Astra Is Spiky, Tibo Resets Everyone, a Claude Code Engineer on Harnesses & a Year to Fix Security

Monday morning after Astra's first weekend in the wild. The tracked accounts were mostly arguing about whether the model is brilliant or maddening, and whether anyone can afford to find out. Karpathy, Boris Cherny, Lee Robinson and Simon Willison were quiet on X; Simon's blog and Bluesky fill in.

Astra Week Two

Theo: the spikiest model ever made

Theo's verdict (1,760 likes, 188 replies) after several days on GPT-6 Astra: it has done incredible things he never thought a model could do and also some of the stupidest things he has ever seen a model do, while "Fable 5.1 just does what I ask." He calls it the spikiest model he has encountered, "sometimes it feels like they invented God, other times it feels like they distilled the Gemini Flash model," and has never sworn at a model more. Asked how he measures it, he gave three questions: did it do what I asked, did it do it well, and did it do something incredibly stupid that I did not ask for. The chart he posted has no axis labels; asked what they are, he said "my blood pressure." To the "your prompts must suck" crowd he says anyone who has not seen these behaviors is not pushing the model hard enough. The replies mostly agree. neovectormind: Astra wrote the best code they got all year and then forgot what a for loop was. nutdistilled's transcript: research complete, did you seed the database, no, then why did I have you do the research. GutenLeinad's theory is an overcorrection after the Hugging Face incident, making the model reluctant about anything that looks autonomous. Theo's alternative framing is that GPT-6.1 will be by far the greatest model ever made if OpenAI does to Astra what it did to 5.5 with the follow-up. A video is coming. He also hit a long Codex /goal run with an unexplained interstitial; replies identify it as the cyber safeguard classifier that pauses a thread while deciding whether to let it continue, and several people say it fires constantly on legitimate work like debugging system libraries or bio research.

Tibo Rickrolls a global reset

Thibault Sottiaux announced a global usage reset for all paid subscriptions "so that you can keep enjoying Astra after burning through all of it doing fun 3D modeling in blender," hidden after six lines of Never Gonna Give You Up. 31,500 likes, 4,500 replies, and it landed at 6pm Pacific. adryt3ch's "a reset a day keeps Claude away" got 2,000 likes on its own. He then teased a 28-page deck (2,800 likes) in which Codex is tracking every individual upcoming launch. The replies to both are the real signal: a Pro 20x user says Codex has been "basically unusable for ~2 days" with capacity errors and only automated support replies, several people say the reset is useless while "selected model is at capacity" persists, and SynKavish reports a single two-hour Astra Medium task burning 80% of a Pro 5x weekly limit where Fable 5 did 90 minutes for 15 to 20% on a £20 Claude plan. There are also repeated requests to drop the five-hour window and a Codex CLI feature request for a finished subagent to wake its parent without polling. Theo is still waiting for OpenAI to explain what Saturday's "3 to 4x less usage" change actually was and says he has a conspiracy theory he will share if no official word arrives. Guesses in his replies: caching changes, slower medium reasoning, or that they simply lowered the multiplier.

Stop using Astra from the Claude Code CLI

Tibo also told people (2,700 likes, 308 replies) that you can use Astra inside the Claude Code CLI if you insist, but you are missing out by not using the Codex desktop app, which gives Astra computer use, subagent management, voice and non-blocking context-gathering questions. Paul Hudson's reply (791 likes) is the counterpoint: Astra beats Sol, but Codex still has a way to go to catch Claude on the CLI, the new recaps are welcome, and he wishes they would "be inspired by" Claude's subagent management and sticky prompt headers. A Chinese-language reply saying they have not opened the Claude Code CLI since switching to the Codex app got 455 likes and Tibo's 你最牛了 in response got 3,600. Linux users want computer use, several people call the desktop app slow and buggy, and one asks him not to make the CLI second class. Benjamin Badejo notes OpenClaw picks up Codex with Astra, including computer use, if it is installed on the same machine.

Browser use, 3D models and SVGs

Browser Use founder Gregor Zunic reports Astra at 77% on their browser benchmark against 56% for Fable 5.1, with the note that Fable "just refuses too much." Every substantive reply asks for refusals to be scored separately from failures, since a refusal is a policy choice and not a capability gap. Andon Labs says Astra ranks first on Blueprint-Bench 2, drawing floorplans from apartment photos, and is near human performance. am.will continued the Astra 3D and SVG run: a tip thread on getting authentic results, which is to have the model replicate an object it can reference obsessively, like the International Space Station, and define the outcome rather than the number of passes; a "draw me" prompt that took 90 minutes on Astra Max; and a PS5 controller SVG that took 32 minutes at max effort. Jerry Liu made GTA 6 at home out of FOMO. cherry_mx_reds says they run no AGENTS.md, no skills and no scaffolding and Astra has been great, so if you are having a bad time, reconsider your setup before writing off the model. Simon Willison added gpt-6-astra to llm 0.35.

Agentic Coding & Agent Harnesses

Thariq Shihipar on how Anthropic engineers actually work

Ryan Peterman's hour-long interview with Thariq Shihipar of the Claude Code team is on YouTube with a full transcript. The chapter list promises "what percent of Anthropic's changes are fully autonomous," and one reply notes that is the number and the rest is a podcast; the honest answer is that Thariq dodges it, saying it depends what counts and that the goal is for Claude to do all the glue work so humans spend their day on work that is unique. The useful parts are elsewhere. On harnesses: the idea that they stop mattering as models improve is backwards, because better models run longer and the harness has to grow to let them do more. Auto mode is a classifier that runs after every action that would normally need a permission prompt, which barely mattered when turns lasted minutes and is load-bearing now that Claude runs for hours. Artifacts exist because eight hours of work has to be represented somehow, and how it is represented is itself a form of prompting. His main prompting tip is permission: models default to what the average user wants, which is to start work fast and not burn compute, so if you want depth you say "this is a hard problem, use subagents, use workflows" or "don't do any work yet, brainstorm with me." He reads Jared's Riemann hypothesis "believe in yourself" story as really being permission to spend compute. On model choice, he uses Fable for planning, brainstorming and finding unknowns and Opus 5 for execution, and thinks nobody will be choosing models by next year. On writing: his rule is that if he would be happy to show someone the prompt, he will send them the output, and before every one-on-one he has Claude read every Slack message and PR and compile a report. On maintenance: naming and stylistic opinions matter less now, the "simplify" skill gets heavy use, but you do not want the model to simplify by default because it may simplify something incorrect. On interns: the work is no longer "write the React code" but the growing list of problems nobody has done before, like evaluating new coding behaviors across millions of users. And on breakages: Anthropic's uptime is not where they want it either, the fix is using Claude to build the dream testing environment, replaying requests across mock databases and fixtures, and chaos-testing everything.

Knowledge work is much harder than code

Matt Pocock argues (2,600 likes, 243 replies) that we underestimate how easy software is for agents: high-quality automated feedback loops in types and tests, well-organized free documentation, version-controlled workspaces, and a long history of systematising and delegating work through sprints, tickets and specs. Knowledge work has none of these. A lawyer replied that this feels correct and adds that law allows no do-overs, since hallucinations mean bar complaints and liability, so verification has to be flawless; Matt's framing is that software changes are mostly two-way doors and many disciplines only have one-way doors. ivanainai's line got repeated a lot: software spent 50 years making itself legible to machines and now we are surprised machines are good at it. A dissenting thread from pitsch cites David Graeber on how much knowledge work is box-ticking with no real output, so agents will eat the empty ritual before the craft. Thariq's podcast answer to the same question is the optimistic version: most knowledge work is reducible to code if you think about it well, and he does his accounting in Python and video editing in ffmpeg via Claude Code. Matt's follow-up, quoting Kent C. Dodds, is that tactical programming is dead and strategic programming has never been more vital.

Why are you making me ping my agent?

Peter Steinberger sent a PR upstream (649 likes, 96 replies), got minor change requests, and asks what the point is: you already wrote the prompt, why make me ping my agent again so your agent then merges. Max Howell says it feels rude to take the contributor's option to make the edits themselves; Peter says the contribution is still fully tracked and the "odd" feeling comes from a time before agents did the coding, and anyone who dislikes maintainer edits can turn the GitHub setting off. He does not run fully automated reply loops on external repos out of respect for contribution policies that are not AI friendly. ccccjjjjeeee's team policy is that whoever requests the change gets it over the line unless it genuinely needs discussion. NotASecretLich describes the absurd version: they review with agents, pull the branch, know exactly what is wrong, and then send the author a prompt for the author's agent instead of fixing it. richardkmichael sends back a pr-review branch with a kick-off prompt for larger changes. lajoiedeslutins: "you've invented a relay race where both runners are the same guy."

OpenAI's internal tools explosion without internal tooling teams

Gergely Orosz reports (3,700 likes) from a visit to OpenAI HQ that ex-Meta hires wanted to build internal tooling teams the way Meta did, leadership refused on the grounds that an AGI-first world has no internal tooling teams, and now Codex has produced a massive internal tools explosion with no such team. A Pragmatic Engineer deep dive is due in about a week. Tibo pushed back (1,460 likes) in the thread: Codex started as an internal tool explicitly built as the highest-leverage way to accelerate infrastructure buildout, and the Meta versus non-Meta angle is not enlightening, since most people from older tech companies were surprised how far OpenAI had taken Python scripts. Gergely says he will reconcile the two accounts. adam_fresko's reply is the individual version: fourteen personal apps that will never be published because they do exactly what one person wants.

Theo on agent memory for teams

Theo cannot see (1,980 likes, 339 replies) why "agent memory for enterprise/teams" is a product category. joshu offers that the repo is the memory for code, so this is repos for non-code things. Theo's position is that it should be system prompt adjustments or, at worst, a manually curated set of skills and knowledge, and that anything that automatically plagues the context of every agent at a company will cause more problems than it solves. When someone described durable architectural context across many repos so it is not re-researched every session, he answered that this is documentation, not an automatic memory product. His codebase video (164,000 views) and the accompanying post (3,059 likes, 399 replies), "if you understand your whole codebase, the codebase isn't that important," drew a crowd he characterises as developers who have only ever worked on codebases alone. The best reply is rshthr quoting Tony Hoare's 1981 Turing lecture, where a customer shouts "you let your programmers do things which you yourself do not understand," and Hoare thinks him out of touch for imagining anyone could understand the whole Elliott 503 Mark II system. Dave Plummer agrees: the last codebase you could hold in your head was MS-DOS, and by Windows NT you were an expert on the scheduler or the allocator, not both.

Other Interesting Stuff

We have a year to fix security everywhere

jyn's essay A year to fix security (HN, 190 points) lays out the argument in three words: GLM 5.3-flash is open-weight, so DeAlignAI's abliterated version scores 0% refusals on HarmBench; it is flash, so it runs at 20 tokens a second on a $6,000 GPU today and roughly 45 on the 256GB M5 Mac Studio shipping September 22; and it is frontier, with GLM 5.3 at 84.5% on CyberGym, above GPT-5.6 Sol, and 54.4% on ExploitBench, where Astra scores 100%. Put together, anyone with savings can run a model that is willing and able to find and exploit vulnerabilities on the open internet round the clock, and the historical three-to-six-month lag means an Astra-class GLM within the year. The policy section is the interesting part: banning the weights is useless, banning Mythos and Astra access makes things worse by removing defenders' best tool, GPU export bans buy a little time, and the real bottleneck is not finding bugs, which Glasswing and Daybreak already do cheaply, but triage, remediation and deployment, so fund security engineers and mandate frequent testing rather than specific techniques. For companies: sandbox your agents, restrict credentials to narrow scopes, use structured prompts like Google's Unsafe Rust Review rather than "look hard for bugs," and give patch-writing models iterated self-review loops rather than one-shot prompts. Simon Willison's HN comment: he does not think we even have a year. Simon's creepy crawlies link is the adjacent pain: git.kernel.org spends more CPU rendering commits for scrapers than on all legitimate access combined, 14 cores across five nodes doing nothing else.

Mistral raises €3 billion

Mistral announced (HN, 360 points) a €3 billion Series D at a post-money valuation above €21 billion, the largest equity round ever by a European tech company. Samsung Electronics led, with EQT's Scaleup Europe Fund and PSG Equity co-leading; BlackRock funds, Advent and the Grand Duchy of Luxembourg are new investors alongside a16z, ASML, NVIDIA and Salesforce Ventures. The pitch is sovereignty: open weights, owned compute and products so that customers are never locked to one vendor's roadmap, with Airbus, ASML and HSBC as reference customers across 20 countries. HN splits along the usual line, with davedx arguing they are deliberately not in a benchmark arms race with China and landing European customers for the right reasons, and others saying they are not on par with Chinese models from a year ago.

Latent Space launched the Frontier AEO tracker, extending AmplifyingAI's "What Claude Code Actually Chooses" to every frontier model across 161 categories, with every prompt and answer pair inspectable. Findings from the writeup: 28 categories have a universally dominant choice across all models; Fable and Opus recommend Claude Code while Sol and Astra recommend Codex, though GPT models do sometimes recommend Claude; Anthropic models search far more sources than OpenAI ones, with Fable's median at 15 against Astra's 5, and Astra is much less likely to change its answer after searching; and markdown content negotiation of the kind measured by Ora and Vercel's is-agentic.com really does affect whether models read your content. Gemini, GLM and DeepSeek were left out of the first run because of errors and rate limits.

Shorter

Simon Willison on how he reads pasted LLM replies in online arguments: "I don't know the answer myself, but I asked a blowhard I know and he took a wild guess, here's what he said." He also notes recent prompt guidelines from Anthropic and OpenAI recommend fewer detailed rules and more "use your own judgement," since models are strict enough rule followers to follow the process to the detriment of the result. Simon also quoted Jakub Pachocki on the strongest argument for continuing to train much smarter models being defence against other AI, paired with "racing forward at all costs seems absurd."
Theo says software is falling apart because the developers do not use what they build, and unlimited tokens make it manifest faster: if developers are not using the product as they ship, bugs and jank expand with velocity. Joel Hooks: "slop is a choice."
Adafruit covered T3CAD, an agent-driven electronics workspace with KiCad PCB, schematic, Gerber and 3D support.
Armin Ronacher's weekend Astra slop thread from yesterday is still running; his software factory has now reinvented let bindings as def.
Theo went live Sunday night for a Fable versus Astra showdown.

Sources: Nitter RSS and thread pages via x.n0g.xyz and nitter.jaydenha.uk for all tracked accounts; Bluesky for Simon Willison; simonwillison.net; Theo's YouTube feed; developing.dev transcript; Latent Space; Hacker News. @potetotes still returns "user not found" and @bcherny, @karpathy, @leerob and @simonw had no posts in the window.

§Astra Week Two