Planet AI Weekly: Planet AI Weekly June 28, 2026

This week we reviewed 242 articles from 20 sources.

OpenAI's GPT-5.6 preview dominated headlines after White House intervention. Mistral OCR 4 shipped structured document output, Liquid AI squeezed on-device inference onto a Raspberry Pi, and the community debated cost-optimization traps and custom chip momentum.

Official Highlights

OpenAI previewed GPT-5.6 with tiered models (Sol, Terra, Luna) and new reasoning modes, but restricted access following a White House request. What early engineers should know about the controlled deployment. Read more

Mistral OCR 4 extracts structured document blocks with bounding boxes, typed classifications, and per-word confidence scores across 170 languages. Built for RAG and enterprise search pipelines that need citation-ready output. Read more

Liquid AI shipped LFM2.5-230M, a 230M-parameter open-weight model that runs at 213 tok/s on a Galaxy S25 Ultra and 42 tok/s on a Raspberry Pi 5. Broad runtime support via llama.cpp, MLX, vLLM, SGLang, and ONNX. Read more

DeepSeek open-sourced DSpark, a speculative decoding framework that accelerates DeepSeek-V4 per-user generation by 60-85% over MTP-1. Combines a parallel draft backbone with confidence-scheduled verification to cut suffix decay. Read more

Samsung Electronics deployed ChatGPT Enterprise and Codex to employees worldwide, marking one of OpenAI's largest enterprise rollouts to date. Read more

From the Community

Ari Joury pitted XGBoost against logistic regression on 358 matches. The boring model won - a concrete bias-variance lesson on knowing when to skip the big hammer. Read more

Pratik R's team slashed AI inference costs by half with a routing layer. Three months later, customer satisfaction tanked. The post-mortem on the Pareto trap in cost-optimization routing. Read more

Sandeep Bharadwaj Mannapur benchmarked where GBDTs beat agents on payment fraud - and where agents earn their keep. Latency, cost, and reproducibility quantified. Read more

Emmimal P Alexander built a context graph layer for multi-agent memory after finding vector RAG insufficient. Benchmarks raw chat history, vector RAG, and the graph against relational retrieval blind spots. Read more

AWS and co-authors detail agentic overlays - thin wrappers that turn REST services into agents for multi-agent workflows without rebuilding legacy systems. Read more

Groq confirmed a $650M raise and re-staffed after Nvidia's $20B not-acqui-hire deal fell through. Pivoting to neocloud with new execs. Read more

Theresa Loconsolo on why OpenAI, SpaceX, Google, and Apple are all building custom chips - OpenAI's Jalapeno with Broadcom joins the race against Nvidia's dominance. Read more

WIRED's Uncanny Valley podcast unpacks Amazon dropping its OpenAI movie, data center worker pushback, and Meta's leaked employee keystroke data. Read more

Featured This Week

Pratik R's "We Built a Routing Layer to Cut Our AI Costs. It Broke the Product." is a rare production autopsy. The team halved inference spend through model routing, then watched customer satisfaction drop in lockstep. The post traces how latency and quality tradeoffs hide in unit economics until they don't, and offers a concrete detection methodology for teams weighing similar optimizations. Essential reading for anyone balancing cost against UX in live AI systems. Read more

Send comments and story tips to tips@planet-ai.net. If you found this useful, share it with a colleague.