Issue #3 · Sent Oct 30, 2026 · Week of Oct 23–29, 2026
GitHub picks your model for you, OpenAI's cheaper Terra, and a Railway security heads-up
This week: what our weekly fact check changed, then eight releases and updates that matter for builders, each with a source. If you host anything on Railway, read the Railway item first.
What our weekly check changed
On Oct 28 we rechecked 66 of our 67 listed tools against their own pricing pages, docs, changelogs and repos. Three things changed. Greptile's Pro plan dropped seat fees and now charges $1 per credit. OpenCode now runs as a desktop app and an IDE extension, not just in the terminal. PocketBase's code moved off GitHub, and it's still MIT licensed. Replit's pricing page couldn't be reached this week, so its price stays as last confirmed on Oct 21.
A reminder: Smithery shuts down its MCP Proxy and managed identity services on October 31. Smithery says it will stay up as a public MCP registry, so it stays our pick for finding MCP servers for now.
Every change is on our public What changed log, with the before and after wording and the sources.
What changed · MCP servers & agent skills
This week's verdicts
GitHub Copilot Fusion is out: pick an effort level and Copilot picks the models
Instead of choosing a model, you choose Copilot Fusion and an effort level (Standard, High or Max), and Copilot decides which models handle each prompt. Sometimes one model drafts and another reviews. It's on all Copilot plans in VS Code, the Copilot app, Copilot CLI and JetBrains IDEs. You pay each model's standard rate, with 30% off on paid plans for a limited time. GitHub says Fusion Max scored within two points of Claude Code with Opus 5.5 Max on the Artificial Analysis Coding Agent Index, at about half the average API cost per task. A separate experimental preview can send some requests to a model on your own machine, though the routing decision still happens in the cloud.
Source · Local models previewMistral releases the weights for Mistral Large 4
Mistral's 1-trillion-parameter flagship is now open weight, and Mistral says it's the best open-weight model for coding. The license is a modified MIT license: free to use, change and sell, except that companies running a model-as-a-service business need a separate license. This isn't a laptop model. Hugging Face's ready-made deployment uses eight H200 GPUs. For most builders the news is the API price: $0.70 per million input tokens and $2.10 per million output tokens, also available on OpenRouter.
Source · Hugging FaceOpenAI launches GPT-6.1 Terra, a faster and cheaper model for computer use
Terra is built for visual reasoning, using desktop apps and clear writing. In the API it costs $1 per million input tokens and $5 per million output tokens, which OpenAI says is half the price of GPT-5.6 Terra and GPT-6.1 Sol. It's available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and OpenAI says the same plan allowance covers more tasks with it. OpenAI still recommends Sol or GPT-6 Astra for the hardest coding work. Codex CLI 0.165.0 added Terra to its model catalog on Oct 27.
Source · Codex CLI 0.165.0Supabase Compute: run agents and background jobs inside your Supabase project
This is for work that doesn't fit an Edge Function: jobs that need more memory, run longer than 400 seconds, or use another language. You deploy Node, Deno or any Dockerfile with the Supabase CLI (v2.120.0 or later) and pick 2 GB with 1 vCPU or 4 GB with 2 vCPUs. It's a free private alpha with a waitlist, only for projects in us-west-2, and not meant for production. Every service gets all of your project's secrets, so keep that in mind before you put an agent there.
SourceRailway users: rotate credentials stored before Oct 21
Railway says an attacker got into some of its internal systems between Oct 20 and Oct 22, and some customers' encrypted secrets were exposed. It recommends that every customer revoke and replace sensitive credentials stored on Railway before October 21, remove public network access to databases, and review GitHub and container registry audit logs. If OpenAI or Anthropic revoked a key you had already replaced, Railway says that was precautionary: make a new key, update the variable and redeploy. New signups are still paused, and Railway hasn't given a reopening date.
Source · Incident report · StatusChroma is joining Replit, and Chroma Cloud is winding down
Chroma, the open-source vector database a lot of AI apps use for search and memory, signed an agreement to join Replit. ChromaDB stays open source under Apache 2.0. Chroma Cloud, the hosted service, is winding down over the next 60 days. If your app stores data in Chroma Cloud, read its wind-down timeline and plan your move now.
SourceLovable moved apps off the retired Gemini 2.5 models
Gemini 2.5 Pro, Flash and Flash Lite retired on October 20 for your app's AI features. Requests for Pro and Flash now go to Gemini 3.8 Flash, and Flash Lite requests go to Gemini 3.1 Flash Lite. Your app keeps working without a code change, but answers can come out differently and cost a different amount per request. If your app used one of them, test its AI features and ask Lovable to switch to a current model on purpose.
SourceVercel Agent sessions now cost 67% less on average
Vercel rebuilt how its built-in Agent (in the dashboard and Slack) manages context, so each session uses fewer tokens and reuses more from the prompt cache. The billing rates didn't change: model inference rates plus a $0.25 per million token Vercel fee. There's nothing to turn on. If you tried Vercel Agent before and found it pricey, it may be worth another look.
SourceIf you already pay for Copilot, this is the easiest way to stop guessing which model to use for each task. Fusion Standard favors versatile models for everyday coding, and paid plans get 30% off Fusion model costs for a limited time. Try it on your normal work this week and check your usage before you move your harder tasks to High or Max.
SourceThe weekly verdict
New AI tools for builders, and which picks changed, once a week. We will email you a confirmation link first. Unsubscribe anytime.