DeepSeek just did it again. On October 17, the Chinese AI lab quietly pushed a new version of its Flash model live — no stage, no keynote, just an announcement on X and new numbers that landed right in the middle of America's most expensive AI race. The update, called DeepSeek-V4.1-Flash-1017, scores 77.0 on DeepSWE v1.1, a respected test of real software-engineering work. That puts it ahead of Anthropic's Claude Opus 5 at 74.0 in DeepSeek's own comparison table, and ahead of OpenAI's GPT-5.6 Sol at 73.0 — while charging a small fraction of what those US models cost per task.
DeepSeek is the lab that shook Wall Street in early 2025 by matching US frontier models at a fraction of the cost. This time it is a price-and-agents play: a cheaper coding model, plus DeepSeek's own Search and Fetch tools for AI agents, opened to every developer the same day.
Key Takeaways
- DeepSeek released V4.1-Flash-1017 on October 17, 2026. It is live now on DeepSeek's website, app and API under the existing model name deepseek-flash.
- In DeepSeek's own benchmark table, it leads on four of eleven tests — including DeepSWE v1.1 (77.0, ahead of Claude Opus 5 at 74.0) and Automation-Bench (60.5) — but Claude Opus 5 still leads on reasoning and several other coding tests.
- Peak API pricing is $0.30 per million input tokens and $1.20 per million output tokens, with off-peak rates at half that.
- DeepSeek also opened its Search and Fetch APIs to all developers at $0.75 per 1,000 calls, with agent-harness calls free until November 1, 2026.
- All headline scores are DeepSeek's own figures, not independent tests.
What DeepSeek Actually Released on October 17
V4.1-Flash-1017 is an updated checkpoint of the V4.1-Flash model launched in September, not a new generation. DeepSeek calls it a "Pareto improvement": higher scores, faster completion and shorter outputs at once. It says the update was trained with reinforcement learning using self-feedback, aimed at building complete projects end to end — from games to 3D models.
The model is built for agent work, tuned for DeepSeek Harness and Tencent's CodeBuddy and WorkBuddy. Reporting based on the V4.1-Flash technical paper describes a mixture-of-experts design — a 552-billion-parameter backbone plus 196 billion Engram parameters, activating only about 8–16 billion parameters per token. Its context window is 1 million tokens, with output up to 384,000 tokens: a very large model that switches on only a small slice of itself per word, which is how DeepSeek keeps it fast and cheap.
One honesty note: the September V4.1-Flash weights are open on Hugging Face under an MIT license, but MarkTechPost noted the 1017 checkpoint's weight details were not disclosed. What is certain is that anyone can use the model today through DeepSeek's app and API.
The Benchmark Story: Where 1017 Leads, and Where It Does Not
DeepSeek published an eleven-benchmark comparison against its previous Flash, Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. The new checkpoint beat the previous Flash on every test, by 2.0 to 11.1 points.
It leads the table on four tests: DeepSWE v1.1 at 77.0, Automation-Bench at 60.5, Agents' Last Exam at 35.4, and Apex Bench at 72.1. DeepSWE matters most for US software teams, because it measures real engineering work, not trivia — and there, the cheap challenger now sits above both Anthropic and OpenAI.
Now the other half. Claude Opus 5 still leads on Humanity's Last Exam (56.3 against 42.6), Terminal-Bench 4.0 (51.8 against 42.3), ProgramBench (37.0 against 29.1) and NL2Repo-Bench (75.3 against 68.2). Posts claiming the new Flash is simply "better than Opus 5" are cherry-picking one column. The accurate headline is narrower and still remarkable: on one major coding benchmark, a far cheaper model now scores higher — while the US frontier still wins on deep reasoning.
The Bigger Move: DeepSeek's Own Search and Fetch APIs for Agents
Buried in the same announcement is the move competitors may feel most. DeepSeek opened its Search API and Fetch API to every developer. Search returns ranked web results; Fetch returns a page as clean Markdown an agent can read. DeepSeek says its engine already serves hundreds of millions of requests a day.
Each endpoint costs $0.75 per 1,000 successful calls — failures are free — capped at 5 requests per second per account. Calls through server-side tools in DeepSeek Harness, Claude Code or Codex stay free until per-request billing begins on November 1, 2026. One API key now buys the model and the web layer underneath it, putting DeepSeek up against the dedicated search-API companies too.
Why This Lands Hard in the United States Right Now
Timing is everything. Just this week, Replit — until recently one of Anthropic's top five customers — made OpenAI its default, citing cost, partnership and performance. Even inside the US industry, platforms are openly shopping on price per task. Into that argument steps DeepSeek, at off-peak rates of $0.15 per million input tokens and $0.60 per million output tokens.
The pressure lands on companies spending at historic scale. Anthropic's revenue and IPO plans — covered in our report on Anthropic's $16.8 billion quarter and its reported IPO ambitions — rest on enterprise customers continuing to pay premium prices. DeepSeek, meanwhile, is reportedly raising $12–15 billion at a valuation of about $75 billion ahead of a planned Shanghai IPO, according to CNBC. For regular US users, the effect will be indirect: the AI tools you pay for may quietly get cheaper or faster as their makers renegotiate which model works behind the screen.
Before You Switch a Workload to the New Flash — A Sane Checklist
- Run your own test tasks first. Every score above is DeepSeek's own measurement; results on your codebase are the only numbers that count.
- Compare cost per finished task, not cost per token. Shorter outputs are part of this release's pitch, so measure completed jobs end to end.
- Keep a premium model for deep reasoning — on Humanity's Last Exam and Terminal-Bench, Claude Opus 5 still leads by a wide margin.
- Check the calendar: free Search and Fetch calls through agent harnesses end November 1, 2026, and off-peak API rates are half the peak rates.
- Review data rules before sending private code or documents to any cloud API, whichever country the provider is in.
Scam Warning: A Famous AI Name Is a Phishing Magnet
Every major DeepSeek release brings fake apps, copycat sites and phishing pages offering "free premium access." Only use DeepSeek's official website and app, and only create API keys inside your real account. Never buy "discounted credits" from a social media ad, and never paste your API key into a third-party page — a stolen key can run up charges in your name. Viral benchmark screenshots on X also overstated the results, so check claims against the published table before sharing them.
That keep-it-local instinct is worth applying to your phone, too. Code, contracts and ID scans you would never paste into a cloud chatbot deserve the same protection at rest. Pro Calculator Secure Vault stores photos, videos, audio, documents and private notes on your device only, behind sign-in protection — nothing is uploaded to a server for an AI, or anyone else, to read.
Frequently Asked Questions
What is DeepSeek-V4.1-Flash-1017?
It is an updated version of DeepSeek's V4.1-Flash model, released October 17, 2026, live on DeepSeek's website, app and API as deepseek-flash — with stronger coding and agent scores than the September version.
Did DeepSeek really beat Claude Opus 5?
On one major test, yes — in DeepSeek's own table it scores 77.0 on DeepSWE v1.1 against 74.0 for Claude Opus 5. But Opus 5 still leads on several other tests, including Humanity's Last Exam (56.3 vs 42.6) and Terminal-Bench 4.0 (51.8 vs 42.3). These are company-reported figures, not independent results.
How much does the new DeepSeek model cost to use?
Peak API pricing is $0.30 per million input tokens and $1.20 per million output tokens, with off-peak rates at half that. The Search and Fetch APIs cost $0.75 per 1,000 successful calls each, and harness-based calls are free until November 1, 2026.
What are DeepSeek's Search and Fetch APIs?
They let AI agents search the web and read pages through DeepSeek's own engine — Search returns ranked results, Fetch returns a page as Markdown — making DeepSeek a one-stop shop for the model and the web access agents need.
Should US developers switch from Claude or ChatGPT?
Not blindly. Test the new Flash on your own coding and agent tasks, compare cost per finished task, and keep a premium US model for the deep-reasoning work where it still clearly leads.
The Bottom Line
October 17 did not crown a new overall AI champion — Claude Opus 5's reasoning lead is intact in DeepSeek's own numbers. What it did was sharper: on real coding and agent work, the cheapest serious option just beat the most expensive ones on several scoreboards at once, with a ready-made search layer for agents attached. With US platforms like Replit already switching defaults over cost, that is exactly the kind of release that reprices an industry.
While the labs fight over your prompts, keep the files that never belong in a prompt under your own control. Pro Calculator Secure Vault keeps photos, videos, audio, documents and private notes behind sign-in protection on your own device, with an intruder selfie, an incognito private browser, and 8 smart calculators (standard, floating, BMI, age, currency, tip, GST & VAT, loan EMI) — local on-device storage only, with no data collected or shared with third parties. More from the maker on the ZAS developer page on Google Play.
No comments:
Post a Comment