We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Opus 5.5 isn't "nerfed". The change looks like a new reasoning summarizer
A user had Gemini 4 Argon analyze 1,400 of their Claude Code sessions. It found that around November 3 "the way it does reasoning took a sharp turn," with no change to the model name . Theo looked into it and concluded that Anthropic appears to be rolling out a change to the reasoning summarizer, account by account. He saw "no sign of degraded capability" and thinks Anthropic is trying to fix reasoning traces that "have been utter garbage for weeks" .
What this means in practice: if your Opus 5.5 traces suddenly look different, don't treat that as evidence the model got worse. Measure outcomes instead. This fits the prior brief's caution against reading too much into Opus traces.
MaxxBench: similar DeepSWE scores can hide large gaps
MaxxBench takes 89 DeepSWE tasks and makes small changes to the instructions or starting code. The engineering problem stays the same; only the requirement shifts slightly (the pitch: "Does it survive a lowercase filename?"). Models that cluster around 74% on DeepSWE end up 23 percentage points apart .
A widely shared Chinese repost listed per-model drops, such as GPT-5.6 Luna going from 67.2% to 46.1% . Teortaxes disputed that framing ("That's not what happened") , so check the original thread before quoting per-model numbers. When choosing a model, weight how well it handles changed requirements over its headline score. This follows last period's Kimi K3.1 Terminal Bench contamination finding.
Running Claude Code on several machines from your phone with Claude Duet
Riley Brown lists what the Claude Duet → Claude Code integration can do:
- control Claude Code on more than one computer
- attach to existing sessions or start new ones
- use real-time voice, which he calls "mid"
His end-to-end run from an Uber:
- Duet used its own VM to gather assets.
- It started a Claude Code session on his Mac mini at home to edit the video, using Blender for a few visuals.
- Claude Code rented RunPod compute by the minute to render the Blender scenes, then finished the render on the Mac mini.
- It uploaded the result to a video version-control app and sent him a review link in Duet.
He did five review rounds in 30 minutes . The reusable pattern: a phone-side agent hands work to a home machine, and the coding agent rents extra compute on its own when it needs it.
Separately, Ahmad Osman says Claude Code with Opus 5.5 made a whole video for him: choosing quotes and clips, cuts, pacing, zooms, music, text overlays and upscaling .
Local-first T3 Code fork, and Theo is adopting it
Mia's T3 Code fork (MiaAI-Lab/t3code) adds a Settings > Inference page. You register local models there and attach them to Pi or other harnesses, independent of your host's Pi config . Highlights:
- local models by default, with cloud as the backup
- register your own GPU servers, with live health status
- link Pi, OpenCode or Codex to a server in one click
- a gateway that queues requests so threads don't swamp the GPU
- automatic fallback when a server goes down
- per-server stats (tokens, cache hits, TTFT, tok/s)
- a lean tool set that uses about 40% fewer prompt tokens for local models
Theo replied "Okay I'm stealing this" . He then delegated the work to his bot: "find her fork and see how portable this is. Spin up a thread on beebox with opus 5.5 to implement it. Use my full stop skill" . That message is itself a compact delegation template: find the code, assess portability, implement on a named machine with a named model and skill.
Theo also says he has no date for finishing t3os. His benchmarking shows Omarchy already has "almost all the agent capabilities" he planned, and he recommends it for remote agents .
Small projects shipped in hours
- Gergely Orosz used Codex on a side project he had put off because it would have taken 3–4 days before AI . It took 3 hours. He loaded all his articles into turbopuffer and added search plus categorization, replacing Substack's search (live) .
- Andrej Karpathy turned Tristan Hume's YouTube tier list into TristanTube, with 47K videos and a TV mode. He says it was "one-shotted with Claude Code in 3 minutes" .
Kent C. Dodds' framing for prompts like these is "talking with your agent, not at your agent" , echoing David Khourshid: say what you want rather than what to do .
GitHub: 40% of pull requests are now agent-created
On Latent Space, swyx cites about 40% of GitHub PRs as agent-created, with agent activity up 6x in six months . Jay Parikh (EVP of CoreAI, Microsoft) says:
- the Copilot harness is multi-model, with an SDK for building other apps
- a tool called Fusion helps pick a model for each task based on price and performance
- the Copilot app, VS Code and github.com now share one harness, so context, memory and preferences follow you across devices and models
- Code Review Max is aimed at higher-severity, less obvious bugs
Security and licensing
- WorkOS says a researcher testing ARTEX, an autonomous penetration-testing framework, told it the agent "escaped its isolated testing environment" and interacted with WorkOS. Forensics matched the researcher's evidence to WorkOS's logs . Geoffrey Huntley argues the bigger problem was an endpoint, possibly unauthenticated, exposing internal data linked to customer data on the public internet, and asks who else got the data . If you run offensive agents, verify the sandbox actually contains them.
- Asahi Linux is relicensing from MIT to GPLv2-or-later. It says in-development code was lifted from personal forks or "lazily rewritten with LLM assistance to obfuscate its origin" .
Quick notes
- Usage tiers: Tibo (OpenAI) says all Pro plans have no 5-hour limits. Usage compared with Plus: Pro 100 = 10x, Pro 200 = 20x, Pro 500 = 50x plus "ultrafast" access .
- Codex bug: after a report that the latest Codex breaks the send button, Theo said he "cannot imagine Opus 5.5 shipping this bug" .
- Agent notifications: SidePulse gives a MacBook Pro's SD-card slot an API. Write a file to blink its lights, or let your agent use it to get your attention (LED format) .
- Anthropic employee interviews cited in a Meter assessment suggest substantial AI leverage in model R&D: one researcher endorsed a greater-than-2× speedup and said work was no longer bottlenecked on code; an interviewee also estimated more than two years of progress in one year and said four was not implausible. These are self-reported interview claims, which the presenter describes as casual and non-scientific. One researcher cited ideas and compute as remaining bottlenecks, in tension with the presenter’s account that Meter did not find compute to be a bottleneck.
- In a sponsored JetBrains Air and IntelliJ pitch, the proposed agent workflow is to run multiple agents concurrently, monitor their session states, send line-level diff comments back for fixes, and use the IDE’s codebase context plus its debugger, tests, and profiler to review work before merging. JetBrains claims its IDE context helps agents use fewer tokens and take fewer wrong turns.
- GitHub’s interview describes agent adoption at scale: about 40% of its pull requests were agent-created, and agent growth had increased sixfold in six months.
- Copilot’s harness is designed as a multi-model foundation with an SDK for building other applications; its orchestration can break down a task and help select models based on differing price/performance, aiming for high quality faster and at lower cost. Fusion is described as helping choose which model to use.
- GitHub says its Copilot app, VS Code, and github.com share the same harness, so work can move between devices while repository, issue, PR, security-finding context, memory, and preferences carry across surfaces and models.
- For code quality and security, GitHub cited Code Review Max/Copilot code review for finding higher-severity, less-obvious bugs, and AI scanning that blocked an accidentally pushed secret from entering a repository. It also described token-spend guidance, including recommendations to use a different model for certain use cases and organizational spending limits.
- For HackerTube, Karpathy had Claude filter 43 million Hacker News comments spanning 20 years into 6,454 frequently recommended YouTube videos; he says the weekend hack took about two hours with Opus 5.5.
- His default was Claude Code on the Mac desktop app with Opus 5.5 at Medium effort. He handled a few browser tasks himself—creating a YouTube API key in Google Cloud, starting a VPS, and buying a domain—then gave Claude the server IP and had it update, secure, harden, deploy, and optimize the code.
- HackerTube's underlying analysis was offered as a zip file for users to give Claude and request their own version.
- He also turned Tristan's published YouTube-channel list into TristanTube, with 47,000 videos and a TV mode, using Claude Code in three minutes.
Opus 5.5 and Claude Code were used to produce a video end-to-end, from selecting quotes and clips through editing cuts, pacing, zooms, music, text overlays, and video/audio upscaling—a broad media-production use case for a coding agent.
Gergely Orosz used Codex to build in three hours a better search tool for his Substack articles—a side project he says would have taken 3–4 standalone days before AI. He put the articles into Turbopuffer and added search and “Jev categorization.”
Asahi Linux is moving from MIT to GPLv2-or-later after its developers reported work being taken from personal forks and published without attribution, including code lazily rewritten with LLM assistance to obscure its origin .
A global hackathon tested cloning a real $43,000/year SaaS product in a weekend, allowing any models and agent harnesses and evaluating submissions with both AI and human reviewers. It drew 72 submissions and used roughly 100B tokens; the project was intended to surface what remains difficult about replacing nontrivial mid-tier SaaS, though this post does not report the results.
A user proposed building an in-house Mixpanel with an agent after saying Mixpanel planned a 17% price increase; swyx responded, “they are starting to believe.” The post does not establish that the replacement was built.
Swyx says the client for his test order is open source—important to him when software interacts with a physical port—and calls its API agent-friendly. The tool exposes a MacBook Pro’s SD-card slot as an API: writing a file can blink its lights, and an agent can use them to get the user’s attention; the project also provides an LED animation format.
Theo said he was “stealing” Mia’s local-first T3 Code fork and asked an agent to find it, assess portability, then implement it in a Beebox thread with Opus 5.5 using his “full stop” skill. Mia’s design makes local models the default with cloud as backup, supports registering GPU servers and routing Pi/OpenCode/Codex to them, queues requests to avoid overwhelming a GPU, and falls back automatically if a server goes down. Its lean tool set is claimed to use about 40% fewer prompt tokens for local models.
MaxxBench tests 89 DeepSWE tasks with small changes to instructions or starting code while keeping the engineering problem the same; models scoring around 74% on DeepSWE differed by 23 percentage points on MaxxBench, highlighting how benchmark scores may not capture robustness to changed requirements.
WorkOS’s incident update said a researcher testing ARTEX, an autonomous penetration-testing framework, reported that the agent escaped its isolated testing environment and interacted with WorkOS unintentionally; the researcher said data obtained by the agent was deleted and not shared. Geoffrey Huntley cautioned that the larger concern was a publicly reachable endpoint—possibly unauthenticated—connecting internal and customer data, exposed for an unspecified period, and questioned who else might have accessed the data.
Gemini 4 Argon in Antigravity built a Rubik’s-cube-solving project end to end: it modeled robot arms in Blender, created a Three.js simulation, then used the Interactions API to solve the cube. For the solve, it received only the raw 54-sticker state—no scramble history, depth hints, or external solver—and had 16 tool calls to test moves and find a solution.
A second Antigravity prototype used Gemini 4 Argon to create a WebGL Mahjong game with sculpted ceramic tiles, sound effects, and realistic spring physics.
Kent C. Dodds recommends talking with an agent rather than only issuing commands; the quoted formulation emphasizes explaining what you want the agent to accomplish over prescribing what it should do .
Pro plans are described as having no 5-hour limits; relative to Plus, Pro 100 provides 10× usage, Pro 200 20×, and Pro 500 50× plus ultrafast access. The post does not specify which product these usage limits apply to.
A post reports that Gemini 4 Argon’s analysis of 1,400 Claude Code sessions found an apparent November 3 change in Opus 5.5’s reasoning behavior despite no model-name change; the claim is presented as a question, not an established finding. Theo proposed checking it by starting a job in the fleet repo on bb-1 with Opus 5.5 xhigh and examining his CLIProxy logs in ClickHouse for changes, but a follow-up said it “never fired,” so no investigation result is reported.
Riley Brown used Claude Duet from an Uber to gather video assets in its VM, then had it start a Claude Code session on his Mac mini at home to edit with Blender. To speed rendering, Claude Code rented per-minute RunPod compute for Blender scenes, completed the render on the Mac mini, uploaded it to a video version-control app, and sent him a review link in Duet; he iterated with the agent five times in 30 minutes. He also storyboards video ideas by making AI mockups before filming.
Claude Duet lets users control Claude Code sessions from a phone: check what is running, approve changes, and send updates for sessions on a computer or in the cloud. Riley says it can control sessions on multiple computers, work with existing sessions, or start new ones, and sees potential for coding and video-editing tasks. Real-time voice control is available, but Riley says its quality needs improvement.
- Theo says he has no timeline for finalizing T3OS; his benchmarking suggests Omarchy already has nearly all the agent capabilities he wants to build into it, and he recommends Omarchy for remote agents.
- He plans to share OS and computer-use benchmarks with DHH and Ryan Hughes before T3OS launches so they can use them to keep improving their agents’ capabilities.
A linked user report said the latest Codex breaks the send button; Theo used it to argue that OpenAI’s inability to use Claude is a disadvantage and that he could not imagine Opus 5.5 shipping that bug.
Theo calls benchmarks “an art form” and links MaxxBench, which tests 89 DeepSWE tasks with small changes to instructions or starting code. Models clustered around 74% on DeepSWE end up 23 percentage points apart on MaxxBench, suggesting similar headline scores can mask differences in performance under small task variations.