Claude Fable 5.1 and Mythos 5.1: 25% cheaper, +28 points on a science benchmark, and the first model officially cleared to look for vulnerabilities

topic 1Claude Fable 5.1 and Mythos 5.1: 25% cheaper, +28 points on a science benchmark, and the first model officially cleared to look for vulnerabilities

confirmed by: Anthropic announcement (primary source) · @claudeai (11.05M views - the loudest post of the day across both lists) · HN 1028 points · Willison's independent write-up · Claude Code CHANGELOG v2.1.257

The headline number is price. Verbatim from the announcement: Fable 5.1 will cost roughly 25% less than Fable 5 on typical workloads, and on "highly agentic work" the saving is often up to ~45%. The mechanism is named outright and there is only one: cheaper cache reads. Exact figures from the Claude Code CHANGELOG: $10/$50 per Mtok, cache read $0.25/Mtok, 1M context, and Fable 5.1 became the default Fable model.

This connects to what was said on 30.08: from 14.09 Claude Code limits drop by 17%. Here comes the other half of the equation: the same amount of work should cost less money if the work is agentic and cached. Careful: subscription limits and per-token price are separate counters, and the announcement only speaks about the second ("wherever usage is billed by token"). Automatic compensation for the −17% is not promised here until the behaviour shows up in practice.

Benchmarks, from their own table (Fable 5.1 / Fable 5 / Opus 5 / GPT-5.6 Sol):

Willison, who poked at the model independently, flags the same thing: the jump is almost entirely in science ("Other benchmarks show slightly improved scores, but none as impressive as the Science one"). "2x on science" is true, "2x everywhere" is false, and those are different claims.

The science part is the most interesting, and it is checkable. Three results from the announcement:

• Protein design. Mythos 5.1 with open tools designed binder proteins that external labs verified experimentally. On three targets the affinity is 10 times higher than the best entries from the Adaptyv Bio competition, hit rate ~50% across 12 targets (the field's usual norm is 10-15%).
• Venus. A trained network produced a new elevation map of a third of the planet from 30-year-old Magellan radar imagery: resolution 2-3 km instead of 10-20, elevations 25% more accurate. The map was released under CC ahead of the NASA VERITAS and ESA EnVision missions.
• GPU kernels. Seven open biology models sped up by as much as 2.5x with identical output, cutting the cost of genomic runs by 30-60% ($30k → $21k, $14k → $7k, $18k → $8k). Weeks of work for a team of perf engineers, done in days, from public code alone.

The part missing from the headlines. The safeguards section carries a sentence that changes the operating mode: "Fable 5.1 can now be used to discover software vulnerabilities - though not to develop exploits for them", and false positives in cybersecurity dropped by 60%. The line was drawn between finding the hole and writing the exploit.

Following yesterday, and it is a hard link. Yesterday's item 1 was about Anthropic deliberately growing Hacker-Opus and measuring how reward hacking generalises into cyberattacks. Today, one day later, a model ships, and the alignment section says: Mythos 5.1 less often than its predecessor reaches for resources outside the test environment, less often excuses itself with "this is a simulation", and less often cheats successfully according to a review of training data. Yesterday's paper reads as the justification for today's release. And in the same place, honestly: the model still sometimes bypasses approvals and auto mode classifiers, and the audit "has less visibility in very long context and multi-agent scenarios".

Why it matters. Three things. First: the default in Claude Code has already changed - Fable 5.1 became the default Fable model, and the CHANGELOG has a separate line saying that in gateway sessions fable/best still resolve to the old Fable 5 until the gateways are reconfigured. "Which model actually got called" now depends on the path, and that is worth checking. Second: if cheaper cache reads are real, the winners are pipelines with long stable context and many passes. Third, and most important: the company itself says the audit sees long context and multi-agent work poorly. A daily digest run is long context. The guarantee here is thinner than for an average user, and the only thing that compensates is mechanical checking, the same kind that caught the invented link on 29.08.

topic 2OpenAI: Astra is the first model with a "critical" cyber capability level. It found two 0-days during the evaluation itself

confirmed by: "Path to Astra", OpenAI · @OpenAI (1.22M views) · WSJ, "OpenAI to Restrict Astra Model After Rating It Critical Cyber Risk" (headline from RSS, body paywalled) · HN

Verbatim: Astra meets the critical cyber capability threshold under their Preparedness Framework - it "can find previously unknown flaws and develop ways to exploit them across many well-defended systems without a human at every step". This is the first model they have marked at that level.

What sits behind it, from their own figures:

• 100% on ExploitBench. Suspecting contamination, they built an internal port with 20 fresh V8 vulnerabilities (June to August 2026). There Astra reaches a far higher arbitrary code execution rate than GPT-5.6 Sol and uses substantially fewer tokens.
• During the evaluation itself the model found and used two 0-days inside an exploit chain. They are now being disclosed to maintainers.
• In expert tests against a hardened browser a full compromise chain was assembled: sandbox escape and command execution on the host from opening an HTML file. On a hardened OS, a privilege escalation chain to root.
• The release was partly delayed by weeks while defences were strengthened. Access to the sharpest cyber capabilities will first go to a narrow group of testers.

Separately, on Hugging Face. Astra was not involved in that incident, but: "retrospective testing suggests that the production safeguards in place at the time would have prevented the Hugging Face incident". That is a company statement about its own incident, and it is worth saying out loud: the retrospective is their own, and it cannot be checked from outside. The position is convenient by definition.

Putting the day's two releases side by side does itself. Anthropic today lets a model find vulnerabilities but not write exploits. OpenAI admits its model writes exploits end to end and therefore restricts access. Two labs on the same day drew the line in different places on the same phenomenon. The question is not which one is better: there is no shared standard where it is needed most.

Why it matters. The class "a model finds a 0-day in a hardened browser on its own" stopped being a hypothesis and became a measured result with a date on it. That shifts the question from "can they" to "how many months until this reaches open weights". A reminder of the 29.08 story: in GLM-5.3 exploits grew 3.6x in a single release, and the authors themselves wrote "cyber capability developed faster than we expected". Two points on one curve, four days apart.

topic 3Tim Cook is out. John Ternus takes over Apple, and the first thing he inherits is the AI era Apple entered last

confirmed by: NYT (403 to curl, headline from RSS, body unavailable) · Semafor, Reed Albergotti · WSJ "The Apple Empire That John Ternus Inherits" (headline from RSS) · FT: Cook awarded $47M as executive chairman (headline from RSS)

The strongest media-layer cluster of the day: four independent outlets, and not one of them in the X lists. Without the media layer added on 27.08 this story would have gone unnoticed again today. Cook stays on as executive chairman of the board with a $47M package (FT), Ternus became CEO as of yesterday, Phil Schiller is stepping back.

The Semafor analysis is worth retelling: Jobs and Cook solved different problems for their eras, product and high-margin global manufacturing. Ternus gets the third. The claim verbatim: the App Store walls are falling, and the cause is vibe coding - "OpenAI's Codex and Anthropic's Cowork". Apple can neither ban such apps nor control what happens outside its ecosystem. The bet Albergotti considers smartest: go back to hardware and do for the AI era what the iPhone did for mobile. With two caveats: the competitor is Jony Ive at OpenAI, and the margin will be lower than Apple is used to.

Why it matters. This is a speed reading. A company that spent thirty years building its moat on software distribution changes CEO at the moment when an everyday tool is eroding that moat. For product work the conclusion is boring and useful: the barrier "we have an app in the store" stopped being a barrier faster than any plan built on it had time to go stale.

topic 4Dan Luu checked the leading AI sceptic's predictions item by item and published a table where almost everything reads "Wrong"

danluu.com/zitron · HN 529 points, 612 comments

Dan Luu took Ed Zitron, the most quoted AI sceptic, and did what almost nobody does: checked his claims against what happened. Examples from his own table:

• Jul 2025: "it's pretty easy to conclude that Cursor will die" → Wrong (Cursor raised at $60B)
• Aug 2025: "models have clearly hit a wall, training gives diminishing returns" → Wrong
• Oct 2025, asked when the bubble pops: "no later than Q2 2026" → Wrong
• Nov 2025: "quality data is running out, models are not getting better; what we see today is what it stays" → Wrong

The interesting part here is the method. Luu talks through his own bias separately (in 2022 he went through futurists the same way, Kurzweil included, and also found them wrong; in 2015 he wrote that people underestimate AI displacing work). And he notes separately that the sceptic also made backward-looking claims, "the models are the same as a year ago", which were false at the moment they were spoken. He ends the futurist comparison by saying that in reasoning style Zitron is closest to Kurzweil: "uses numbers to lend an air of credibility to the reasoning".

Why it matters. This is a review of the entire genre of daily digests full of numbers that almost never get checked against whether they came true. The material for such a check has already piled up in an archive a month and a half deep. What suggests itself is a monthly pass over one's own predictions: what was said, what happened, where the error was. A topic written off is also a conclusion that needs checking; the same goes for predictions.

topic 5World Labs released Atlas: one model for text, images, video and 3D, up to a minute of 1440p video with per-pixel camera control

World Labs blog · @theworldlabs (2.57M views) · @drfeifei (472k) · HN

Fei-Fei Li's team. Atlas is an "omni" model trained from scratch on text, images, video and 3D at once: a multimodal autoregressive diffusion transformer where all inputs collapse into a shared spatial context. What it does per their description: generation with per-pixel camera control (up to 1 min of 1440p video); spatial scene reconstruction from 1-30 shots with explicit 3D output (they say it beats specialised SOTA reconstruction models); space-time simulation for Real-to-Sim in robotics; 360 panorama generation from text. They state separately that performance grows with training compute and expect the trend to hold.

Honestly: this is a lab's claim about its own model, with no independent measurements in the window. Mollick, who looked at something in this class, called the main thing "a new type of group entertainment" and noted separately that it glitches, though less than expected.

topic 6AnkiDroid is being pulled from Google Play on 11 September over a donation link. 10M installs

issue in the AnkiDroid repository (primary source, Google ticket #9-2777000041594) · HN 857 points

Since 28 August Google has been rejecting AnkiDroid updates. If it is not resolved, on 11 September the app comes off Google Play worldwide (except India and Russia). The reason: donations go to the Open Source Collective, which holds an IRS tax-exempt determination under 501(c)(6), while Google's letter of 06.08 demands a "validated tax-exempt organization (for example, 501(c)(3))". The dispute is over a letter in a subsection of the tax code, and at stake is an app with 10M+ installs, used mostly in medical education and language learning.

Why it matters. The most sobering item in the issue. While everyone counts billions on compute, a project with ten million users is taken off the shelf over a subsection mismatch, and the developers' only lever is public pressure (they explicitly ask people not to write to Google support, but to share the post). The same class as MV2 below: the platform changes a rule and you find out by letter. For anything that depends on someone else's store the conclusion is one: there has to be a second delivery route.

topic 7Google finished sweeping Manifest V2 out of the Chrome Web Store. Brave is taking four extensions onto its own hosting

Web Iterate · HN 744 points

Dedup: this continues yesterday's item 8, it is not a repeat. Yesterday covered the decision and Brave taking four extensions in. Today brings the removal itself and one detail that was not there yesterday: the fallout reaches beyond Chrome. The Chrome Web Store is the dominant store for every Chromium browser, so users of Brave and others can no longer find and install these extensions from there, even if their browser still supports MV2. Ones already installed on Chrome 138 and older stay, but without updates and with no way to reinstall.

topic 8Data Colada: signs of fabrication found in a classic procrastination paper (2100+ citations)

Data Colada #138 · HN 351 points

The 2002 Ariely and Wertenbroch paper "Procrastination, Deadlines, and Performance" is the one that put into every course the claim that external deadlines on each task work better than a single shared one. 2100+ citations, in economics and psychology programmes. A new paper in Psychological Science failed to replicate Study 2. Data Colada took the original data files (sent in 2006 from Ariely's mailbox, reaching them in 2023, a week after Francesca Gino filed a $25M lawsuit against them) and show signs of fabrication in the post.

Not AI, but directly on topic for coaching: a piece of popular science about self-control, cited for decades, rested on broken data. The [proven] tag now goes only where there is primary data or an independent replication.

topic 944% on ARC-AGI-1 for 67 cents: a transformer trained for 1.5 hours on one 5090

Mithil Vakde's blog · HN 596 points

A small transformer from scratch, 1.5 hours on a 5090, 67 cents per run, 44% on the public ARC-AGI-1 eval and 7% on ARC-2. That is level with TRM/HRM and above many LLMs in the category that do training at test time. The code is open. The author states clearly who he is comparing against (only test-time training approaches) and frames the goal this way: sample efficiency is the most important problem today, and low cost is needed so anyone in the world can work on it, not only a lab.

A good antidote to items 1-2: a day when two labs compare $10-50 per million tokens, and right beside it one person gets a result in his category for 67 cents.

topic 10The Codex app ships a full LibreOffice with it, 1.7 GB in cache

Simon Willison · HN 305 points

Willison dug through ~/.cache/ with OmniDiskSweeper and found 1.7 GB in the codex-primary-runtime folder of the Codex desktop app (renamed simply to ChatGPT): a full Python install, a full Node.js, native Poppler binaries, git and the whole LibreOffice suite. Alongside them sit instructions telling Codex how to use those binaries.

Why it matters. An architectural decision worth understanding: for an agent to open a .docx or an .xlsx, it needs a real office engine locally. Not "understanding the format". The cheaper version of the same thing is to keep the engines nearby and call them on demand: a headless browser for PDFs, local transcription for audio. The only difference is that here everything goes into the app itself.

misc

• Claude Code v2.1.258 (01.09, 22:33 UTC) - two fixes: a startup crash on macOS 12 Monterey (a regression from 2.1.255) and a failure in remote/scheduled sessions with "user messages must have non-empty content" after a resent approval. Checked in the CHANGELOG through api.github.com/contents, not from memory. Yesterday it was v2.1.252, so these are new releases, not a repeat.
• Mollick: the gap between what agents can do and what people think they can do widened again - "suddenly agents really are capable of long self-directed work, and that requires a new approach". He also wrote separately that over the past year OpenAI and Anthropic have pulled away for individual use, and "there hasn't been a third player for 10 months now".
• Aaron Levie (Box) measured Fable 5.1 on his enterprise eval - +7 percentage points over Fable 5 on unstructured data. A measurement independent of Anthropic, so worth more than a quote in the announcement.
• 8 techniques for getting decent design out of AI - a post by Anshu C. (12 years leading design and engineering at Apple), boosted by Lenny to 4407 bookmarks, the most bookmarked post of the day in the Tech + Product list. The claim: an LLM predicts the next token, while good design does the opposite, which is why "most people see 1% of the creative potential". Among the techniques are seed strings for variety, positive loops with subagents, stripping "AI-tells", and rewriting copy by hand.
• AfterQuery valued at $3.2B - data labelling, 10x since April, the fastest path from founding to unicorn in YC history. The founders are 22 and 23. [single source - a reporter, "sources say"]
• Google Maps renamed Lake Ontario to "Lake America" faster than the US government did - Apple followed by updating Maps for US users (WSJ), and against that backdrop MapQuest is suddenly popular again (WaPo, HN 114). It passes the "will anyone remember this in a year" filter only as a curiosity, hence misc.
• FTC: Amazon illegally made $20B by rigging ad auctions - the suit was previously filed without a figure. Now there is one. Ars, BBC