Simon Willison’s Weblog

Newsletter
Atom feed

Blogmarks

Filters: Sorted by date

External sandboxes for coding agents. I've been exploring new ways to sandbox code that is executed by agents, with a focus on making it safer to try out local models without worrying that a mistake or prompt injection attack might cause them to take damaging actions on a machine.

I worked on this project with Prime Radiant, the AI research lab behind Superpowers and other projects.

My post on the Prime Radiant blog introduces the external sandbox pattern, where code executed by the agent runs inside a dedicated sandbox while the agent itself lives outside of that sandbox.

To demonstrate this pattern we've released three sandbox server implementations:

Each of these implementations offers the same localhost API. You can then configure Pi, Claude Code, or Codex to send API calls to that sandbox any time they want to execute code.

The sandboxes also include an optional HTTP/HTTPS allow-list proxy, so you can control exactly which hosts your agent's code is allowed to communicate with - plus a credential replacement mechanism to make authenticated API calls without sharing a secret with the agent. Here's the UI you can use to reconfigure that proxy while the sandbox is still running:

The network policy page with allowed destinations, the policy JSON editor and recent proxy activity

I continue to be fascinated by the challenge of safely running untrusted code. This external sandbox pattern provides a neat way to experiment with different approaches.

# 17th November 2026, 10:52 pm / sandboxing, security, ai, webassembly, generative-ai, llms, coding-agents, claude-code, codex, pi

GPT‑6.2 Astra and GPT‑6.2 Sol. OpenAI:

We’re introducing GPT‑6.2 Astra, our new frontier model for your hardest projects, and GPT‑6.2 Sol, our efficient, versatile model for everyday work.

They're available in the API as gpt-6.2-astra and gpt-6.2-sol. I shipped LLM 0.38 with support for the new models.

6.2 Astra is 80% of the price of 6 Astra - $8/MTok input and $40/MTok output, the same price as Claude Fable 5.2 (which was also 80% of its predecessor). 6 Astra and Fable 5 were both $10/$50. The cached price is an even bigger discount - $0.40/million compared to $1/million for the older model. This is great news for longer running agentic sessions.

GPT-6.2 Sol is the same $2/$10 price as 6.1 Sol and 6 Sol - which are half the price of GPT-5.6 Sol ($4/$20).

Here are the pelicans for GPT-6.2 Sol and for GPT-6.2 Astra. The 6.2 Astra max pelican looks like this:

Good bicycle, good pelican, one wing is on the handlebars with fingers, the other wing is behind. Feet on the correct side of the frame. Nice gradiants. Good pouch, hooked bill.

I have a full comparison grid here, but here's how the high/xhigh/max pelicans compare with the previous Sol and Astra releases:

Twelve pastel illustrations compare four models drawing a pelican riding a bicycle at three effort levels. All twelve have a recognizable pelican and bicycle. The clearest differences are in the feather detail and how convincingly the bird steers and pedals. gpt-6.2-astra has the most detailed Max drawing (88.71¢), with layered feathers, a trailing red scarf, and a wing on the handlebars. Xhigh (17.70¢) adds a green cap and keeps the seaside setting. High (11.55¢) simplifies the feathers while keeping the riding pose clear. gpt-6.2-sol shows a convincing pedaling pose at both Max (13.69¢) and Xhigh (6.74¢), with one foot higher than the other and a wing reaching forward. At High (2.41¢), the bird sits lower and the handlebars are harder to distinguish. gpt-6-astra gives the bird a long wing that follows the curve of the handlebars at Max (63.21¢). Xhigh (33.85¢) shows two bent legs on the pedals. At High (18.37¢), one leg stretches almost straight down, making the pose look stiffer. gpt-6.1-sol uses nearly the same long-necked bird with a wing on the bars at Max (13.48¢) and Xhigh (5.39¢). At High (3.66¢), the bird folds its wing against its body and appears to ride hands-free.

I also ran my most recent benchmark prompt against the two models, using "Ultra' effort levels in the Codex CLI:

Make a three js animation where yonkers stands up as a big mech (cool stand up motion) and stomps around manhattan and queens and Brooklyn, use openstreetmap mapping data

Here's Sol and here's Astra. Neither of them did a good enough job of morphing Yonkers into a big mech, at least in comparison to Claude Fable 5.2 and Gemini 4 Carbon.

This video showing the Astra and Sol animations playing out at the same time:

I created that video using GPT-6.2 Astra (Ultra) in the Codex desktop app by prompting:

https://static.simonwillison.net/static/2026/yonkers-gpt-6.2-astra-ultra/index.html and https://static.simonwillison.net/static/2026/yonkers-gpt-6.2-sol-ultra/index.html  are two videos of the same thing by two different models

I need a video of the first 10 seconds where the top half of the video plays the first and the bottom half plays the second

I need an overlaid label of "GPT-6.2 Astra" on the top half and "GPT-6.2 Sol" on the bottom half. These labels should be tastefully positioned.

Make the video square and 2000x2000 pixels

Then later:

OK make it 20s

Here's the exported transcript.

The buzz around these new models is mixed at the moment. Astra 6.2 isn't scoring quite as high as Fable 5.2 on various benchmarks, and I've seen a few complaints that its communication style can be confusing. Theo Browne's review (in the form of a feedback letter he sent OpenAI during pre-release) has this useful example:

When this model is working on long tasks, it can pull in a lot of subagents and manage them pretty well. It will then proceed to randomly ask you some weird internal implementation detail bullshit that you could not under any earthly circumstance understand, much less have input on.

I sent "/goal make a new level for Outer Wilds". 2 hours in, it's "blocked" because it wanted to ask me "does r24's reviewer write it's clips to 24_g521.mp4 or 24_p11.mp4?"

# 17th November 2026, 10:38 pm / ai, openai, generative-ai, llms, llm-pricing, pelican-riding-a-bicycle, llm-release, theo-browne, gpt, yonkers-rising

Gemini Decisions API (via) Google join Jev and OpenAI in the decision model category with "Gemini Decision 1". This new model API is almost identical to Jev: it supports the same three types of questions (yes/no, choice, and rubric-scoring) and even reuses Jev's term "Noul" for a yes/no question.

It's priced the same as the GPT-6 Luna decision model: 10 cents per million input tokens, and output is free. Jev remains the cheapest of the three, at 4.2 cents per million.

Jev is text (and JSON) only. OpenAI can handle text and images. Gemini Decision 1 is fully multi-model: it covers text, images, audio, video, and PDFs.

I tested it against video by passing these questions to it along with my Yonkers stomping robot video file:

dragon: Does this video have a dragon in it?
robot: Does this video have a robot in it?
france: Is this video set in France?
usa: Is this video set in the USA?

Jev and OpenAI both return 1.0 or 0.0 for yes/no questions, but Gemini insists on being a little less confident than that:

{
  "dragon": {
    "type": "yes_no",
    "probability": 0.04976021841520312
  },
  "robot": {
    "type": "yes_no",
    "probability": 0.9234055387792485
  },
  "france": {
    "type": "yes_no",
    "probability": 0.08329171353166241
  },
  "usa": {
    "type": "yes_no",
    "probability": 0.9230459995157706
  }
}

# 11th November 2026, 10:03 pm / ai, generative-ai, gemini, llm-release, jev, decision-models

aws-emulator. Finally, an official emulator from AWS to run tests against a local clone of their many different services without having to provision a paid API key to do so.

It's written in Rust by a huge swarm of agents - 29,684 commits in the first three weeks of the project. Swaminathan (Swami) Sivasubramanian wrote about how it came to be in a guest post on Werner Vogels' blog.

It currently requires Docker to run on macOS or Windows. I'm hoping they fix that in the future - it would be great if you could download a single binary and use that as part of a standard test suite in other languages and on other platforms.

# 11th November 2026, 9:18 pm / amazon, amazon-web-services, testing, werner-vogels, ai, rust, docker, generative-ai, llms, agentic-engineering

Browse tools.simonwillison.net by tag. I've been wanting a better way to browse my collection of vibe-coded tools (244 and counting) for a while. I had a few Fable 5.2 credits left over before my weekly reset so I set it to work using tree-sitter to analyze the code for JavaScript APIs, libraries, and web API hosts used by each tool.

The homepage now features tag clouds for browsing by each of those categories of tag:

Screenshot of two tag cloud sections where tags are sized by count. Browse by JavaScript libraries - Libraries loaded from a CDN by two or more tools. Select a tag to list its tools. @google/generative-ai 3, @jsquash/avif 3, @jsquash/webp 4, Babel 2, Chart.js 2, Cropper.js 2, D3 2, DOMPurify 7, js-yaml 3, JSZip 3, Leaflet 4, markdown-it 4, Marked 12, pako 2, PDF.js 3, Pyodide 11, React 3, Tesseract.js 2, three.js 3. Browse by browser APIs - Web platform features used by each tool's own code. Select a tag to list its tools. <dialog> 5, AbortController 6, Blob URLs 38, BroadcastChannel 1, Built-in AI 1, Camera / microphone 9, Canvas 2D 44, Clipboard API 69, Compression Streams 3, createImageBitmap 2, crypto random values 4, DOMParser 8, Drag and drop 30, Dynamic import() 12, ES modules 92

# 11th November 2026, 6:42 pm / javascript, tools, ai, generative-ai, llms, claude, vibe-coding

Anthropic’s IPO roadshow video. Anthropic kicked off their IPO roadshow yesterday, and their roadshow video has some previously unreleased information about how their business works.

The section on "Approach to compute" with co-founder Tom Brown is particularly interesting, describing how Anthropic train and serve Claude across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs.

I'm not sure why there's no official video transcript, but here's one that I extracted using MacWhisper. Highlights below:

We train and serve Claude now across all four major chip families. That's only possible because early on we made an unconventional bet on a multi-chip strategy. At the time, many people thought that running AI models on different chips wasn't worth the effort. Each chip needs its own tailored software stack and inference stack, which is expensive and difficult to build. We've now been doing it for years, and we've gotten incredibly good at it. [...]

As an example, this year, we started working with AMD GPUs for the first time. In the past, bringing on a new chip platform took dozens of engineers months of work This time, we had Mythos try it out. It got Claude running on AMD in just three days, with basically no human intervention. Claude ported or wrote almost 100% of the kernels. [...]

The second driver of our compute advantage is our rate of discovery of compute multipliers. We focus on total effective compute, which is the multiplicative product of physical capacity times the compute multipliers that we discover. Each one we discover makes our pool of total effective compute bigger. For example, say you need 100 units of compute to train a model of a certain intelligence. A 4x compute multiplier means you can use 25 units of compute to train the exact same model. Or you can use those hundred physical units to train an even more intelligent model that would have previously required 400 units.

Compute multipliers create a competitive advantage for us that's hard to replicate. They build on each other. Each new multiplier we discover has a positive effect on the multipliers before it. They stack. We believe that this can accelerate even further as we approach recursive self-improvement. Claude will be able to discover compute multipliers entirely by itself, allowing the number of compute multipliers to grow faster and faster.

We got a hint at Anthropic's multi-chip strategy last September in their postmortem of three recent issues, where they described bugs that affected just their TPU deployments.

# 11th November 2026, 3:11 pm / ai, generative-ai, llms, anthropic, gpus, claude-mythos-fable, recursive-self-improvement

Introducing Our Smallest, Most Efficient Model: Muse Ember. Meta AI just launched their entry in the Claude Haiku/GPT Luna category: Muse Ember, which they describe as:

[...] our smallest and most efficient model designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.

For shorter prompts it's priced the same as GPT-6 Luna and Claude Haiku 5.5 - both are $0.10/million input tokens and $0.50/million output. Today's other big model release, Claude Fable 5.2, is 80x that price at $8/$40.

The pricing is a little less straightforward as prompts get longer, though. Luna, Ember, and Haiku all support up to a million input tokens, but they also all have a threshold after which the price goes up. Taking that threshold into account Ember is priced between the two others:

Model Prompt size Input ($/M) Cached input Output
GPT-6 Luna ≤272k $0.10 $0.01 $0.50
>272k $0.20 $0.02 $0.75
Muse Ember 1.0 ≤200k $0.10 $0.004 $0.50
>200k $0.30 $0.012 $1.50
Claude Haiku 5.5 ≤100k $0.10 $0.01 $0.50
>100k $0.50 $0.05 $2.50

But comparing price-per-token doesn't really make sense these days, when reasoning tokens mean that different models might use dramatically different numbers of tokens to solve the same problem.

For that, one measure that's useful is the Artificial Analysis.

Stacked bar chart from Artificial Analysis comparing a cost in dollars across six models, each bar split into four unlabeled stacked segments, sorted from lowest to highest total. Claude Haiku 5.5 (low): $0.04* total, bottom segment $0.02, three thin unlabeled segments above. GPT-6 Luna (non-reasoning): $0.07 total, bottom segment $0.06, thin unlabeled segments above. GPT-6 Luna (xhigh): $0.10 total, segments from bottom $0.07, $0.02, $0.02, plus a thin unlabeled top segment. Muse Ember (xhigh): $0.12 total, segments from bottom $0.06, $0.02, $0.02, plus a thin unlabeled top segment. Muse Ember (max): $0.14 total, segments from bottom $0.07, $0.03, $0.03, $0.02. Claude Haiku 5.5 (xhigh): $0.25* total, segments from bottom $0.14, $0.04, $0.05, $0.02. Anthropic bars are terracotta, OpenAI bars are dark gray, Meta bars are blue, each with the company logo beneath.

And even more useful than that is their Intelligence Index vs. Cost per Task chart, which shows the following for those three models:

Scatter plot from Artificial Analysis titled "Intelligence Index vs. Cost per Task", plotting Artificial Analysis Intelligence Index (y-axis, 5 to 40) against cost per task in USD on a log scale (x-axis, $0.04 to $0.3), with a green "most attractive quadrant" in the upper left and a dotted Pareto line. Claude Haiku 5.5 (low)* scores about 19 at roughly $0.044. GPT-6 Luna (non-reasoning) scores about 9.5 at roughly $0.07. GPT-6 Luna (xhigh) scores about 23.5 at roughly $0.10. Muse Ember (xhigh) scores about 34 at roughly $0.12 and Muse Ember (max) about 36 at roughly $0.14, both on the Pareto line. Claude Haiku 5.5 (xhigh)* scores about 34.5 at roughly $0.25, well to the right of the Pareto line. Three unlabeled gray dots also sit on the Pareto line.

The Muse Ember models sit right on that Pareto line - the line at which you can only improve one dimension by giving up another. Meta call this out in their announcement:

It also sits on the efficiency frontier of the full Artificial Analysis leaderboard: no model scores higher for less per task.

(That grey dot on the Pareto line in the middle of the green area is GPT-6.1 Terra (low) - clearly I should be paying more attention to the Terra family!)

Muse Ember has six reasoning levels - minimal, low, medium, high, xhigh, and max - so I rendered six pelicans riding bicycles. The "high" reasoning level failed to return a full SVG, and the other SVGs were surprising in that the reasoning level didn't cleanly map to the visual quality of the response.

Here's the "max" one, but the others are worth a look too:

Bicycle frame is OK, basket has fish, pelican has a weird blue hat on.

The Ember pelicans are comparable in quality to those I got from Muse Wonder, which is 40x the price of Muse Ember.

# 10th November 2026, 7:28 am / ai, generative-ai, llms, meta, llm-pricing, llm-release, artificial-analysis, muse

datasette #3046: Don’t serve link: rel=next HTTP headers to Firefox for HTML pages. I was looking at the Firefox network panel today when I noticed that requests to a Datasette table page were triggering an HTTP prefetch to the next page of rows in that table.

This was happening because Datasette pages serve an HTTP link header like this:

link: <https://latest.datasette.io/fixtures/compound_three_primary_keys?_next=a%2Cd%2Cv>; rel="next"

This is bad. A tiny fraction of the visitors to one of those pages will click the "next" link, which means for everyone else their browser is fetching unnecessary data and the Datasette server is having to render an entire page that then gets thrown away.

I dug in and learned that Firefox has an issue that has been open for 24 years about this behavior...

And one of the early comments on that issue links to my blog post from 2002 complaining about this feature. I hadn't just forgotten that I had written about it, I had forgotten the feature entirely!

# 7th November 2026, 11:57 pm / firefox, html, mozilla, datasette

I discovered a sound that leaks your exact location (and it’s everywhere)—YouTube (via) Genius piece of OSINT by Colsto, who demonstrates how you can take almost imperceptible background noise of planes from a video soundtrack, figure out their Doppler curves, then cross reference them against ADS-B data to find the exact location of the video based on their speed, distance, and gaps between different flights.

# 3rd November 2026, 5:32 am / privacy, youtube

How We Took Over ChatGPT Accounts (via) Parameter Labs describe an exploit they found against ChatGPT's implementation of the artifacts pattern (OpenAI used to call this "Canvas" but it looks like the latest revision is called HTML previews), where interactive custom HTML apps are displayed in a sandboxed iframe.

It's important that HTML in these stays tightly sandboxed to avoid untrusted code interacting with OpenAI APIs and accessing private data.

Parameter found quite a complex attack chain. OpenAI use the MathLive JavaScript library as part of their graph support, and that library includes its own virtual keyboard, which can be triggered in the parent page via postMessage() from the sandboxed iframe.

That message receiving code uses innerHTML, so anything that sends a message can add custom HTML to the parent - and there was no mechanism to prevent other sandboxed frames, including HTML preview frames, from calling that method.

This meant untrusted code in an iframe could insert HTML, and while innerHTML doesn't directly execute script elements it could be used to inject a new iframe without sandbox restrictions.

OpenAI’s CSP rules blocked inline scripts entirely, and only allowed script loading from a handful of domains, but one of those was YouTube - and the YouTube oEmbed endpoint could be used to reflect JavaScript back in a way that executed it.

Once the attack had JavaScript execution on ChatGPT, it was trivial to fetch private user data from APIs and exfiltrate it via an image reference.

The end result: you could create a ChatGPT share URL which, when visited by another signed-in user, could exfiltrate their private data or take actions on their behalf.

OpenAI closed the hole and paid out a $5,000 bug bounty, which I think was a bit cheap for a vulnerability this serious.

# 31st October 2026, 11:15 pm / iframes, javascript, sandboxing, security, xss, openai, chatgpt, content-security-policy

dbreunig/nesformer (via) Inspired by qwenputer by Divyansh Agrawal, Drew Breunig baked a working emulator for a Nintendo Entertainment System 6502 CPU into the weights of Qwen3.5-4B. You can now play NES ROMs such as Super Mario Bros (very slowly) running inside Qwen!

In a comment on Hacker News Drew reveals the Opus 5.5 prompts he used to build this system:

Read this article and any relevant source code: https://divagr.com/blog/qwenputer. Then explore if you could emulate the chip and ram in a Nintendo Entertainment System (as people do today with FPGAs in MISTRs) using this technique. If so, what size model would you have to build and if you could use smaller faster ones, could you get it to be at the standard clockspeed.

Followed by:

I think we should build it! Just for fun. It's okay if it's not fast. We want to be able to have the weights, and a cli to run them that we can point at a ROM.

Apparently Opus found and then used this collection of conformance tests for the NES CPU released 15 years ago by Christopher Pow.

# 30th October 2026, 9:59 pm / games, ai, generative-ai, llms, drew-breunig, qwen, conformance-suites

API credits benefit for eligible ChatGPT subscriptions. Earlier this month Anthropic announced a generous new benefit for their Max and Team subscriptions: $100, $200, and $500 per month subscribers would get the equivalent quantity of (expiring) credits to use with the Anthropic API.

OpenAI have now launched pretty much exactly the same deal: if you have a $100, $200, or $500 subscription plan with them you now get that same quantity of credits to apply to one of your API organizations.

I maintain subscriptions with both companies, so this is great news for me! I may even ship some public-facing demos now that the credits are already covered by my existing subscriptions.

# 29th October 2026, 10:48 pm / ai, openai, generative-ai, llms, anthropic, llm-pricing

How I Built the Most AI-Pilled Company in the World. Don't be put off by the clickbait headline, this piece by Jacob Sansbury describing how they are building Spawn is a fascinating exploration of the bleeding edge of what's possible with agentic engineering against frontier models.

Spawn is a massively multiplayer gaming environment where players can use prompts (to the in-game agent, Savi) to modify the game world.

I fired it up on my phone for a few minutes and was able to prompt a castle into existence, then a dragon, then ride the dragon to fly around that castle.

Two mobile screenshots side by side from a 3D voxel-style game. Left: a first-person view from the saddle of a small teal dragon, its neck with pale diamond markings and two curved white horns in the foreground, flying toward a dark stone castle with tall teal-gray conical spires, glowing windows and teal pennants, on a rocky island in a lake with green and snowy hills behind. Top bar: profile icon, "keep it", "your den" dropdown, "+ new game" button and a "..." menu. Right side: icon buttons and a portrait of a white-haired, freckled character labeled "tap to talk". Bottom right buttons: "JUMP" and "USE". Right: the same game with the chat panel open over the castle scene. Top bar: "your den" dropdown, "+ new game". Button: "earlier ↑". User message: "Make the dragon smaller and land it next to me again". Replies from the white-haired character: "on it. shrinking it to a pony-sized dragon and landing it right beside you." Then: "it's a little dragon now and you're already up on it over the lake. your view sits right in its saddle, horns ahead. fly back to the house and look down. saw your screen just now: flying over the water, the neck and horns ahead, the house on the far shore" Then: "oh, idea: USE while flying could breathe fire instead, with JUMP held to land. want it?" Input box: "what do you want to bring in?" Buttons below: upload, a small pixel-art icon, "lock in" and a microphone.

By Jacob's telling, Spawn is now being developed autonomously by a swarm of Fable-powered agents costing over $3 million in tokens per month ("we spend 30x payroll on AI tokens").

Our agents change the live code directly and ship to production hourly, so they read every error, bug report and metric themselves. About 60% of our work starts from a production error, a Savi bug report (Savi is the AI you build games with in Spawn), something odd in our logs, or a feature request.

Savi files bugs herself. When a user hits one, or Savi hits one in the product, she files it. It goes straight to an agent that decides whether it's a real bug to fix, something we won't fix, or intentional design that needs a decision. Usually it becomes a change, delivered to that user within an hour, with a message in the same chat.

This is continuous deployment turned up to 11:

A loop this tight only works if it never stops, so we treat an hour without a ship as a SEV-0, our most severe kind of incident. Unshipped changes pile up, and every problem already in production lasts longer. The feedback slows down too. An agent can only keep a few hours of work in its head, so shipping faster than that actually helps it stay coherent. [...]

We can roll back a broken change in 30 seconds, exactly where it's broken and nowhere else. When a monitor or a log alert fires, it starts or wakes an agent that owns the incident, and we break things on purpose to test that those recovery paths work.

My eyebrows raised at least a dozen times reading this. I have trouble imagining this style of development working on software with higher stakes for when something goes wrong, but for a game this appears to be working surprisingly well.

Jacob says that Fable 5 (and then 5.1) was the model that made this level of autonomy possible. Opus 5.5 doesn't quite hack it for them, and when they tried GPT-6 Astra to apply fixes from a security review everything stopped working:

Every feature had been disabled and marked to be turned on again once the security fixes were done. Astra had gone through our whole product and made the insecure things secure by turning them off.

These kinds of extreme experiments are why I'm excited about game development as a substrate for exploring the edges of what today's coding agents can do.

# 29th October 2026, 5 pm / game-design, ai, generative-ai, llms, claude, coding-agents, agentic-engineering, claude-mythos-fable

GitHub Octoverse 2026. I'm at GitHub Universe today (GitHub gave me a free ticket through their "GitHub Stars" program). EVP Jay Parikh just announced this year's Octoverse report, and highlighted these numbers in particular:

Photo of a conference keynote stage: a wide curved screen shows three line charts side by side, with a speaker in a black sweater and dark jeans standing at its right edge beneath an overhead lighting rig. Each chart has a large colored headline number above it and a shaded line that runs low and nearly flat across the left side, then curves sharply upward on the right to end in a dot at its peak. Left, in green: "220M" "Merged PRs per month", y-axis 0M, 50M, 100M, 150M, 200M. Middle, in magenta: "7.38B" "Commits per month", y-axis 0B, 2B, 4B, 6B, with the most dramatic, almost vertical final spike. Right, in orange: "30.3M" "New repos per month", y-axis 0M, 10M, 20M, 30M, with a bumpier, more gradual climb. All three share the x-axis labels '23, '24, '25, '26.

This is a pretty stark representation of the impact of agentic engineering on GitHub. In 2026, merged pull requests tripled in less than a year to 220 million per month. Commits more than tripled to 7.38 billion per month, with most of that increase in the second half of this year. New repository rate doubled this year, to 30.3 million per month.

The report includes new numbers by industry, and breaks out numbers for repositories that involve "interoperability or reverse engineering". The methodology section includes notes about how they did this:

Top countries by pushes to public repositories, adjusted by profile location. Pushes from each country were summed across the Octoverse year. To adjust for biases caused by VPN usage, LLM classification was used to extract the ISO 3166-1 alpha-2 code from the profiles of GitHub users who contributed pushes during this period [...]

Industry. Public repositories created 2022 to 2026 were sorted into NAICS sectors by a fine-tuned multilingual embedding model using name, description, topics, and README, from random samples per year. [...]

Interoperability. Counts public repositories created each year whose name, description, topics, or README mention interoperability or reverse engineering, or related terms, such as clean room implementation, drop-in replacement, or protocol re-implementation.

# 28th October 2026, 4:45 pm / github, ai, generative-ai, llms, coding-agents, disclosures, agentic-engineering

Game Boy Color on My Conference Badge. I'm at GitHub Universe this week and, similar to last year, the conference badge is a full-blown color screen computer.

I've been hacking around with it again, and then I ran into with Steve Francia... who worked with Gemini to get a full NES emulator working on it, and then went a step further and deployed an emulator for both the Game Boy and the Game Boy Color.

Which means, if you can obtain the ROM, you can run The Legend of Zelda: Link's Awakening on your conference badge!

Photo of the badge running Zelda, in full color, with visible touch buttons below the screen.

A neat feature of the badge is that you can turn it upside down and then use game controller touch buttons that are built into the badge body.

# 27th October 2026, 7:34 pm / games, github, hardware-hacking, ai, generative-ai, llms, gemini, steve-francia

GPT-6.1 Terra (via) It's always interesting to note which aspects of a new model are emphasized in the release announcement.

In the case of GPT-6.1 Terra, OpenAI went with visual reasoning and computer use:

We’re introducing GPT-6.1 Terra, a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing. A significant step up from GPT‑5.6 Terra, it excels at the communication and computer-based tasks that fill the workday, from drafting proposals and interpreting technical drawings to filling out forms and editing videos.

Terra sits in the same position as Sonnet in the Anthropic model collection - not the cheapest or the most expensive, making it harder to decide when to use it.

Computer use is a good example of a task that benefits from a less expensive model, since you might put it to work on automated, repetitive tasks.

OpenAI's Ted Sanders provided some useful advice in a comment on Hacker News:

In the case of GPT-6.1 Terra, I'd say Terra is probably better than Astra and Sol at technical communication, similar on vision and computer use (but much faster!), and a bit worse on most other things. For API developers, it's a nice midrange option if you want something cheaper than Sol but a lot smarter than Luna (especially for vision / computer use / communication). For subscribers, it's not a bombshell release - if you're hitting usage limits, maybe it helps you go a bit further, and if you want clearer communication or faster computer use, it can be a nice option to select.

Here are the pelicans for the different reasoning levels. This pelican for "max" cost 11.3161 cents and took 4 minutes and 11 seconds:

It's a good pelican and a good bicycle - correctly shaped frame, feet n pedals either side of the frame, has the same color scheme as the others in the GPT-6 and GPT-6.1 families.

Here's a new comparison grid showing pelicans for every reasoning level of OpenAI's current frontier lineup: GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, and GPT-6 Luna.

GPT-6.1 Terra is priced at $1/million input tokens and $5/million output - almost half the price of GPT-5.6 Terra and GPT-6.1 Sol, and around 10x the price of GPT-6 Luna.

# 26th October 2026, 11:53 pm / ai, openai, generative-ai, llms, llm-pricing, pelican-riding-a-bicycle, llm-release, gpt, ted-sanders

Introducing Kimi K3.1 Pro (via) Kimi K3.1 Pro landed on the Kimi API today - it's not yet open weights, but Moonshot promise "Full model weights will be released by Nov 4, 2026".

They're also offering a "Highspeed" version that's 2-3x faster, at twice the price (except for input token cache hits, which are priced the same).

The US dollar pricing is $2.00/million tokens for input and $10/million for output - the same as GPT-6.1 Sol and half the price of Claude Opus 5.5.

Moonshot's own benchmarks show Kimi as approximately level with Opus 5.5 and GPT-6 Astra - beating them on a few benchmarks, coming close on others.

Artificial Analysis scored Kimi at 55 - ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon (all at 53) and behind only Claude Opus 5.5 (58).

This is the first time a (soon-to-be) open weight model has scored this highly against the frontier.

I tried it out via the Moonshot API using the llm openai endpoint command I added to LLM 0.32 back in August:

llm openai endpoint \
  https://api.moonshot.ai/v1 \
  --key kimi \
  -m kimi-k3.1-pro \
  "Generate an SVG of a pelican riding a bicycle" \
  --responses \
  -o reasoning_effort low \
  > kimi-k3.1-pro-pelican-low.stdout.txt \
  2> kimi-k3.1-pro-pelican-low.stderr.txt

llm openai endpoint can run prompts against any OpenAI-compatible endpoint but doesn't log the results, hence the extra options to save stdout and stderr (the reasoning trace).

Here are the pelicans I got for Kimi K3.1 Pro on the three reasoning levels it supports - low, medium, and max. This is max:

It's pretty solid - feet are on the pedals, bicycle frame is good, nice beak. Both legs are on the same side of the Fram though.

For comparison, I ran fresh pelicans against Kimi K3 on those three reasoning levels.

# 23rd October 2026, 7:40 pm / ai, generative-ai, llms, llm, llm-pricing, pelican-riding-a-bicycle, llm-release, ai-in-china, moonshot, kimi

DE STAAT—FREEDOM (Official Music Video) (via) OK well this is a first for me... at 00:44 in this new music video by Dutch alternative rock band De Staat:

The visual effects are credited to Matties Grooten of Fake it or Leave it.

# 22nd October 2026, 7:42 pm / music, ai, generative-ai, pelican-riding-a-bicycle

Donating Agent Skills to the Agentic AI Foundation. News today from Anthropic:

We’re committed to keeping Agent Skills open, community-driven, and vendor-neutral. Today, we are donating the standard to the Agentic AI Foundation (AAIF) under the Linux Foundation, which we co-founded with Block and OpenAI last year, and to which we donated the Model Context Protocol (MCP).

I'm pleased to note I called this one last December, when Anthropic first declared skills to be an "open standard":

I wouldn't be surprised to see this end up in the AAIF, recently the new home of the MCP specification.

Fun side note: since my blog content is available to Claude via the datasette-mcp plugin I found that post by prompting Claude to "Use simonwillison.net MCP to find if I ever talked about how I want skills to be donated to a standards body".

# 22nd October 2026, 6:09 pm / standards, ai, generative-ai, llms, anthropic, model-context-protocol, skills

Today’s PHP Security Release, and How We’re Strengthening PHP Security (via) If you're running anything on PHP it's vital you upgrade - today's updates (across the entire PHP 8.x line) address two remote code execution holes, and there have been exploits circulating since last night, thanks to a leak in the coordinated disclosure process.

This post from the PHP Foundation explain what went so wrong:

On Tuesday, October 20th, at around 15:00 UTC, the two CVE IDs began circulating on online forums and social media. They came with a wrong release time and a wrong CVSS score of 10.0 (the actual score is 9.2). No technical details leaked. In response, the PHP project brought its pre-announcement forward and published it at 19:30 UTC. It also moved the release from October 22nd, 14:00 UTC, to October 21st, 14:00 UTC [...]

As a result of the leak and pre-announcement around 20:00 UTC, Calif, one of the security firms that had reported the issues mistakenly published a video. It showed their proof of concept working against a number of targets and revealed a few additional, though limited, details about the vulnerabilities. Copies of the deleted video started circulating on social media soon after. The video contributed to accelerated development of various PoCs (exploitation proof of concepts), which then increased circulation on October 21st.

This article from the team that discovered the original vulnerability includes a more detailed timeline of the original intended responsible disclosure:

  • August 19, 2026: Wiz Research discovers the first PHP vulnerability using Wiz Atlas, Gemini 4 Argon, and GPT-5.6 Sol.
  • August 20, 2026: Unauthenticated RCE is confirmed against Drupal.
  • August 21, 2026: Unauthenticated RCE is confirmed against WordPress.
  • August 24, 2026: Wiz Research reports the vulnerabilities to the PHP security team.
  • September 21, 2026: PHP acknowledges the reports and begins developing a coordinated fix.
  • October 21, 2026: Patched PHP releases become available.
  • October 21, 2026: Public disclosure.

# 21st October 2026, 6:41 pm / open-source, php, security, ai, ai-security-research

Panache — RTS Belle Époque pour mobile (via) Well this is delightful: I released Battle Drones yesterday under an Apache 2.0 License and there's already a French language fork by Gabriel Turpin which reskins the game to take place in steampunk 1900s France and adds an animated intro!

Panache est un jeu de stratégie en temps réel pour mobile et navigateur, dans la France de la Belle Époque (vers 1900) : dirigeables, Fardiers à vapeur, Creusot cuirassés, automates en képi, Gargantua la forteresse marchante — et beaucoup de panache.

You can play Panache here - it's worth clicking that link just to experience the new intro video.

So far it's mostly just aesthetic updates as opposed to gameplay changes. The music has had a French twist added to it as well.

The new steampunk models are so cool!

The new buildings are gold and red, there are dirigibles and biplanes, and the Goliath stomping mech has had a total makeover.

Here's a page with the original Battle Drones models and that same page with the new Panache alternatives.

This is a comparison of the original and the new design for the Goliath:

On the left: blue stomping mech robot, missile launcher shoulders, classic sci-fi robot look. On the right: gold, steampunk, much more 1900s French look.

# 21st October 2026, 4:25 pm / game-design, open-source

eMudhra Intermediate CA Compromise. This is very, very bad news:

The private keys of eight eMudhra intermediate CA certificates have been compromised. They chain to roots trusted by the Mozilla, Microsoft, Apple and Chrome root programs and by many downstream trust stores.

Any software that bundles a list of trusted certificates needs to be updated. This site is tracking software updates via this GitHub repository - Linux distros need to update their ca-certificates packages.

In Python world uv 0.13.4 and pip 26.2.2 and certifi 2026.10.19 have been released with upgraded certificate lists.

There's a very informative thread about this on the official Python forum.

# 20th October 2026, 11:50 pm / encryption, https, pip, python, security, uv

My obituary for programming (via) Jeremy Scheff poignantly captures the key loss caused by coding agents for career programmers who enjoyed the puzzle of developing code:

Programming turned into one of the greatest jobs you could imagine. You got paid to work on tricky little logic puzzles. The trickier the puzzle, the more you got paid. But even pretty simple puzzles were worth a lot of money because regular people just couldn't do it! No, they needed us. They would give us a problem and we would solve it the way we wanted. We set the strategy, we made the timeline, we defined what was and was not possible.

If you're a programmer who thrives on building working systems and doesn't care too deeply about the problems solved along the way, LLMs are the gift that keeps on giving.

If directly solving those problems was a big part of your enjoyment, the loss is very real.

# 19th October 2026, 11:58 pm / programming, careers, ai, generative-ai, llms, coding-agents, deep-blue

Inline rendering with document​.currentScript. Lean Rada describes a neat trick that takes advantage of the document.currentScript property. You can use that to drop in a tiny inline <script> element that replaces itself when the page loads:

The year is <script>
document.currentScript.replaceWith(
  new Date().getFullYear())</script>

If you define a shortcut function like this:

$=(...n)=>document.currentScript.replaceWith(...n)

You get a tiny template language:

The year is <script>$(new Date().getFullYear())</script>

# 17th October 2026, 9:11 pm / javascript

Anthropic tells investors quarterly revenue up nearly 50% to $16.8bn ahead of IPO. The phrase "told investors" appears four times in this FT article, which reports on Anthropic's day of hosting investors on Wednesday. Lots of details in here from "people with knowledge of the matter".

Key things I learned from the article:

  • Q1 revenue (Jan/Feb/Mar) was "about $4.8bn"
  • Q2 (Apr/May/Jun) was $11.5bn
  • Q3 (Jul/Aug/Sep) was $16.8bn
  • In Q3: $1.70 in revenue for every $1 spent on "computing power"
  • Adjusted operating margins (excluding stock-based compensation) were 11%

This point about distillation is worth quoting in full (highlight mine):

The start-up has also clamped down on “distillation” — a controversial practice whereby rival AI labs train their models on the outputs of more advanced ones.

When rival companies “distil” their models from Anthropic’s models, that involves heavy usage and spending.

Anthropic told investors on Wednesday distillation accounted for 5 per cent to 7 per cent of revenue and that they had cut it back “materially” in the third quarter of the year, said four people with direct knowledge of the matter.

Napkin maths says that if distillation was 5% of their Q2 revenue (before they cut it back in Q3) then unnamed "other labs" spent over half a billion dollars distilling Anthropic's models in that quarter!

Also notable: Anthropic appear to no longer be leaning on "annualized run-rate revenue" (previously) in these communications to their investors.

# 15th October 2026, 1:04 pm / ai, generative-ai, llms, anthropic

AI is mass decompiling games. That’s both awesome and scary. (via) Indie game developer Popcar talks about the recent wave of AI-powered decompilation of existing games, enabling weird mashups between existing games, and producing a boom in game preservation projects to get old games working on new hardware (often in a browser using WebAssembly).

I like how this piece captures both the positives and the negatives of this new development:

Most people already realized the scary part by now. If AI can decompile games and software, then the entire tech industry is about to explode. Anybody can decompile other software and use it any way they'd like, and nobody can stop them. Companies can steal from competitors. Your paid game or app can be open-sourced against your will. This isn't exclusive to old video games, we can apply this to any software. [...]

This is also a big win for preservation. Old games can be ported, modified, enhanced, and made much more easily accessible than before. [...[

With console makers going all in on digital and games becoming entirely unplayable due to servers shutting down, this is fantastic news. We could own software we bought regardless of what the company thinks.

# 14th October 2026, 4:26 pm / computer-history, game-design, ai, webassembly, generative-ai, llms, ai-ethics, coding-agents

Redesigning source control at Anthropic (via) Interesting piece from Anthropic describing why and how they built their own custom alternative to Git, based on the dramatic increase in rate of changes to their code after the past six months.

In May 2026, Anthropic averaged around 2,500 code changes merged per day; as of September, we merged over 30,000 per day.

They included a chart showing those changes. I had Claude modify their chart (code here) to add markers for the public releases of each of their models in 2026:

Stacked bar chart titled “Changes merged per day”, for the Anthropic monorepo, January–September 2026. Each day’s bar is split into pull requests (blue) and antique change requests (orange), with a black line for the 7-day average. Weekdays start at around 1,000 merges in January and climb slowly to about 3,000 by late May. Antique change requests appear in late May, overtake pull requests in mid-June and have almost fully replaced them by mid-July, when weekdays reach 6,000–8,000. A labelled “Quiet week” at the end of July dips below 1,000 a day. After that, growth turns steep: over 10,000 a day by mid-August, about 17,000 in early September, and a peak of 30,401 on Sep 9. Dotted lines mark Claude model releases: Opus 4.6 (Feb 5), Sonnet 4.6 (Feb 17), Mythos Preview (Apr 7), Opus 4.7 (Apr 16), Opus 4.8 (May 28), Fable 5 and Mythos 5 (Jun 9), Sonnet 5 (Jun 30), Opus 5 (Jul 24, just before the quiet week), and Fable 5.1 and Mythos 5.1 (Sep 1).

This shows public release dates, not the date when models became available for Anthropic's internal use. My guess is that the introduction of Fable/Mythos class models is a key factor in the increased coding velocity.

"Antique" is their new code review system. The source control system is called Scion, and sounds to me like a reinvention of centralized source control along the lines of Subversion.

(Update: Anthropic credit Google's Piper and Meta's Sapling as "prior art" - it looks like the core concept of "derived data" was heavily inspired by Mononoke, the server component of Sapling.)

They had some pretty extreme requirements:

The history of our main branch alone is almost 100 GB, and a compressed working copy is over 25 GB. We perform on the order of 25 million code checkouts per day for CI, before counting developer environments, agent sandboxes, and research jobs. Fetched naively, that would be over 600 PB of git traffic per day.

The new system was built by a team of seven people, and the migration from Git was completed in five months.

When asked on Hacker News about the lack of offline support, Scion engineer Nick G replied:

Since Claude is by far the heaviest user of Scion and is necessarily online, this just wasn't important to do.

As an aside, I've been trying adversarial review in my own projects recently, where you have a model try and find flaws (I often play Claude and GPT off against each other). This note from the article reflected my own experience with that:

Anthropic also makes extensive use of adversarial review, and, while this is incredibly effective in many ways, it can also be myopic, so skilled engineers needed to steer Claude away from over-engineering parts of the system in response to hypothetical concerns.

# 13th October 2026, 11 pm / git, migrations, version-control, ai, generative-ai, llms, anthropic, agentic-engineering

Python 3.15.0 added to actions/python-versions. Bit of a niche link, but I've been waiting for this for a couple of days - I even told ChatGPT to check for it:

Clone https://github.com/actions/python-versions and git pull once an hour until they add the stable 3.15 - then tell me about it

Now that this has landed, you can add "3.15" to a GitHub Actions testing matrix to run tests against the new Python 3.15.0 release.

# 10th October 2026, 11:58 pm / github, python, github-actions, chatgpt

Deno is joining Cloudflare (via) The Deno team released the first version of celld back in August - their open source implementation of the Durable Objects pattern from Cloudflare Workers.

Today, Cloudflare are acquiring Deno outright, with the goal of building on celld to "make workerd self-hosting a first-class supported way to build and run apps using the Workers programming model" (see the Cloudflare blog.)

The bad news is that Deno itself will not be maintained by Cloudflare beyond the next year:

We will support the Deno runtime for another year with monthly releases containing bug fixes and security updates. After that year we will end our development of the Deno runtime. Deno will remain open source, and we welcome others who want to continue its development.

Deno (and Node.js) creator Ryan Dahl explained that decision in a comment on Hacker News:

It's a joint decision and I agree with it. I'm most invested in its success and have put the most work into it - and I no longer think it's where I can do the most important work. There are some good ideas in Deno and it's well engineered - but it ultimately is not solving big problems. It has been sucked into the gravity well of node compatibility, which forces it to behave exactly as Node does. Why reimplement Node? It works. Marginal performance or UX or security benefits are not enough.

I'm interested in building powerful new abstractions. celld has been working remarkably well, depending only on object storage for coordination and persistence. It is not just a slightly different API to interact with the file system or network - it's an entirely new model for server development.

My favorite feature of Deno has long been the permissions system, where you can run a Deno script and specify exactly which files and folders it can read and write to, and which network hosts it can access.

Node.js has a similar permissions model these days, added in Node v20.0.0 in April 2023 and declared stable in Node v22.13.0 in January 2025. They don't yet support allow-listing specific network hosts though - networking is either on or off.

# 9th October 2026, 10:48 pm / cloudfront, javascript, nodejs, ryan-dahl, sandboxing, deno

Anti-Patterns in Software Blogging (via) Some excellent writing advice from Michael Lynch. Michael warns against "meandering intros", misjudging your reader's existing knowledge, assuming they'll read your previous posts, and excessive formality.

He also warns against overreliance on links as an excuse not to explain terminology. This one hurt! I do this all the time, but I have a nagging suspicion that almost nobody ever clicks on them.

(In a Lobste.rs comment Michael clarifies that "My rule of thumb is that my article should still make sense to the reader even if they don't click any links". That works for me.)

This point about using your own voice is crucial:

Beginner software bloggers suffer from a mass delusion that you have to write in a stiff, overly formal way for people to take you seriously [...]

Just write the way you talk.

With so many developers delegating their writing to AI, software blogging is becoming bland and homogenous. Readers are hungry for writing with personality.

# 7th October 2026, 2:53 pm / blogging, writing, michael-lynch

Years

Tags