Good Morning! Here's what I have for you in today's newsletter:
OpenAI adds multi-account support across all ChatGPT plugins
Meta puts SAM 3.1 on Model API, detect, segment, and track in one call
Qwen-Image-2.1 lands in ComfyUI, 7B open model with 2K and RGBA output
Learn the five habits that protect your Claude Code tokens every session.
4 new AI tools worth trying today
AI TOOLS
OpenAI enabled multi-account support across most ChatGPT plugins, so a user can connect personal, work, and side-project accounts for the same service and use them inside one conversation.
Extra accounts are linked from the plugin directory, and ChatGPT picks the right account for each request automatically, respecting that account's permissions, so no switching or logging out is needed.
Multi-account support applies to existing plugins without any developer changes since ChatGPT handles the account linking, not the plugin itself.
Developers can add a profile tool to their MCP server so each linked account appears with a clear label, such as "Acme workspace" instead of a generic entry.
If your work and personal tools live in separate accounts of the same app, ChatGPT can pull from both in one thread, pick the right one for each task, and keep their permissions apart. Check the plugin directory for the ones you use, since most already support it.
AI MODELS
Meta released SAM 3.1 on Meta Model API, serving its perception model that detects, segments, and tracks objects in images and video from a short text phrase in a single call.
One request returns bounding boxes, pixel-precise segmentation masks, and identity-preserving video tracks, with inference tuned for SAM 3.1's DETR architecture and nothing to host or configure.
Object Multiplex tracks up to 16 targets in one pass, and the model shares keys, docs, and billing with Muse Spark and Muse Voice Transcribe on the same API.
Pricing is $2.50 per 1,000 images and $0.20 per 1,000 video frames, the API accepts calls through the OpenAI SDK, and open weights remain available for visual prompts.
Any app can add object detection, background removal, or video tracking by sending an image or clip plus a text phrase like "red bicycle" to Meta's API. There is no model to download, no GPU server to run, and the price is fixed per image or per frame, so a developer can test it for a few dollars and ship it the same day.
AI MODELS
ComfyUI added day-0 support for Qwen-Image-2.1, Alibaba's open-sourced generation and editing model whose visual component runs on 7B parameters, shipping with three official workflow templates.
Native output reaches 2048 by 2048 at 1:1 and 2752 by 1536 at 16:9, generated at that size directly, with improved typography, portrait lighting, and fine detail.
Editing accepts up to 10 reference images in one pass, with local edits marked by circles, painted annotations, or a mask, and identity preserved for people and products.
The model writes RGBA output with alpha, so stickers, logos, and product cutouts need no background removal, and Comfy-Org's INT8 repackage fits the whole pipeline on a 4090 or 5090.
A 7B image model that generates at 2K, edits from ten references, and outputs transparency in one checkpoint is a full asset pipeline on a home GPU. Sprites, logos, and product shots come out ready to composite. Read the license before shipping anything, since this release is research-only and commercial work needs a paid agreement.
HOW TO AI
Learn the five habits that protect your budget every single session.
Claude Code doesn't need more context, it needs better context. These five habits stop it from wasting tokens on things that never needed to happen.
Step 1: Find the code directly
Without direction, Claude can scan an entire codebase before finding the one file that matters. Point it there first.
Use semantic search to find the file handling [the specific feature], don't scan the full codebase first
Step 2: Filter the noise
Command output and web pages both dump far more into context than any task needs.
Run npm install, but only show me the final result, filter out warnings, progress output, and notices
Fetch this URL, extract only the title, the relevant section, and any code examples, don't paste the raw page
Step 3: Route by size
Simple work doesn't need the most expensive model available.
# 3-Tier Model Routing
Haiku → locate code, grep, diff review
Sonnet → iterative development
Opus → architecture decisions only
NEVER use Opus for:
- File lookup
- Simple function questions
- Basic diff reviewStep 4: Stop repeating itself
Every new session starts from zero, and files already read once can get read again for no reason.
Index my previous project conversations so relevant context carries into new sessions
Don't re-read files already loaded this session, reuse what's already in context unless I say the file changed
Step 5: Keep the output short
Long preambles and repeated explanations cost tokens without adding anything useful.
Answer directly, skip the preamble, don't repeat what I asked, keep explanations minimal unless I ask for more
P.S. You can access all the AI courses and trainings (including the full version of this one), prompts and workflows if you upgrade.
Prism ML released Ternary Bonsai 2 27B under Apache 2.0, built on Qwen3.8 27B and 9x smaller at a 5.9 GB footprint while keeping 98.2 percent of benchmark performance, with strong gains in agentic coding and tool use.
Notion showcased a field-sales workflow at Perfected Energy, where reps send Claude a voice note after each store visit covering what sold and what was restocked, and Claude updates the right account in Notion through MCP before the next shop.
Perplexity published a Search SDK cookbook that fans out focused searches, filters results to official documentation, extracts the relevant passages, and writes a source-linked brief a coding agent can read before touching the codebase.
📣 Fastlane: Jev for Marketing takes a website URL and generates social accounts and short promo videos for the product in seconds.
🎬 Motion: added a Muse workflow, where Muse researches a website and writes the launch video prompt, then Motion builds the scenes from it.
🌵 Cactus Compute: Needle 3 is an 8-29MB tool-calling model that runs at up to 4k tokens per second on a Raspberry Pi 5, with a browser demo.
🧩 Cline: added Union Alpha as a free model, 256K context and multimodal, landing near GPT-6 Astra and Opus 5 on agentic coding at 18x lower cost.
THAT’S IT FOR TODAY
Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!
See you tomorrow :)
- Dr. Alvaro Cintas
✓ Full archive of premium guides with ready-to-use prompts
✓ Structured AI courses (step-by-step, start-to-finish)
✓ Every upcoming premium tutorial