GLM-5.2 Matches Claude Opus at One-Sixth the Price, OpenAI Builds Its Own Chip

News

• Snowflake’s CEO ran a direct benchmark: Chinese model GLM-5.2 nearly matched Claude Opus 4.7 (66% vs 67% task solve rate) at just $4.40/M output tokens — about one-sixth of Opus’s $25/M. GLM burns twice as many tokens, which narrows the gap, but the pricing pressure on Western AI labs right in their core coding use case is real. https://the-decoder.com/snowflake-ceo-finds-glm-5-2-competitive-with-opus-4-7-at-a-fraction-of-the-cost/

• OpenAI and Broadcom unveiled Jalapeño, a custom chip built from scratch for LLM inference. Designed in nine months with OpenAI’s own models helping accelerate the process, Microsoft is taking 40% of output and a gigawatt-scale deployment is planned for late 2026. Performance-per-watt claims are self-reported; independent verification is still pending. https://the-decoder.com/openai-and-broadcom-unveil-jalapeno-a-custom-chip-built-for-llm-inference/

• Codex crossed 4 million weekly active users — a 5x jump in three months, with Germany up 720% since January. OpenAI’s DeployCo subsidiary embeds engineers directly inside corporate clients to build AI systems hands-on, positioning itself between frontier research and enterprise deployment. https://the-decoder.com/openais-deployment-chief-on-codex-growth-falling-ai-prices-and-the-roi-question/

• Figma’s Config 2026 brought code, motion, 3D, and shaders onto the canvas — a response to Anthropic’s Claude Design threatening to cut Figma out of the workflow. The catch: Figma’s own AI features run on Anthropic, OpenAI, and Google models. Figma’s answer is to stay model-agnostic and keep the designer inside the canvas. https://the-decoder.com/figma-bets-on-human-judgment-at-config-2026-while-the-ai-powering-its-canvas-belongs-to-someone-else/

• SignalFire data: engineering was the most AI-resilient job function in 2025. Engineers made up 55% of new hires at big tech (up from 46% in 2019), and early-stage startups hired 7% more engineers than in 2019. Classic Jevons paradox — productivity gains expand the work rather than replace it. https://techcrunch.com/2026/06/24/ai-was-supposed-to-kill-engineering-jobs-but-new-data-suggests-theyre-the-most-resilient/

• Cerebras dropped ~20% after its first post-IPO earnings report. Revenue grew 94% YoY, but gross margin guidance of 38–41% disappointed markets. CEO says investors misread the numbers. https://techcrunch.com/2026/06/24/cerebras-stock-plunges-after-earnings-as-ceo-says-margin-outlook-was-misunderstood/

GitHub Trending

• calesthio/OpenMontage — open-source agentic video production system: describe what you want and the agent handles research, scripting, asset generation, editing, and final composition. Produced a 60-second animated short for $1.33. https://github.com/calesthio/OpenMontage

• interviewstreet/hiring-agent — parses résumé PDFs, enriches with GitHub signals, and outputs a transparent, scored evaluation. Runs fully local with Ollama. https://github.com/interviewstreet/hiring-agent

• revfactory/harness — type “build a harness for this project” in Claude Code and it generates a full agent team architecture plus skill files, choosing from six team patterns (pipeline, expert pool, fan-out/fan-in, and more). https://github.com/revfactory/harness

• JCodesMore/ai-website-cloner-template — point Claude Code at any URL, run /clone-website, and it reverse-engineers the site into a Next.js codebase with parallel component builders. https://github.com/JCodesMore/ai-website-cloner-template

• ZhuLinsen/daily_stock_analysis — LLM-powered multi-market stock analysis (A-share, US, Japan, Korea) with GitHub Actions scheduled runs and push notifications to Telegram, Slack, or email. https://github.com/ZhuLinsen/daily_stock_analysis

Community

• Tom MacWright via Simon Willison: “LLM-written résumé → LLM portfolio → LLM commits… I don’t know anything about this person.” AI-polished applications are making candidates anonymous — the perfected generated résumé says nothing real. https://simonwillison.net/2026/Jun/24/tom-macwright/#atom-everything

• VibeThinker-3B paper: exploring verifiable reasoning in a 3B-parameter model. Evidence that small models can go further than expected with the right training setup. https://arxiv.org/abs/2606.16140

• Lobsters thread on a fully local voice assistant setup — no cloud, no data leaving your machine. A practical guide for anyone who wants an AI assistant without the privacy trade-off. https://blog.platypush.tech/article/Local-voice-assistant