Tech • AI • Robotics • Game
DeepSeek Fires Again! (New #1 Open Source Coder Beats Claude)
7/10TL;DR
DeepSeek jolted China’s AI market with the surprise V4.1 Flash 1017 update, sharply improving coding and agent performance while cutting task costs, as rivals including MiniMax moved quickly to answer.
KEY POINTS
Surprise model jump
DeepSeek quietly renamed V4.1 Flash to V4.1 Flash 1017 on October 17, and the minor-looking update delivered major benchmark gains. Terminal Bench 4.0 rose from 31.2 to 42.3, while Program Bench for reverse engineering climbed from 20.3 to 29.1, flipping several coding and terminal-work tests into wins over leading Western models.
Cheaper coding performance
On DeepSWE 1.1, the model reached about 77% pass@1, up from roughly 74% for the prior V4.1 Flash and 63% for V4 Pro, at around 1.3 yuan per task. Average task cost reportedly fell from about 0.49 yuan to 0.20 yuan, a drop of roughly 60%, helped by shorter outputs and less wasted token use.
Open-source leaderboard gains
By October 19, the Max variant had overtaken Kimi K3 Max as the top open-source model on Code Arena WebDev with 1,670 points. It ranked ahead of several closed models and sat just 12 points behind GPT-6 Soul Max, while early independent testing placed its bug-fixing ability close to premium Claude Opus tiers at a fraction of the price.
Pricing strategy undercuts rivals
Peak API pricing was listed at 2 yuan per million input tokens and 8 yuan per million output tokens, with off-peak rates at 1 and 4 yuan. Cached input could fall as low as 0.04 or 0.02 yuan, making usage about a tenth of some domestic rivals and reinforcing DeepSeek’s position as a price disruptor.
Post-training, not a new architecture
DeepSeek did not label the release V4.2 because the base architecture was unchanged. The model kept its 552 billion backbone parameters, with 8 billion active on input and 16 billion on output, and the gains were attributed to post-training, especially reinforcement learning with self-feedback to improve planning, coding and self-checking.
Agent tools and “wish-based programming”
The update was paired with a push toward agents that can take a plain-language request, plan work, write code, test it and return something runnable. DeepSeek has been training that workflow in DSAC, a sandbox for large-scale agentic learning, and showcased browser games, interactive 3D scenes, narrated science animations and small simulations built from very few prompts.
Consumer expansion despite earlier skepticism
Although founder Liang Wenfeng had previously downplayed consumer products, the company has been broadening its reach. An internal desktop tool was open-sourced in August under MIT, then quickly gained one-click macOS and Windows installers, while holiday promotions and credits helped attract new users.
Outage and aggressive refunds
On October 15, DeepSeek suffered a roughly five-hour outage affecting chat and major APIs. The company later apologized, pledged to remove single points of failure, and refunded user spending from October 8 to 14 as credits, with screenshots showing reimbursements ranging from small sums to 3,000 yuan.
Search enters the stack
The same weekend, DeepSeek opened two APIs from its in-house search engine: Search, which returns structured JSON results with summaries and scores, and Fetch, which converts URLs into clean markdown. Pricing was set at 5 yuan per 10,000 successful requests, materially below many third-party search APIs, though critics noted tighter rate limits and a smaller index than Google or Bing.
MiniMax counters with M3.1 Flash
Rival MiniMax responded with M3.1 Flash, positioning it as the second-best coding value in the market. On DeepSWE 1.1 it scored 73.7, above GPT-5.6 Soul’s 73.0, and on Zapier’s Automation Bench it led with 62.8, nearly 12 points ahead of Opus 5, while maintaining low task costs.
Inference stack goes open source
On October 20, DeepSeek open-sourced DeepInfer, the inference framework behind its own service, under Apache 2.0, along with a codec called DeepSeek Recipe. The company said one cluster serves 155 trillion tokens a day, average requests see roughly 230 tokens per second, and cache hit rates near 96% can lift throughput by 25 times.
Funding surge and broader implications
Investor interest has accelerated alongside the technical gains. DeepSeek reportedly closed a funding round of about 100 billion yuan, or roughly $15 billion, with Tencent and CL among major backers, underscoring how pricing power, distribution and infrastructure are becoming as important as raw model quality in China’s AI race.
CONCLUSION
DeepSeek’s October releases showed that marginal-looking updates can reshape the competitive order when paired with lower prices, stronger agent tooling and infrastructure scale. The result is a faster-moving contest in which Chinese labs are competing on both model quality and the economics of real-world deployment.
Ask a question