--- title: Qwen 3.8 27B floods every feed, Anthropic admits its... description: Alibaba's Qwen 3.8 27B dominates today's posts, with a dozen creators independently testing the new 27B dense model on consumer GPUs and getting wildly... canonical: https://daily.dev/posts/qwen-3-8-27b-floods-every-feed-anthropic-admits-its-bioweapon-filters-were-off-for-11-months-zgsc5jhy9 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Qwen 3.8 27B floods every feed, Anthropic admits its bioweapon filters were off for 11 months | daily.dev og:description: Alibaba's Qwen 3.8 27B dominates today's posts, with a dozen creators independently testing the new 27B dense model on consumer GPUs and getting wildly... og:url: https://daily.dev/posts/qwen-3-8-27b-floods-every-feed-anthropic-admits-its-bioweapon-filters-were-off-for-11-months-zgsc5jhy9 og:image: https://api.daily.dev/og/posts/ZgsC5jHY9.png og:image:alt: Qwen 3.8 27B floods every feed, Anthropic admits its bioweapon filters were off for 11 months og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Qwen 3.8 27B floods every feed, Anthropic admits its bioweapon filters were off for 11 months **[Agentic Digest](https://daily.dev/sources/agents_digest)** · 4 min read · 4 upvotes · 0 comments ## Summary Alibaba's Qwen 3.8 27B dominates today's posts, with a dozen creators independently testing the new 27B dense model on consumer GPUs and getting wildly different verdicts on whether it actually rivals Opus. Anthropic's second Risk Report disclosed that bioweapon-blocking classifiers were silently disabled for eleven months across 133 million contractor chats. Two research papers landed with practical implications for anyone building coding agents: one shows code-based tool calling beats JSON at scale, the other shows verbose 'consider multiple approaches' prompts can waste 18x the compute for zero benefit. Google also quietly shipped Gemini 3.7 Flash at half the price of its predecessor. ## Content **TLDR:** Alibaba's Qwen 3.8 27B dominates today's posts, with a dozen creators independently testing the new 27B dense model on consumer GPUs and getting wildly different verdicts on whether it actually rivals Opus. Anthropic's second Risk Report disclosed that bioweapon-blocking classifiers were silently disabled for eleven months across 133 million contractor chats. Two research papers landed with practical implications for anyone building coding agents: one shows code-based tool calling beats JSON at scale, the other shows verbose 'consider multiple approaches' prompts can waste 18x the compute for zero benefit. Google also quietly shipped Gemini 3.7 Flash at half the price of its predecessor. --- ## Qwen 3.8 27B ships, and testers can't agree on whether it's actually good Alibaba's Qwen 3.8 27B is a dense, Apache 2.0 multimodal model with native vision, a 262K context window (extendable to 1M via YaRN), and adjustable reasoning effort. It runs on a single consumer GPU (RTX 4090, 16GB cards via quantization) and Alibaba's own benchmarks claim it beats Opus 4.6 on SWE-Bench Pro, LiveCodeBench, and OSWorld Verified. But independent testers are split: one comparison found it missed a real bug that the previous Qwen 3.6 caught and produced a fix that failed Copilot's review, while others ran clean one-shot games and got near-perfect agentic tool-calling scores. Decode speed on 16GB cards is rough, 3-6 tokens/sec, meaning some of these 'one-shot' wins took over an hour to generate. [Read more](https://daily.dev/feed-by-ids?id=IVRmYAfnh&id=wR5ASB5u0&id=ZhgzJwoTA&id=NE7DqAELb&id=0gLV9a5Ty&id=q5cxtD6rc&id=qKBGOpVtl&id=uUn4qzn8P&id=TVkSZbzgJ) ## Anthropic disclosure: bioweapon filters were off for 11 months, nobody noticed Anthropic's second Risk Report reveals that bioweapon-blocking classifiers on its contractor feedback platforms were silently disabled from May 2025 to April 2026, covering roughly 133 million exchanges from about 50,000 vendor-vetted contractors, some without screening capable of stopping serious threat actors. An internal flag disabled both the blocking and the logging at once, so nobody could review what slipped through after the fact. A Claude Sonnet 5 review flagged 1,197 high-risk transcripts; manual review found no clear misuse but Anthropic now believes similar undiscovered issues are more likely than not. The company also downgraded its February safety verdict and disclosed a separate incident where contractors used a leaked API key to access an unreleased model, Mythos Preview, for two weeks. [Read more](https://daily.dev/posts/ulYjzriyL) ## Two papers rewrite how you should prompt and structure coding agents A Bitter-Lesson-style paper found that letting capable models write Python scripts to invoke tools beats standard JSON tool calling in 11 of 14 models tested, with the gap widening sharply at scale: Claude Sonnet 5 using JSON dropped to 0% accuracy at 100 chained calls while Python stayed at 100%. Separately, a study on prompt-induced waste found that asking agents to 'consider several approaches' inflated reasoning compute 2.4-7.4x with zero success gain, and 'be absolutely certain' triggered redundant re-verification costing up to 18x the median run. A bounded instruction, start with the failing test, make the smallest sufficient change, stop, kept costs at baseline without hurting correctness. Worth stealing for your own system prompts. [Read more](https://daily.dev/feed-by-ids?id=W16t9RHrY&id=eCSnZYPUT) ## Gemini 3.7 Flash undercuts its own predecessor on price Google released Gemini 3.7 Flash at $0.75/$3.75 per million input/output tokens, half the cost of Gemini 3.6 Flash, positioned as its workhorse model for coding and agentic tasks. It's available across the Gemini API, Android Studio, Antigravity, and Gemini Enterprise. The pricing is introductory through the end of the year. [Read more](https://daily.dev/posts/dJQVViXMv) --- ## Also notable - **AI store manager fires an employee, but only after being reminded of its own policy:** Luna, an AI store manager built on Claude Sonnet 4.6 for Andon Labs' Andon Market experiment, fired a human for lateness only after staff prompted it to look up an attendance policy it had created and then forgotten months earlier. [Read more](https://daily.dev/posts/PC6HrcWzG) - **Microsoft kills Excel's =COPILOT function after one year:** Microsoft is removing the =COPILOT formula function from Excel on September 14, 2026, folding its capabilities into the Copilot side pane instead. [Read more](https://daily.dev/posts/SnahFIF9I) - **IBM's BenchDrift shows top models are the most wording-dependent:** Testing eight models on GSM8K, MMLU, and MATH-Hard, IBM researchers found that meaning-preserving rephrasings flip correctness more often for strong models than weak ones, meaning leaderboard-topping scores may partly reflect phrasing luck rather than capability. [Read more](https://daily.dev/posts/W84mUd7d5) - **Repeated sampling beats self-reflection at equal token cost:** A study on Qwen2.5 models (1.5B-7B) found that across 36 comparisons, elaborate self-critique methods like Self-Refine and Reflexion never reliably beat simply resampling the same prompt multiple times and taking the majority answer at equal token budget. [Read more](https://daily.dev/posts/R9NmqBGPQ) - **Command Code's $10/month plan undercuts a tightening model-subscription market:** Command Code's new 'Goat' plan bundles $70 in monthly credits across 33 models for $10, launching as GLM raises its own coding plan to $18-160/month and Moonshot pauses new Kimi subscriptions after hitting GPU capacity limits. [Read more](https://daily.dev/posts/ZsRSzI3sY) --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#anthropic](https://daily.dev/tags/anthropic), [#prompt-engineering](https://daily.dev/tags/prompt-engineering), [#ai-safety](https://daily.dev/tags/ai-safety), [#qwen](https://daily.dev/tags/qwen) [View this post on daily.dev](https://daily.dev/posts/qwen-3-8-27b-floods-every-feed-anthropic-admits-its-bioweapon-filters-were-off-for-11-months-zgsc5jhy9) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"DiscussionForumPosting","mainEntityOfPage":"https://daily.dev/posts/qwen-3-8-27b-floods-every-feed-anthropic-admits-its-bioweapon-filters-were-off-for-11-months-zgsc5jhy9","headline":"Qwen 3.8 27B floods every feed, Anthropic admits its bioweapon filters were off for 11 months","text":"Alibaba's Qwen 3.8 27B dominates today's posts, with a dozen creators independently testing the new 27B dense model on consumer GPUs and getting wildly different verdicts on whether it actually rivals Opus. Anthropic's second Risk Report disclosed that bioweapon-blocking classifiers were silently disabled for eleven months across 133 million contractor chats. Two research papers landed with practical implications for anyone building coding agents: one shows code-based tool calling beats JSON at scale, the other shows verbose 'consider multiple approaches' prompts can waste 18x the compute for zero benefit. Google also quietly shipped Gemini 3.7 Flash at half the price of its predecessor.","url":"https://daily.dev/posts/qwen-3-8-27b-floods-every-feed-anthropic-admits-its-bioweapon-filters-were-off-for-11-months-zgsc5jhy9","datePublished":"2026-08-16T04:17:42.605Z","dateModified":"2026-09-13T21:34:09.336Z","author":{"@type":"Organization","name":"Agentic Digest","logo":"https://media.daily.dev/image/upload/s--V91DY4ls--/f_auto,q_auto/v1772617267/logos/agents_digest","url":"https://daily.dev/sources/agents_digest"},"interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":4},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"isPartOf":{"@type":"WebPage","url":"https://daily.dev/sources/agents_digest","name":"Agentic Digest"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Agentic Digest","item":"https://daily.dev/sources/agents_digest"},{"@type":"ListItem","position":3,"name":"Qwen 3.8 27B floods every feed, Anthropic admits its bioweapon filters were off for 11 months"}]} ```