Tag: coding

Lawyers scramble to shut down vibe-coded AAA game ports for the browser

Some games work better than others, and it is currently whack-a-mole with url takedowns from publishers.
Thanks to Claude Opus 5.5’s gains in coding, it can now quickly recompile entire popular games to the web, and lots of AAA titles are currently getting posted online.

That includes games like Modern Warfare 2, Black Ops 2, Halo CE, GTA 5 and Vice City, Half Life and Diablo.

The games are responsive and playable, work just fine in the browser, and are popping up on obscure URLs all over the place.

The games mostly run very well, Kotaku notes, while also saying that they are of course DMCA-protected intellectual works.

That means they can be taken offline with a simple lawyers notice, but Destructoid points out they can be harder to stop than that.

Once anyone can get Claude to port a game to the web in an easy and fast fashion, they start emerging everywhere on the internet, and not always on .com addresses, but privacy-protected and hard to reach ones.

The industry is already hard at work taking websites down, Destructoid writes, but this might of course be better handled at the source — with Claude itself.

Read more: Kotaku, Destructoid, and Medium. Discussion on Hacker News and r/Singularity.

Anthropic reaches 90% AI use in model development, «leading» 26% of cases

AI use for improving models has not yet reached fully autonomous levels, but «leads» more often. (Picture: generated)
The AI lab has determined a methodology for measuring how AI is being used to build the next models, detail the extent and method of use, test their ability to intervene, and measure the resources used.

Anthropic randomly sampled 20% of research and development staff’s Claude usage for each week of July, and then used Claude to organize them. This was done using an «automation rating scale» from Epoch AI, classifying tasks from AL0, meaning no AI involvement, to AL5, which is fully autonomous AI.

They found that none of the usage reached AL5, but AL4 was 26% of the tasks — meaning AI is «leading» from «high-level prompts» with human supervision, while more than 90% of the tasks were AL3 — which is AI «collaboration» under «close human direction.»

AI «leading» is up from 1% in February 2026, and Anthropic estimates that they now have 30,000 agents in use at any time for research and engineering, that use real-time monitoring that can block «dangerous actions.» As agent use eventually grows ever larger, even rare use cases become more likely, Anthropic says.

The measure is a way of determining how close we are getting to recursive self-improvement, meaning AI writing itself, which could lead to a lack of control of the technology. Anthropic says every lab could publish such reports if there was a standard, which they hope to contribute to. They will also keep publishing these.

It’s also a step toward better third-party monitoring at Anthropic, which plans to install independent evaluators with employee-level access soon, as promised in their call for «pacing».

Read more: Anthropic’s report. Coverage on Reuters, ABC News, The Washington Post, and Bloomberg. Discussion on r/Singularity.

Meta releases Muse Code in beta, competing on price and performance

Muse code competes primarily on price and general «good enough» quality. (Picture: Meta)
The new coding model from Meta comes close to Opus 5 in Claude Code while costing less than half as much to run, at $1.25 per million input tokens and $4.25 in output. It is available today in beta.

— Muse Code takes on complex software engineering tasks across large repositories: planning changes, writing code, and validating the results, Meta says.

It can coordinate subagents that are persistent and run in the background through each session, instead of respawning for individual tasks, reducing latency and feedback loops.

The agents also use advanced logging of every call, tool run, approval and edit, meaning they should run directly after restarts and crashes, picking up from the latest log entry like nothing happened.

The underlying model, Muse Spark 1.2, scores well on benchmarks for its price point. It clocks in just behind Opus 5’s 86.7% on Terminal-Bench with 82.9%, beating GPT-5.6 Terra and Gemini 3.6 Flash. On DeepSWE it scores 59.3%, ticking in behind GPT Terra and Opus 5 Max. On Artificial Analysis, a more general benchmark, it scores a little behind GPT-5.5.

— Muse Spark 1.2 was extensively trained on long-horizon coding tasks, including whole-repository generation, large end-to-end projects, and auto-research, Meta says on the model.

Muse Code is available with a «one-line install» from the terminal on macOS and Linux, and is available on dev.meta.ai on the web.

Read more: Meta’s launch page, X post by Zuckerberg, X post from Artificial Analysis. Writeups on CNBC, TechCrunch, and Reuters. Discussion on Hacker News and r/Singularity.

Anthropic scrambles to contain fallout from Claude Code source code leak

The leak quickly spread to all corners of the internet and will be hard to contain. (Picture: Anthropic)
Anthropic has sent 8,000 copyright takedown notices after their source code leaked on March 31, worrying that competitors or bad actors might try to reverse engineer their coding agent, The Wall Street Journal and TechCrunch write.

Some 2,000 files and 512,000 lines of code had been inadvertently put on a public server, then copied to Github and «forked» (copied) at least 50,000 times, according to Ars Technica.

The code itself outlines Claude Code’s memory architecture, instructions for the AI bot and an upcoming feature with an always-on background agent, The Verge reports.

Developer and writer Gabriel Anhaia hailed Anthropic’s craft, calling it «both inspiring and humbling,» amid a deluge of comments all over x.com.

The leak was not due to malicious actions, but a «human error,» Axios says, and no user data was exposed.

Read more: The WSJ, TechCrunch, Ars Technica, and The Verge.

OpenAI developer releases Codex plugin for Claude Code

Codex for Claude Code might be a tad cheeky, but it’s useful. (Picture: screenshot)
Thanks to OpenAI’s Dominik Kundel, you can now call up OpenAI’s coding agent Codex within the Claude Code environment.

The plugin is fairly easy to install and use, so long as you have a ChatGPT account to log in with.

It’s handy for people who switch between the two models, and Codex on Claude can do things like review code, do an adversarial review, or hand off the entire task — where you should be able to switch apps and finish the work in Codex.

— This plugin is a simple way to keep your Claude Code workflow and still use Codex where Codex is strong, writes OpenAI developer Vaibhav (VB) Srivastav in the instructions.

Whether Anthropic will like Codex integration in its flagship coding product is anyones guess.

Read more: Announcement tweet, OpenAI dev community, and instructions for use.

Github Copilot to start training on user interactions from April 24

Github is coming for your code, after a successful trial on internal Microsoft data. (Picture: Github)
If you ever used Github to complete your code, your data can now be «used to train and improve our AI models,» Github says.

This comes after a trial period where Copilot has been feeding on internal Microsoft engineers’ data, which they say «improved model performance.»

They will not train on your entire code repositories, and will only use your interactions with Copilot — including accepted outputs, inputs sent to the model and «code context.»

Github is hardly alone in doing this, as Anthropic and OpenAI have been doing this for more than half a year. It’s common industry practice.

If you don’t like Copilot training on your data, you can opt out on the Copilot features page.

Read more: Github’s announcement, How-to Geek.

Anthropic launches Code Review for all those bothersome Claude pull requests

Drowning in pull requests from Claude Code? Anthropic has an answer. (Picture: Anthropic)
If you ever used Claude Code, you’d probably notice a mountain of pull requests asking for review on any decent code base. This takes time and effort for developers — but now Anthropic offers a solution.

— One of the questions that we keep getting from enterprise leaders is: Now that Claude Code is putting up a bunch of pull requests, how do I make sure that those get reviewed in an efficient manner? Cat Wu, Anthropic’s head of product, tells TechCrunch.

The answer is the newly launched Code Review tool that uses multiple agents to scan code changes, comment, and rate them for severity.

Using it internally, Anthropic found something of note in 84% of automated code reviews with more than 1,000 lines, they say.

The only problem is that it takes quite a few tokens to run a lot of agents on code changes, and the average cost is between $15-$25 per pull request, depending on complexity, Anthropic writes.

The tool is available as a research preview for Team and Enterprise plans as of today.

Read more: Anthropic’s announcement, writeups on TechCrunch and The Register.

Anthropic finds most agent use in software, with users interrupting often

Anthropic’s agents are overwhelmingly used for coding, but is also making inroads elsewhere. (Picture: Anthropic)
The AI lab has analyzed millions of human-agent interactions with Claude Code and their API. Unsurprisingly, they found most of the usage to be for coding work, with uptake in other sectors lagging far behind.

They discovered that while most of the usage is for one-shot code snippets, more users are letting Claude Code work autonomously, up to 45 minutes at a time after three months.

Continue reading “Anthropic finds most agent use in software, with users interrupting often”

Anthropic unlocks fast mode for Claude Code, 2.5x faster at 6x the price

For «work like rapid iteration or live debugging» Anthropic is letting users go fast, by toggling «/fast» on their console.

The feature is available in a «limited research preview» in the API for paid users with «extra usage» enabled, and there is a waiting list to get on it.

It’s the same model with the same capabilities and there is no change other than the speed — and cost.

The price for using fast mode is $30 for input and $150 in output for a million tokens, but there is a 50% discount until February 16.

Read more: Details on Claude, launch post.

Apple’s Xcode gets full agentic coding support for Claude and Codex

The new Xcode frees the developer from coding, so you can focus on innovation. (Picture: Apple)
With Xcode 26.3, there is full support for Claude and Codex — from idea iteration to file creation and structural editing.

Agents can also verify their work, and «collaborate throughout the entire development cycle,» Apple says.

They can search documentation, update project settings and «explore file structures.» Additionally, users can track what the agents are doing in the sidebar and adjust progress.

It only takes a single click to switch between models and pick the one best suited for the task, MacRumors writes.

Although Apple has worked exclusively with Anthropic and OpenAI to implement the features, any agent can be used — so long as they are using the Model Context Protocol.

Read more: Apple’s release, writeups on MacRumors, Engadget.

Google partners with Replit to bring «vibe coding» to the enterprise

Replit tightens its integration with Google models and Cloud.
Vibe Coding is comping for the enterprise. (Picture: Google, modified)
Decade old Replit has a valuation of $3 billion dollars and is a «leader» in the AI Vibe coding space, writes CNBC, and they are now tightening their integration with Google Cloud and the Gemini models.

—The goal for us, and Google, is to make enterprise vibe-coding a thing, Replit founder and CEO Amjad Masad said; — We want to show the world that these tools are actually going to transform businesses and how people work.

Under the new agreement, Replit will expand its Google Cloud use and «further integrate Google’s models into its platform,» Google writes on the deal.

Replit will gain access to all of the Gemini models, and the deal will «help enterprise customers embrace vibe coding.»

— Our mission is to enable the next billion software creators — from hobbyists to entrepreneurs to enterprises, Masad said.

Read more: Google’s announcement, writeup on CNBC.

Weekend roundup; expanded Sora, security research and the battle for India

For a limited time, Sora is available without invite codes for select countries, but the 30 generations per day limit may have to go.
No more invite codes for select countries in Sora 2, and bevy of new features. (Picture: generated)

Sora 2 expands, is now available without invite codes
Following the massive success of the Sora 2 video generator, OpenAI is opening up the service for those without invite codes in the USA, Canada, Japan and Korea «for a limited time.» Simultaneously, they are announcing reusable characters that can feature in more than one video and an easier way to stitch videos together. If that wasn’t enough, OpenAI is adding more video generations for power users hitting the 30-per-day generation limits and letting them pay for more gens. They are also musing about letting rightsholders get compensation for the reuse of their characters, as a means of getting paid for your work on the platform. They do warn that 30 gens needs too many GPUs and will be throttled at some stage.
More at: MacRumors and a Twitter announcement, list of available countries.

OpenAI reveals security research agent in beta
The new agent, Aardvark, will look through code repositories at scale almost like a human would, and find errors and exploits before the bad guys do. It will continually analyze your source code and find vulnerabilities. The agent has already been used to find «numerous» vulnerabilities in open source software, and OpenAI will provide pro bono scanning to «select, non-commercial» OSS systems. Aardvark is not being widely released, existing instead as a private beta inside OpenAI’s offices, kind of like Google’s CodeMender.
More at OpenAI’s announcement and ZDNet.

Read on for more!

Continue reading “Weekend roundup; expanded Sora, security research and the battle for India”

OpenAI’s 2025 Dev Day with Altman livestream incoming

Speculation is rife as to what Altman might announce at the livestream.
Will it be a browser, a new image model, or the highly anticipated AI device? It’s too soon to tell. (Picture: generated)
The October 6. stream will be an excellent moment to announce product news for the OpenAI CEO, but only a few items remain on their to-do-list.

Nothing has been announced but a cryptic tweet promising «new ships,» which could mean anything from new models to new modalities:

Continue reading “OpenAI’s 2025 Dev Day with Altman livestream incoming”

OpenAI tops ICPC coding contest for students, Google finished second

OpenAI solved 12 of 12 problems with vanilla GPT-5. Google had a custom model and solved 10.
OpenAI says they will now focus on scientific discovery. (Picture: OpenAI)
ChatGPT solved all 12 of 12 problems in the 2025 International Collegiate Programming Contest (ICPC) — an algorithmic programming contest for university students.

That result would have given it first place if it were human, as the best college teams only solved eleven.

Google also participated with a custom Gemini 2.5 Deep Think and earned Gold status, solving 10 of the problems and finishing second, Google claims.

Continue reading “OpenAI tops ICPC coding contest for students, Google finished second”

OpenAI announces GPT-5 Codex

GPT-5 Codex is slightly better than vanilla GPT-5 in benchmarks.
OpenAI is especially proud of the code review function in the new Codex. (Picture: OpenAI)
Savvy users have been using GPT-5-high with the Codex CLI (Command Line Interface) on their terminals for weeks, and consensus seems to be that it competes well with Claude.

Now, OpenAI is launching a custom, optimized version of GPT-5 for the Codex coding agent that they say is faster, more reliable and more steerable than before.

Continue reading “OpenAI announces GPT-5 Codex”