Artifacts Go Public, Kimi K3 Weights Land, and Roblox Puts AI Game-Making on Phones
July 28, 2026
This week's real story is capability, not commentary: Claude Code artifacts can now be shared with the public and edited by a team in real time, Moonshot's 2.8T Kimi K3 finally has downloadable weights, and Roblox is testing prompt-to-game creation on a phone. We skipped the usual pile of model-benchmark roundups and hardware keyboard hype unless they actually change what you can ship this week.
code.claude.com · 2026-07-13Claude Code Artifacts: public links, multiplayer editing, and live MCP data
Anthropic shipped public sharing links and editor roles for shared editing on Team and Enterprise plans, plus artifacts created from Claude Tag sessions. A published artifact can now call MCP connectors on every view so a dashboard shows live data instead of a frozen snapshot from the build session, with each viewer authenticating through their own connections. On Pro and Max, a public link is the only way to share an artifact — no org audiences or editors — while Team/Enterprise gets org-only sharing until an Owner flips on public links.
Why it matters: If you've been prototyping client dashboards in Artifacts, you can now actually hand someone a working link instead of a screenshot — but know the catch: a connector-backed live dashboard cannot be a public link on any plan.
platform.claude.com · 2026-07-24Claude Sonnet 5's cheap pricing window extended to August 31
Anthropic confirmed Sonnet 5's introductory API pricing of $2/$10 per million input/output tokens holds through August 31, 2026, after which it steps up to the standard $3/$15 rate. The model is available by default in Claude Code and across Pro/Max/Free plans.
Why it matters: If you're running agentic coding workloads on Sonnet 5 through the API, you've got about five more weeks at the discounted rate before costs jump 50% — good time to lock in usage or benchmark whether you actually need it.
techtimes.com · 2026-07-27Kimi K3's 2.8T open weights are live on Hugging Face
Moonshot AI published the full Kimi K3 weights — a 2.8-trillion-parameter MoE model with a 1M-token context window — a day ahead of schedule, landing on Hugging Face under a bespoke license rather than MIT, with a revenue-triggered clause for hosted API operators and a UI attribution requirement above 100 million monthly users. Running it locally realistically requires multi-GPU server hardware, not a single workstation, so most builders will use hosted access via Together AI, Modal, or similar providers.
Why it matters: This is the first time you can actually inspect, fine-tune, or self-host a genuinely frontier-class coding model instead of only calling an API — check the license before you build a product on it, though, since it's not a clean MIT grant.
roblox.com · 2026-07-28Roblox Build: prompt-to-game creation lands on mobile, alpha starts today
Roblox's Build feature enters public alpha in New Zealand today, letting age-verified users 9+ turn a text prompt into a playable game directly in the mobile app, with games publishable globally once users are 16+ and pass safety review. It shares the same backend as Roblox Studio, so a project started on a phone can be picked up and extended on desktop.
Why it matters: If you build for the Roblox ecosystem, this is the first real test of whether prompt-built mobile games can compete in the same discovery feed as hand-coded ones — worth watching how the retention-based ranking treats them.
techcrunch.com · 2026-07-15OpenAI's $230 Codex Micro keyboard sold out in 12 hours
OpenAI's first hardware product, the Codex Micro, packs light-up agent-status keys, a reasoning-level dial, and a joystick for managing fleets of Codex coding agents, built with Work Louder. It sold out within 12 hours of launch and is now reselling on eBay for roughly $1,850.
Why it matters: Not something you can actually buy this week, but it signals where OpenAI thinks agentic coding workflows are headed — physical controls for juggling multiple background agents rather than tab-switching.
Researched with AI assistance against primary sources, then reviewed and published by a human. Spotted an error? Tell us.