Google releases Gemini 3.8 Flash, its third Flash model in six weeks

Anthropic slashes agentic workload costs 45% with Fable 5.1 as token prices crater industry-wide, while Perplexity and CrowdStrike race to solve agent privacy and security gaps.

• Anthropic launched Claude Fable 5.1 and Mythos 5.1: 25% cheaper for typical workloads, 45% cheaper for agentic tasks (via cache-read cost cuts), same $10/$50 per-million-token sticker price but far more token-efficient than Fable 5, beating GPT-5.6 Sol on coding/knowledge benchmarks.
• Fable 5.1 adds Enterprise Frontier Safeguards (ZDR-equivalent data retention) and reduced false positives in cybersecurity use cases; early testers (Cognition, Ramp, Canva, Block) report ~2x speed and half the token usage vs Opus 5.
• Market-wide token prices dropped from $2.07/M in May to $0.97/M by August 31 (LLM Token Expenditure Index), signaling frontier models are becoming commoditized as routing services let buyers ignore which model they're actually using.
• CrowdStrike unveiled Falcon Guardian, a new 'AIDR' (AI Detection and Response) category product that monitors, controls, and shuts down rogue AI agents on enterprise networks — a direct response to recent uncontrolled agent incidents at OpenAI, Anthropic, and Meta.
• Perplexity launched Hybrid Compute: an agent that auto-detects sensitive data (legal, health, customer PII) and routes it to local models (Gemma E4B, Qwen3.6 35B-A3B) on-device while sending general tasks to cloud frontier models — a privacy pattern likely to be copied industry-wide within 12-18 months.

Anthropic open-sources a shopping-agent blueprint with proven conversion lifts, while Qwen and Google ship incremental but concrete model upgrades at unchanged pricing.

• Anthropic open-sourced Claude Commerce Agents: a ready-to-deploy shopping + merchant agent kit with retail/travel/telecom/entertainment demos, showing 35% larger carts and 60% higher purchase completion in pilots.
• Alibaba updated Qwen3.8-Max to Qwen3.8-Max-0902 (2.4T params, 95B active, 1M context) with post-training focused on coding/collaboration, beating GPT-5.6 and Claude Fable 5 on PaperBench and OSWorld-Verified, same $2/$6 per 1M token pricing.
• Google shipped Gemini 3.8 Flash with deeper internal reasoning and repeated tool calls, scoring 89.4% on Terminal-Bench 2.1 (beating Claude Opus 5) at the same $0.75/$3.75 per 1M token price as 3.7 Flash.
• Meta released Muse Spark 1.3, cutting token usage 25% while improving agentic coding performance.
• An 8-year research paper claims LLMs build symbolic structure inside their internal vectors without being trained to, hinting at future interpretability tools beyond prompting.

Cursor's new Claude Fable 5.1 tops its coding benchmark at 73.4% by self-verifying code before finishing a task, while a new paper argues agents—not codebases—are becoming the software itself.

• Cursor shipped Claude Fable 5.1, scoring 73.4% on CursorBench 3.2 and beating all prior models tested; it self-verifies code and runs multi-step tasks with less babysitting, plus 75% cheaper cache reads than Fable 5
• New paper 'The End of Software Engineering' argues agentic software has no permanent codebase — the agent IS the product, generating and discarding code on the fly; developer role shifts to 'intent architect'
• Nous Research released Hermes Agent v0.21.0 with Bot Mode: named AI agents collaborate like a Slack team, plus memory-aware cron jobs, live subagent steering, and ~50% lower context usage
• Gemini's new video model cuts inference costs 66% by selectively processing only relevant frames instead of the whole video
• Google published a training technique that makes LLMs explicitly signal uncertainty/confidence about their own outputs
• An Anthropic hackathon winner open-sourced a 68-agent, 286-skill engineering team built on Claude, showing multi-agent orchestration is becoming replicable at scale

llm-gemini 0.34

• <p><strong>Release:</strong> <a href="https://github.com/simonw/llm-gemini/releases/tag/0.34">llm-gemini 0.34</a></p> <blockquote> <ul> <li>New model <code>gemini-3.8-flash</code> for <a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/">Gem