Key Signals

• GitHub Copilot removes manual model selection for Free and Student plans, making auto-selection the only option. This forces millions of free-tier users onto dynamic model routing that picks the best model per task across multiple families. The retirement of `` labels from Microsoft models signals that Copilot's model routing is now considered production-ready. This move standardizes the experience across the largest segment of Copilot's user base and reduces user confusion over model choices. [1]

• Claude Code achieves a 37% reduction in CPU usage during streaming, addressing a key performance bottleneck for long coding sessions. The v2.1.191 release coalesces text updates to 100ms intervals and reduces memory growth from terminal output caching. Combined with MCP reliability improvements, this makes Claude Code substantially more resource-efficient for developers running extended agentic workflows. [2]

• Engineering hiring proves resilient to AI disruption, with engineers comprising 55% of new hires at major tech companies — up from 46% in 2019. SignalFire's analysis of millions of careers shows a Jevons paradox at work: AI-powered productivity expands work rather than eliminating it. Early-stage startups hired 7% more engineers than in 2019, and Nvidia CEO Jensen Huang stated that "software engineers are busier than ever" since adopting agentic AI. [3]

• The "tokenmaxxing" era is ending as companies scramble to impose AI budget controls after encouraging maximal usage. Accenture, which previously threatened promotion consequences for employees not using AI, is now trying to prevent workers from depleting token reserves on basic tasks like PDF-to-slide conversions. The shift reflects a broader industry reckoning where AI must prove cost-to-value rather than simply demonstrating novelty. [4]

• OpenAI unveils Jalapeño, its first custom inference ASIC co-developed with Broadcom, designed from scratch for LLM inference. Engineering samples are already running GPT-5.3-Codex-Spark workloads, with deployment planned at gigawatt scale in Microsoft data centers by year-end. The nine-month design-to-tapeout cycle — accelerated by OpenAI's own models — represents OpenAI's bid to control the full AI stack from silicon to services. [5]

• Claude Code team lead advocates replacing Markdown with HTML outputs for agentic coding workflows, citing cognitive bottleneck relief. Thariq Shihipar argues that as agent outputs grow in complexity, HTML's interactivity and visual density help developers stay engaged during review rather than rubber-stamping results. The idea has polarized the community, with supporters praising scan-ability and critics warning about security risks, token overhead, and poor git diff integration. [6]

AI Coding News

• Independent benchmarking of Anthropic's Mythos 5 model for security bug detection yields a "resounding maybe." Developer Joe Cooper built a custom benchmark suite testing whether Mythos can identify real-world security bugs without guidance. Mythos found four bugs that no other model detected, lending credence to Anthropic's claims about its vulnerability-finding capabilities. However, multi-file bugs remain extremely difficult for all reviewers — human or AI — and security experts caution that raw model capability alone doesn't constitute comprehensive security coverage. [7]

• AI is expanding beyond code generation into PRD governance, with Uber, DoorDash, and Cloudflare deploying AI earlier in the software lifecycle. Uber now uses AI to evaluate product requirement documents for clarity, completeness, and execution risks before they reach engineering teams. DoorDash built an AI code reviewer designed to "earn trust, not create noise" with fewer but more actionable comments. Cloudflare deploys specialized multi-agent review systems where each agent handles security, performance, or correctness independently — finding that "specialized agents outperform a single general-purpose reviewer when each is tightly scoped in responsibility." [8]

• Figma adds code layers directly to its collaborative canvas, deepening integrations with Claude Code and Codex. Teams can now clone repositories and iterate on code within Figma's spatial environment alongside designs, animations, and 3D transforms. The update also introduces AI-generated shader effects, custom agent skills via text prompts, and connections to tools like Notion and GitHub for richer context. Figma's CPO emphasized the canvas is designed for rapid exploration where "you don't really care about the quality of the code." [9]

Feature Update

• GitHub Copilot CLI v1.0.65 ships persistent working directories, a /skill subcommand, and CI status integration. The /cd command now persists across session resumes and automatically discovers custom agents in the new directory. A new opt-in status bar shows CI check status for the current branch. The copilot skill subcommand enables managing skills from files, URLs, or directories. Additional fixes include shell command history in normal mode via Ctrl+R, silent MCP OAuth scope reuse, and reliable inline image rendering in tmux. [10]

• Claude Code v2.1.191 adds /rewind support and resolves background agent persistence bugs. The /rewind command can now resume conversations from before /clear was run. Background agents that were stopped from the tasks panel can no longer resurrect — a fix for a frustrating issue in multi-agent workflows. The sandbox network permission dialog now remembers allowed hosts for the session duration instead of re-prompting on every connection. MCP error messages now include the URL and point to config for 404 errors. [2]

• OpenCode v1.17.10 adds MCP resource tools, managed provider integrations, and a --mini CLI mode. MCP server instructions are now included in session context, and new resource template listing and read tools expand MCP interoperability. The managed provider integration enables OpenCode-hosted provider support. Desktop improvements include per-tab draft state, a session progress indicator, mobile bottom navigation, and collapsible server sections. The SDK now exposes provider integration IDs, and the new V2 plugin API supports Effect and Promise patterns. [11]

• GitHub Copilot auto model selection becomes the sole model experience for Free and Student plans. Auto dynamically selects the best model for each task across multiple model families, removing manual model choice entirely. The `` label is retired from Microsoft-released models, reflecting that auto-selection's routing intelligence is now considered mature enough to guide all model decisions without user intervention. [1]