AI News — October 18, 2026: DeepSeek Flash Beats Claude on 10 Benchmarks, Yelp and Delta Lock Out Agents

DeepSeek released DeepSeek V4.1-Flash-1017, a minor update to V4.1-Flash that reportedly beats Claude Sonnet 5.6 on 10 of 11 coding benchmarks while producing roughly 20% fewer tokens at the same intelligence level, and users on Hacker News say they keep choosing it for daily work despite weaker paper scores. Separately, Yelp, Delta, JPMorgan Chase, Charles Schwab, Cars.com and Amtrak are actively blocking AI agents from OpenAI, Meta, xAI and others, citing security, licensing and lost ad revenue, while Replit CEO Amjad Masad told TechCrunch Disrupt that OpenAI is now Replit's default model and Meta removed commercial-terms language that lawyers read as permitting it to scrape publisher sites for AI training.

Good morning. DeepSeek quietly dropped a Flash update that people seem to actually prefer in daily use even when it loses on paper, and the open-weights world keeps getting more crowded with Black Forest Labs and Fable both shipping. Meanwhile, websites are starting to slam the door on AI agents, Replit is breaking up with Anthropic sort of , and Meta had to walk back some terms-of-service language that lawyers found a little too ambitious. DeepSeek V4.1-Flash-1017 lands, and users keep picking it. DeepSeek released a minor update to V4.1-Flash https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017 that reportedly beats Claude Sonnet 5.6 on 10 of 11 coding benchmarks while producing roughly 20% fewer tokens at the same intelligence level. The HN thread is full of people describing a pattern: benchmarks say Claude or Gemini is smarter, but they keep reaching for DeepSeek because the speed and token efficiency change how it feels to work with. Several commenters mentioned pairing Flash with Opus 5.5 for the hard stuff — fast cheap default, expensive brain on tap. Black Forest Labs ships FLUX 3 Image Klein. BFL’s new 5.6B open-weights image model https://bfl.ai/blog/flux-3-image-klein does native 4K, supports up to 10 reference images, runs on 8GB VRAM, and generates a 1MP image in about 4 seconds on an RTX 5090 — priced at $0.009 per 1K images via API. It reportedly outperforms larger open models like Qwen-Image 2.1 and Krea 2 in internal benchmarks, though users on Reddit note it’s noticeably bad at generating people, apparently because NSFW-adjacent content was aggressively scrubbed from training. Fine text rendering is also weak, and there are scattered reports of body horror. Fable’s Ultramarine 1 Flash puts RL on top of GLM-5.3. Fable released Ultramarine 1 Flash https://www.paperinstruments.com/ultramarine , an agentic model aimed at finance, law, and consulting work, built by applying large-scale RL to GLM-5.3 Flash. The pitch is competitive performance with Claude Fable 5 and GPT-5.6 Sol at roughly 60x lower cost — $0.13 versus $8.19 per task. HN was skeptical: the benchmark charts conveniently omit stronger models like GPT-6 Astra and Opus 5.5, and one commenter asked flatly whether this is meaningfully different from GLM-5.3 with context tricks and guardrails. Websites are blocking AI agents. The Verge reports that Yelp, Delta, JPMorgan Chase, Charles Schwab, Cars.com, and Amtrak https://www.theverge.com/ai-artificial-intelligence/1013195/ai-agents-blocked-web are actively blocking agents from OpenAI, Meta, xAI and others, citing security, licensing, and lost ad revenue. It’s the same fight Amazon picked with Perplexity, just broader: AI companies want agents treated as user proxies, platforms want humans in the funnel. How this gets resolved will shape what the agentic web actually looks like. Replit swaps Claude for ChatGPT as default. Replit CEO Amjad Masad told TechCrunch Disrupt https://techcrunch.com/2026/10/17/replit-is-no-longer-working-with-anthropic-ceo-says/ the company has made OpenAI its default model, citing cost, performance, and partnership strength. Replit was reportedly a top-five Anthropic customer before an August OpenAI deal. The story got messy fast — Anthropic insiders disputed the “no longer working with” framing, and Replit’s own late-September blog posts still mentioned Claude running in production. Meta quietly backs out of a scraping clause. Meta added language to its commercial terms that lawyers read as giving it rights to scrape any publisher’s site with a Meta business account for AI training, potentially overriding robots.txt. After the backlash, Meta removed the language https://www.theverge.com/tech/1013409/meta-removes-language-that-may-have-allowed-it-to-scrape-news-sites and insisted it only ever applied to ads and business products — a position TechCrunch notes https://techcrunch.com/2026/10/17/meta-removes-language-that-raised-scraping-and-ai-training-concerns/ is awkward given active copyright suits over exactly this behavior. Two for the reading list. Alex Wang’s post arguing you could have invented transformers from memory https://alexwang.ai/posts/associative-memory/ reframes each layer as an associative memory store, walking from linear layers ~64 memories to MLPs ~2,700 to self-attention over a million . HN liked the framing but pushed back on the assumption of random unit-vector keys — real data has structure, and that structure is where transformers actually earn their keep. For the hardware-curious, Llmspeed https://llmspeed.dev/ lets you preview local LLM token speeds across hardware configs before buying, though commenters warn the numbers age fast as inference engines like Strata keep squeezing more out of the same silicon — and that just renting a GPU for 50 cents is probably more honest. That’s the morning. Watch the blocked-agents story — it’s the kind of slow-moving fight that quietly decides what the next version of the web looks like.