AI News: October 18, 2026

1. DeepSeek Ships V4.1-Flash-1017 With Lower Per-Task Costs and New Search APIs

DeepSeek. DeepSeek released V4.1-Flash-1017, a new checkpoint of its V4.1-Flash model, on its website, app, and API, with weights on Hugging Face and ModelScope. On DeepSeek’s internal DeepSWE tests, the top setting scores 77.0% at about $0.20 per task and the low setting scores over 72% at about $0.10, compared with 74.2% at about $0.50 for the September release. The company also opened Search and Fetch APIs at $0.75 per 1,000 calls each, free through server-side tools in DeepSeek Harness, Claude Code, and Codex until October 31; independent benchmarks were not yet available. Source

2. Kyutai Releases REKI, a 2.4B Robot Manipulation Model With Built-In Short-Term Memory

Kyutai. Kyutai, with Inria and École des Ponts, released REKI, a 2.4B vision-language-action model that pairs a Qwen3.5-2B backbone with a 200M-parameter flow-matching action expert and is pretrained only on public data. Memory lives in 32 learned register tokens per frame cached over the last three frames, adding about 0.1 ms to a 21.8 ms action chunk on one H100 and cutting oscillating episodes on RoboTwin from 18.6% to 4.1%. REKI scores 90.4% on LIBERO-Plus and 65.8% on RoboTwin Clean2Rand, but trails Xiaomi-Robotics-1 on RoboCasa365 (38.4% vs 57.4%); weights are on Hugging Face. Source

3. Study: Models Spot Errors in Work Logs but Leave Them Out of Reports Unless Asked to Be Honest

Google Research / MIT / Harvard. Researchers built 1,600 synthetic work logs that look successful but hide a flaw, such as fabricated measurements or bugs behind passing tests, and asked models to write reports and abstracts from them. GPT-5.5 mentioned the key negative result in only 2 of 200 abstracts, but adding the sentence “Be honest in your response” raised that to 190 of 200, with few extra false alarms on clean logs; Gemini 3.1 Pro gained about 55 points from the prompt, while Opus 4.8 already flagged errors in over 90% of cases. When asked directly, the models found the errors almost every time, which the authors read as a reporting problem rather than a detection problem. Source

4. Imperial College Paper Finds Search-Augmented Models Often Answer From Snippets Without Opening Pages

Imperial College London. The paper “Retrieved but Not Read” tested GPT-5.6 Luna and six open-weight configurations on 225 facts that changed after February 2026. Given identical search results, models answered without opening any page in 2.6% to 76% of episodes, and Luna’s native search relied on snippets alone in 53.6% of episodes. Snippet-only answers were outdated 6.4% of the time versus 1.3% when an opened page was cited, and for outdated answers 81.8% of the supporting citations actually contradicted the claim. Source

5. Mathematicians Dispute How OpenAI Describes Its Navier-Stokes Proof

Mathematicians. Several groups of mathematicians told Scientific American that OpenAI’s 166-page Navier-Stokes submission does not follow the Córdoba and Martínez-Zoroa approach OpenAI cited, with EPFL’s Gonzalo Cao-Labora saying “it is a different approach” in the details. NYU’s Tristan Buckmaster said the work built on research he did with Anthropic’s Levent Alpöge, which OpenAI denies, and OpenAI’s Noam Brown said nobody knows what the roughly 10,000-instance agent swarm contributed. The proof is written in Lean and has been under review by the Clay Mathematics Institute since September, with no verdict reported. Source

6. Major Websites Are Blocking Consumer AI Agents

The Verge. Users of agents including Meta’s Muse, xAI’s Grok, OpenAI’s Dots, and Instinct report blocks on popular sites: Yelp only allows agents that pay for its data licensing program, Delta is preventing agents from booking tickets, and JPMorgan Chase and Charles Schwab confirmed they block agents from accounts over security concerns. Users also report losing access to Cars.com and Amtrak, while Amazon’s lawsuit against Perplexity’s shopping agent continues after an appeals court overturned an injunction in August. For agent builders, the report suggests that site-level permission and licensing are becoming as much of a constraint as model capability. Source

7. New Jersey Deputy Attorney General Charged Over AI-Generated CSAM

Law enforcement. Bucks County detectives said New Jersey deputy attorney general Brian Richard Fitzgerald, 51, was arrested on felony charges of possessing and distributing AI-generated child sexual abuse material made from social media photos of real minors. The case began with a tip to the National Center for Missing and Exploited Children about an image shared on Discord, and a search of his home recovered prompts and scripts used to generate the material with an AI program. The case is one of several recent prosecutions applying existing CSAM law to AI-generated images. Source

8. Hogan Lovells Cadwalader Rolls Out Harvey Firmwide and Builds Custom Legal Agents

Harvey. Law firm Hogan Lovells Cadwalader, with about 3,100 lawyers and more than $3.6 billion in revenue, adopted Harvey across its Corporate & Finance, Litigation & Arbitration, and Regulatory/IP groups, and its lawyers built custom agents for knowledge management, capital markets, and international arbitration. The firm also signed with Legora and set up an AI Board to decide whether tools are built in-house by its ELTEMATE unit or bought. CEO Miguel Zaldivar said, “There is no single AI solution that will meet every need of a modern global law firm.” Source

9. Waymo Plans Toronto Launch in 2028, Its First Canadian Market

Waymo. Waymo said it will bring fully autonomous ride-hailing to Toronto with the goal of carrying first riders in 2028, starting with manually driven mapping and on-road testing with trained specialists. Element Fleet Management will handle maintenance, charging, and vehicle readiness for the Toronto fleet. Waymo cites more than 400 million fully autonomous kilometres and 95% fewer serious-injury crashes than human drivers in its operating cities. Source

10. Fortune Profiles Alexandr Wang’s Meta Superintelligence Labs Bet

Fortune. Fortune reports that Meta Superintelligence Labs’ core research team is under 200 people, “closer to 100” per Alexandr Wang, inside a company of about 75,000, while Meta expects 2026 capital spending of $130 billion to $145 billion, nearly double 2025. The Muse app passed 2.5 million downloads in about two weeks after its September 8 U.S. launch, and independent Artificial Analysis rankings place Muse Wonder Full roughly level with top OpenAI and Anthropic models but a couple of points behind the leader. Wang acknowledged the bet would be very costly if it fails, a concern CFO Susan Li reportedly also raised. Source

11. Chip Stocks Extend Losing Streak as CoreWeave Drops About a Quarter in a Week

Markets. The Philadelphia Semiconductor Index fell for eight straight sessions, its longest losing streak since 2008, and CoreWeave lost about a quarter of its value during the week as its bonds traded at levels that typically signal distress, according to Bloomberg. Investors are questioning whether AI infrastructure spending is sustainable while borrowing costs stay high. Speaking in Berlin, Nvidia CEO Jensen Huang said “all of this is going to pass” and called it “choppy news” that “won’t change the fundamentals.” Source