported, by margins ranging from 2.0 to 11.1 points.\n\nThat's the whole story in one sentence: a model most people had never heard of a year ago just got meaningfully better at the one thing founders building AI agents care about most, finishing real tasks, and it did it quietly, the way DeepSeek has done nearly everything since its DeepSeek-R1 moment rattled markets in January 2025.\n\nV4.1-Flash-1017 keeps the same 552-billion-parameter backbone as its predecessor, built on what DeepSeek calls a causal encoder-decoder design. Only 8 billion parameters activate while the model reads a prompt, and 16 billion while it writes a response. That asymmetry is the point: most of the cost in running a large model comes from generation, not reading, so DeepSeek keeps the expensive part lean. The model also natively handles images and supports context windows up to 1 million tokens, according to BigGo Finance, which also reported that this release marks the first time DeepSeek has opened its proprietary Search API to the public.\n\nThe Search API matters more than it sounds. Agentic workflows, the kind where a model has to look something up, reason about it, and act, live or die on retrieval quality. Giving outside developers direct access to DeepSeek's own search infrastructure, rather than making them bolt on a third-party tool, is a bet that agent builders will pick the stack that's cheapest to wire together, not just the one with the highest leaderboard score.\n\n[DeepSeek is making its 75 percent API discount permanent](https://startupfortune.com/deepseek-is-making-its-75-percent-api-discount-permanent/)\n\nDeepSeek has made its 75 percent V4-Pro API discount permanent after the promotion period, resetting the cost expectations for AI startups and developers. The move puts pricing pressure on OpenAI, Anthropic and Google while raising fresh questions about DeepSeek’s margin strategy as it pursues a major funding round. - [deepseek api pricing discount permanent](https://startupfortune.com/deepseek-is-making-its-75-percent-api-discount-permanent/) - [cheapest frontier ai model for startups](https://startupfortune.com/deepseek-is-making-its-75-percent-api-discount-permanent/)\n\nPrice is where DeepSeek keeps pressing its advantage. Peak-hour API pricing for V4.1-Flash sits at $0.30 per million input tokens and $1.20 per million output tokens. Off-peak, that drops to $0.15 for uncached input and $0.60 for output, roughly half the peak rate. Compare that to Anthropic's Claude Opus or OpenAI's frontier-tier models, which routinely charge multiples of that for comparable agentic performance, and the math starts to explain why DeepSeek's downloads and API usage keep climbing even as Western labs tout bigger benchmark wins.\n\n## Why cost-per-task beats raw benchmark wins\n\nHere's the thing a lot of benchmark coverage misses. A founder shipping a customer support agent or a code-review bot doesn't care if a model scores five points higher on an abstract reasoning test. They care what it costs to run that agent a million times a month, and whether it finishes the job without burning through a context window or needing three retries. DeepSWE specifically measures whether a model can resolve software engineering tickets end to end, which is a much closer proxy for what agent builders actually ship than a multiple-choice exam.\n\nThat's exactly why V4.1-Flash-1017's jump from 74.2 to 77.0 on DeepSWE matters more to a startup's burn rate than it does to an academic leaderboard. A 2.8-point gain on a task-completion benchmark, paired with a price tag that's a fraction of GPT-5-class or Claude-class API costs, is the kind of upgrade that shows up directly in a company's cloud bill, not just in a marketing deck.\n\nDeepSeek has run this playbook before. V4.1-Flash itself launched on September 10 specifically to compete with the costlier V4-Pro on price, and DeepSeek initially planned to route V4-Pro API traffic to the cheaper Flash model starting September 14 before reversing course a day later and keeping V4-Pro available at its original pricing due to user demand. That back-and-forth is itself a company optimizing for adoption through price, not headlines. Frankly, that's the more dangerous competitor to watch than one that ships a flashy demo once a quarter: a lab that ships small, frequent, cheaper upgrades forces everyone downstream, OpenAI, Anthropic, Google, to defend their pricing every few weeks instead of once a year.\n\nNone of V4.1-Flash-1017's benchmark numbers are independently verified outside DeepSeek's own reporting, and on harder reasoning tests like Humanity's Last Exam, the model still trails top-tier closed models by a wide margin. But for the specific job agent-building founders are hiring these models to do, write code, call tools, finish multi-step tasks, DeepSeek just made that job cheaper again. It did it on a Saturday, with no announcement thread, and that's becoming the whole pattern: DeepSeek doesn't need a launch event when the price tag does the talking.\n\n**Also read:** [Accenture weighs $1.5 billion sale of its China consulting business](https://startupfortune.com/accenture-weighs-15-billion-sale-of-its-china-consulting-business/) • [Banks Add Sam Altman's World ID Badge to Calls to Spot Deepfakes](https://startupfortune.com/banks-add-sam-altmans-world-id-badge-to-calls-to-spot-deepfakes/) • [Meta blames Anthropic's Claude for wrongly suspending No Kings Instagram](https://startupfortune.com/meta-blames-anthropics-claude-for-wrongly-suspending-no-kings-instagram/)\n\n*This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.*\n\n## Join the discussion\n\n[Open in the community →](https://startupfortune.com/community/)\n\nAlmost there. Sign in and your reply posts straight away.", "url": "https://wpnews.pro/news/deepseek-quietly-ships-v4-1-flash-1017-beating-rivals-on-agent-tasks", "canonical_source": "https://startupfortune.com/deepseek-quietly-ships-v41-flash-1017-beating-rivals-on-agent-tasks/", "published_at": "2026-10-19 04:50:08+00:00", "updated_at": "2026-10-19 05:21:00.639220+00:00", "lang": "en", "topics": ["large-language-models", "ai-agents", "ai-products", "ai-startups", "ai-tools"], "entities": ["DeepSeek", "V4.1-Flash-1017", "V4.1-Flash", "Hugging Face", "ModelScope", "MarkTechPost", "BigGo Finance", "DeepSWE"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/deepseek-quietly-ships-v4-1-flash-1017-beating-rivals-on-agent-tasks", "markdown": "https://wpnews.pro/news/deepseek-quietly-ships-v4-1-flash-1017-beating-rivals-on-agent-tasks.md", "text": "https://wpnews.pro/news/deepseek-quietly-ships-v4-1-flash-1017-beating-rivals-on-agent-tasks.txt", "jsonld": "https://wpnews.pro/news/deepseek-quietly-ships-v4-1-flash-1017-beating-rivals-on-agent-tasks.jsonld"}}