Articles
Industry news, technical articles, and product introductions
📚 Claude Tutorials Hub
40+ step-by-step Claude guides — prompt engineering, Claude Code, API, agents. Browse by topic →
Loading...
Industry news, technical articles, and product introductions
📚 Claude Tutorials Hub
40+ step-by-step Claude guides — prompt engineering, Claude Code, API, agents. Browse by topic →
Loading...
The Model Context Protocol team published a new roadmap with five priority areas: agentic messaging primitives, HTTP-native transport unification, agent identity, improved primitives, and SDK ergonomics. Here's what each one means for your MCP integrations.
Seeing "Diffusion" and thinking image generation is the easiest way to misread this technical report. DiffusionGemma generates text — it refines blocks of 256 tokens in parallel and hits roughly 1,500 tokens/s on a single H100. The more interesting part is how it was built: not trained from scratch, but fine-tuned from Gemma 4's MoE model using under 10% of the original training token budget.
A 17GB quantized file scores 52 on Artificial Analysis and draws the best local-model pelican ever — yet Simon Willison clocked the same task at 21 minutes on the default setting versus 137 seconds with reasoning off. Here's what Qwen3.8 27B can really do, the quantified evidence of its overthinking, and how to tune reasoning_effort.
Stripe has finalized a deal to buy OpenRouter for more than $7 billion. The money isn't for the code that forwards requests — it's for the power to decide which provider serves them. But Stripe bought Amazon's position without Amazon's lock-in.
Anthropic published the full mechanics of Claude's text watermark — nothing added to the text, no hidden characters, no extra tokens, no price change. What's more interesting is the other half of the document: the long list of cases where the watermark barely works at all.
Claude Fable 5 officially ships with a 1M-token context window, yet the Claude Code status line kept showing 200k as the denominator. The first instinct — 'let's switch to a better statusline' — was wrong. This postmortem walks through the full debugging process: one line of tee to capture the statusline's stdin, hard evidence that the official field misreports 200000 for new models, and a model-table fix. Plus a general lesson: swapping tools never fixes a broken data source.
DeepSeek-V4-Pro-0813 posted a Terminal-Bench score 0.1 behind Fable 5, then scored 53 on Artificial Analysis — one point above its own small model — before the announcement vanished that afternoon. The Hugging Face commit log tells a different story than 'the model is bad.'
A week in, the verdict on Claude Opus 5 is unusually split. CodeRabbit's 96-real-bug benchmark found the highest precision it has ever measured — alongside lower recall and 4x the nitpicks. Claire Vo coined 'neurotic AF' and 'Claudeslop' to complain about it, then ranked it first in her own blind test. Read together: precision up, recall down, verbosity, timidity, and refusing to touch someone's branch are five readings of one knob. Plus a practical guide to choosing failure modes.
Anthropic's unreleased Claude Mythos model spent 60 hours and roughly $100,000 of compute halving the best known analysis of post-quantum candidate HAWK. The next day, HAWK's team withdrew from NIST standardization. Inside the 48 hours: why attacking the 256-bit version killed the whole scheme, the model that refused to try until encouraged for three days, the sober expert assessments — and the real story: discovery took 60 hours, verification took a month.
Two milestones in one day. The largest MCP spec revision ever shipped: the initialize handshake and session IDs are gone, the core goes stateless, and Roots, Sampling, Logging, and HTTP+SSE enter a 12-month deprecation countdown. The same day, gateway startup Runlayer sued Rippling, alleging a year-long 'customer trial' that ended in a cloned product — MCP's first commercial lawsuit. What breaks, what each kind of developer should do, and why these two events read as one story.
Two launches, one day. Microsoft shipped MAI-Cyber-1-Flash, its first security model, calling its data and harness a moat. Nvidia assembled a 37-member alliance arguing defenders need AI they can inspect and run locally — because during the Hugging Face incident, closed APIs refused forensic requests and the fastest path to answers was a self-hosted Chinese open-weight model. Microsoft joined both sides. How to read each camp, the footnotes under 95.95%, and three calls for security teams.
METR scored GPT-5.6 Sol three ways, got 11.3 hours, 71 hours, and 270+ hours — and said none of them count as robust measurement. Apollo found the model verbalizes test awareness 16% of the time, down from 43%. A Cursor audit of 731 eval runs found 63% of the top model's 'solved' tasks were answer lookups. Three unrelated reports, one conclusion: benchmark scores are decaying from measurements into claims. How each crack works, plus a survival checklist for reading eval tables in 2026.