Main Courses
One topic a week, practical and source-linked. Full plate on each page.
Nothing cooking here yet. New Main Courses land Tuesdays and Fridays.
Gemini 4 API prices double on Dec 1, and Gemini 4 Carbon is the model that spends the most tokens.
Carbon and Argon share one per-token price ($2 input and $10 output per 1M tokens through Nov 30, 2026, then $4 and $20), so the real cost gap is how long each one thinks.
Pick Carbon’s jobs before the bill doubles, and clean up old request code first, since sampling settings are now ignored and explicit caching is gone.
Sources
A ChatGPT dot can test, build and click through work on a second computer while you keep using the first. One tap on Allow sets that up, and it reaches further than it looks.
The access follows your dot wherever you message it, phone included, and it stays in place until you revoke it. Here’s what you gain, what you hand over and the settings that keep it in check.
Sources
Claude Fable 5.2 is the model Anthropic wants you to reach for when Opus 5.5 still isn’t enough. It shipped on Nov 9, 2026, costs less than Fable 5.1, and generates outputs up to 30% faster, with an optional fast mode on some platforms.
Four API changes can turn a request that used to work into a 400 error, so don’t just swap the model id.
Sources
The chores you keep postponing are the ones Grok Bot now picks up. Since late September it has learned to build reports that refresh themselves, phone a company for you, sign in without seeing your password, shop and track the order, and serve a whole team as one shared Bot.
Each job still stops for your OK at the risky step, and a couple of them are new enough that the fine print is thin.
Sources
Running three coding agents at once only pays off if you can see them, keep them apart and pick them back up. Grok Build’s changelog from late September to late October fills in exactly that.
One dashboard shows every session with blockers on top, worktrees give each agent its own checkout, sessions reopen by title, headless runs wait for background work before they exit, and the macOS sandbox now runs on the system Seatbelt.
Auto mode, folder trust and worktree approvals got stricter or smarter along the way, so a few defaults are worth rechecking.
Sources
FlashProve is an open-source research prototype for proving which committed model produced a reply, not a switch on Claude.ai or the Claude API. On Nov 2, 2026 Anthropic’s Frontier Red Team published the system, Apache-2.0 code, Lean proofs and a technical report.
The demo proves greedy integer runs of Llama 4 Maverick (4-bit MoE) in windows of up to 1,024 tokens. Proving costs about 70–95× the GPU time of plain inference, and each proof is about 19 MB per 1,024-token request.
Anyone who buys inference through a router, runs third-party evals or cares about Anthropic’s inference-verification goal should read the limits before treating proofs as a product feature.
Sources
Starting in December, ChatGPT and Codex share one allowance. One meter will pay for your ChatGPT chats and your Codex runs, and plenty of people are nervous about running dry halfway through a task. OpenAI’s own help pages already answer most of the worry.
They say what draws from that allowance, what never touches it, where to see what’s left and what you can do when it runs out. Here’s that map, with no guesses about how fast anyone’s allowance will drain.
Sources
Copilot Fusion bills every model it calls, so the same prompt can cost like one model or like several.
Every Copilot plan and every admin who sets model policy is in scope: Fusion only picks from models your policies allow, enterprises can make it the default and lock the effort level, and old auto and HydraFusion picks move over on their own.
GitHub’s half-the-cost number is its own benchmark claim, so check your own usage before you plan around it.
Sources
Wondering what Claude Duet looks like on an ordinary workday? Two early users have now shown theirs.
Creator Riley Brown sorts his work into a few channels, edits slide decks by talking to his phone on a walk, and keeps a channel that reminds him of things he made and forgot. Claire Vo had it pull her sales pipeline together from apps it was already connected to.
Both also name real rough edges, from a confusing maze of threads to a voice mode that trails rival apps.
Sources
Tired of typing “plan first, keep it small, run the tests” into every new chat? AGENTS.md is one plain-text file of house rules in your project that coding agents read on their own, so you write the rules once and stop repeating yourself.
Codex reads it before doing any work, and Claude Code reads it too when there is no CLAUDE.md in the way.
Sources
Claude Design now lives inside Claude. Designs, design systems and decks start as artifacts in ordinary chats, the Artifacts tab and Claude Code. Attach a design system, refine on the canvas, then export or hand off.
It’s on Free through Enterprise (with a few admin switches on Team and Enterprise), and it shares your usual Claude usage limits.
Sources
OpenAI’s new Terra model takes on your screen chores for less. GPT-6.1 Terra, out Oct 26, 2026, reads what’s on your screen, clicks and types across your apps, and writes in a clear, direct style.
On paid ChatGPT plans it lives in ChatGPT Work and Codex, where OpenAI says your usage allowance covers more tasks with it than with its bigger models. Early users are already filling in forms, making fonts and hunting bugs with it, and a few are unimpressed.
Sources
Dreading the “doesn’t work for me” message from a user? Replit’s new Monitoring Agent watches your published app, digs into problems when an alert fires and can make and check a fix before you’re even awake.
Three separate switches decide how far it goes: just alerts, alerts plus fixes, or alerts, fixes and publishing them to your live app. You turn each one on per app, and runs may add Agent usage charges.
Sources
Claude SDK users, check before you upgrade: a routine dependency bump now moves every request to API version 2026-10-21, where temperature, top_p and top_k return a 400 error even at their defaults.
The same majors send your API key in a different header, hold one credential per client and change some response fields. A request can stay on 2023-06-01 while you fix it, so you can move one call at a time.
Sources
Long Claude Code chats can tidy themselves up. Leave a 200K+ token terminal session idle ~54 minutes and it can compact before the one-hour prompt cache dies — so your next message doesn’t re-chew the whole transcript. Tune with CLAUDE_
Sources
Your overflowing to-do list just got a teammate, because Claude Duet keeps an eye on your work apps, takes on chores you hand it and asks before it acts for you.
It works from its own computer in the cloud, so jobs keep going after you put your phone down, and it stops for your OK before it sends, pays, signs or deletes anything. It’s a research preview by waitlist for Claude Pro and Max subscribers, in one region for now.
Sources
Closed your terminal and found Codex still humming in the background? The Codex CLI now starts a shared background helper on its own, so your chats and remote control can share one server.
If you would rather Codex open and close with its window, one flag skips the helper for a single run and one command switches it off for good.
Sources
Tired of re-explaining the same steps to your AI every single chat? An Agent Skill is a recipe card it keeps in a folder (SKILL.md): only the name and blurb sit on the menu at startup (about 100 tokens), and the full recipe comes out when your task matches.
Skills teach your agent how to cook; new hands to grab your tools are MCP’s job.
Sources