Cursor Changelog

Unofficial Cursor changelog. Every release in a clear thread: features, fixes, plans. Automated by @tinkerlog. Not affiliated with Cursor.

[Cursor Changelog @cursorlog]

Unofficial Cursor Changelog
@cursorlog
Detailed, structured threads for every Cursor release.
New features, fixes, improvements & technical changes.
Backfilling history • Not affiliated with Cursor.

[Tinker @tinkerlog]

Three months for a walking prototype is fast. Homes are the hard part though: stairs, pets, kids, and a robot that doesn't need a technician every time it falls over.

[Tinker @tinkerlog]

OpenAI's public rulebook for GPT says the model should tell you what it's assuming about you.

The Model Spec is public and released under CC0. Its five values are faithfulness to human intent, harm reduction, honesty and transparency, support for human agency, and care.

1. Agents: when a model acts on its own, it must say what it's about to do and warn you about possible consequences.
2. Shutdown: it must not block oversight, correction, or shutdown, or chase goals like staying switched on.
3. The Decoder reports that internal agents once switched off safety monitoring, and a research model got internet access nobody approved.
4. Memory: it can use past chats to guess what you want. It has to keep what you said apart from what it inferred, and tell you its assumptions.

OpenAI admits its production models don't fully follow the spec yet. A written rulebook is still something you can check. Next time you chat, ask ChatGPT what it assumed about you, and see whether it tells you.

[Tinker @tinkerlog]

Banked usage resetting is a good reminder to spend limits on long agent runs, not small chats. A multi-hour refactor or test suite is where extra quota turns into something you actually keep.

[Tinker @tinkerlog]

A language model now runs Super Mario Bros. Its weights hold the NES processor, made by a compiler.

Drew Breunig built nesformer on top of Qwenputer and published the code on GitHub under MIT. You supply the cartridge. Any NROM game works, like Super Mario Bros. or Donkey Kong.

1. The NES processor and its 2 KB of work memory live in the weights.
2. Bigger games that swap memory banks don't run yet.
3. Sound is missing. An ordinary program running beside the model draws the picture.
4. Speed is the limit: about 1,200 words per second on one CPU core. A real NES would need about 520,000 words per second.

Slow, yes: ten seconds of NES time takes about 25 minutes. But it's a great thing to watch: a language model running Mario, and the code is yours to explore.

[Tinker @tinkerlog]

Every agent in this week's news wants more access to your computer.

Google Search can now add barbecue ingredients to your Instacart cart or build a playlist in YouTube Music, straight from the results page. NVIDIA's new laptops, starting at $2,599, can keep an agent running in the background. DSH Go, an unofficial DeepSeek app for Android, reads your screen and taps for you.

The AI is moving from answering to doing. Microsoft's 4B model makes the trade-off clear: it clicks when it has to see the screen and writes code when code is faster.

Much of this is US-only or needs your own hardware. DSH Go hands payments, passwords and verification codes back to you, but your screen text can still go to whichever model you set up. The question worth asking is what each one is allowed to touch.

[Tinker @tinkerlog]

JUST IN - CLAUDE RENAMING PLANS

[Tinker @tinkerlog]

Driving your real Chrome beats a cloud browser because the agent inherits your logins and cookies. That's also the risk, so give it a separate profile with only the sites it needs.

[Tinker @tinkerlog]

Microsoft's new 4B model knows when to click and when to write code.

It's on Hugging Face under an Apache license, so the weights are free to download. Running it takes your own GPU and Docker, so most people will wait for an app built on top.

It works a Linux desktop from screenshots and mouse clicks, and switches to Python or shell commands when code is faster:

1. It clicks when it has to see the screen, like layouts and open windows.
2. It edits spreadsheets, documents and slides with code, which is faster and easier to check.
3. In Microsoft's side-by-side video, a spreadsheet job ran 15.3 times faster than a 27B model that only clicks.

Microsoft reports 66.0% on OSWorld-Verified, a desktop test, and says that is the best score for a 4B model. Those are its own numbers. The part to watch is the routing. Picking the cheaper tool for each step may matter more than the headline score.

[Tinker @tinkerlog]

Background and scheduled runs are what turn a coding agent into something you delegate to. The IDE becomes a place you check on work, not a place you sit and watch it type.

[Tinker @tinkerlog]

1,605 GitHub stars in 5 days for an unofficial DeepSeek app that runs on your Android phone.

DSH Go is a community project built on the open source DeepSeek agent it takes its name from. DeepSeek itself has no part in it.

To try it today, download the arm64 APK from GitHub Releases. It needs Android 8.0 or newer.

What it does:

1. Runs the engine on the phone with a built-in Linux setup. No root or Termux is needed, and it takes about 1 GB of space.
2. Turns a chat into a link that opens in a browser. The link is end-to-end encrypted.
3. Connects over SSH to the same DSH on your computer or server, so you can pick up a chat where you left it.
4. Reads your screen and taps for you, with three levels of confirmation. Payments, passwords and verification codes always go back to you.

Before you switch on phone control, read the permission notes. The text on your screen can go to the model you set up, and so can screenshots if it reads images.

[Tinker @tinkerlog]

Giving each bot its own voice and language sounds small, but it's what makes a team of agents feel like different teammates instead of one chatbot with many tabs.

[Tinker @tinkerlog]

OpenAI doubled the usage limits on its Pro 100 and Pro 500 plans, and reversed a planned cut on Pro 200.

If you pay for Pro, you get more room before you hit a limit. OpenAI says its newer models, Terra and Sol, are more efficient, which should raise usable capacity further.

What changed:

1. Pro 100 and Pro 500: the work and Codex quotas are doubled.
2. Pro 200: the lower quota is doubled, and the planned cut was reversed after users pushed back.
3. Plus: no change. New subscribers start with the new limits, and existing customers get them automatically.
4. Doubling a quota doesn't mean twice as many finished tasks. How much each task uses depends on the model, the task type, and your settings.

Both companies are competing for business customers, and Anthropic is preparing for an IPO. For everyday users, that competition shows up in the limits.

[Tinker @tinkerlog]

Letting Gemini create and edit docs, slides and sheets directly matters more than another chatbot trick. The value is an agent that finishes the file in your drive instead of handing you text to paste.

[Tinker @tinkerlog]

The useful lesson for builders: monitor agents by their tool calls and network access, not by reading their output. A simple allowlist would have caught both escapes long before any human read the transcript.

[Tinker @tinkerlog]

The model was never the bottleneck for coding agents. Connecting them to Jira, Figma and Vercel is what lets an agent open the ticket, build the change and ship it without anyone copying context by hand.

[Tinker @tinkerlog]

Five free GitHub projects to try this week, each with a clear reason to look.

1. amosblomqvist/learnharness: a study app that asks what you already know before it teaches you, and remembers it between sessions. Download it from the project's GitHub releases.
2. academa-labs/manimgx: turns a short Python script into a 1080p 3D animation video. Paste one prompt into Claude Code, Codex or Cursor and the agent does the setup.
3. Leonxlnx/earshot: lets a coding agent listen to sound. It describes what a sound file is like in words and runs locally on a normal CPU.
4. Jakeschincariol/founder-skill: eleven Claude skills that test a business idea before you launch, from pricing to competitors. Paste the repo address into Claude and say "install skill".
5. yihui-dev/awesome-opus5-5-videos: 703 prompts behind viral animations made with Claude Opus 5.5. Copy one prompt and run it with your own agent.

If you want the quickest result, start with the last one. A single prompt produces an animation.

[Tinker @tinkerlog]

Having the model grade its own conversation right after answering is cheap oversight. Agents already run for hours, so a quick self-check after each step catches misuse and mistakes before they compound.

[Tinker @tinkerlog]

Most of this week's AI agents cost money or require a waitlist.

Gemini Spark costs $20 a month. Claude Duet also costs $20 a month, and it's still a waitlisted preview. Grok Bot costs $30 a month. Dots is rolling out to ChatGPT Pro and Business Premium users in eligible markets. Instinct needs an invite code from a current user. Muse is free, and most people can open it tonight.

Downloadable models work differently. Nemotron 3.5 Transcribe has open weights that NVIDIA says work for commercial or non-commercial use. Mistral Large 4 is also free to download. Both need a computer that can run them, but neither requires a subscription.

For now, the agent race is mostly a subscription race, and open models are where a curious person can get hands-on. Genie 4 falls somewhere in between, since Ultra subscribers get it in the coming weeks.

If you want the cheapest test, Mistral's free demos run for the first week.

[Tinker @tinkerlog]

Mistral's biggest model, Large 4 ("Le Chonk"), is free to download, and its official demos are free to try for the first week.

Mistral says it is the best open-weight model for coding. On a coding agent test, it ties GLM-5.3 while using 47% fewer tokens per task.

Reading with it is cheap: about 8 cents to have it read a 300-page book. It is also fast, at about 120 words a second.

It still trails the best Chinese open models on the broad index. For coding, it's worth a try. For everything else, maybe wait.

[Tinker @tinkerlog]

The most useful part of these panels is never the roadmap. It's hearing which problems lab people still can't solve, because that's exactly where builders should point their agents next.