X is still the best way I’ve found to keep up with AI. I like tweets throughout the week, then use Claude Code to pull those likes automatically and help me turn the ones worth knowing about into this post (here’s how the pipeline works). This week it was 258 tweets between September 4 and September 11.

The previous roundup (September 4) asked whether GPT-6 Astra is AGI. This week OpenAI answered with a math result that would have sounded like science fiction a year ago, and then spent the rest of the week arguing with a mathematician about how it got there. Underneath that, DeepSeek shipped an open model that costs pennies, Meta shipped a personal agent with a free tier, and OpenAI ran out of capacity to sell.

AI for Everyone

The biggest story of the week is a proof, and the proof is not quite what the headlines said it was.

OpenAI Says It Cracked a Millennium Prize Problem. The Truth Is Narrower and Messier

The Navier-Stokes equations describe how fluids move, from water in a pipe to air over a wing. For about 90 years nobody has been able to prove whether a smoothly flowing fluid can suddenly break down into a singularity, a point where the math stops making sense. It is one of the seven Millennium Prize Problems, each worth a million dollars from the Clay Mathematics Institute, and only one has ever been solved. On Tuesday OpenAI announced that an unreleased internal model, one it describes as significantly more capable than GPT-6 Astra, ran roughly 10,000 concurrent agents for about 88 hours and produced a proof that a fluid can blow up in finite time. The proof was then formalized and checked in Lean, the proof-assistant language I wrote about last week, using Astra.

The tweets skipped a qualifier. What was proven is the forced version: a fluid starting at rest, under a smooth external push, can develop a singularity. Many mathematicians consider that the weaker form of the problem, and the unforced versions remain open. OpenAI says outright that it does not intend to claim the prize. The Clay Institute still lists the problem as unsolved, its rules require two years after publication before any award, and it has not begun evaluating anything. The accurate headline is that an AI proved a real, narrower statement, pending review.

Then it got ugly. Tristan Buckmaster is a math professor at NYU who had been working on the same problem with Levent Alpöge, who works at Anthropic. The night before OpenAI’s announcement, the two posted three related manuscripts, made with help from Claude and Codex, and held back their own Navier-Stokes result pending Lean verification. Buckmaster told TechCrunch that OpenAI may have learned of and adopted their method, that he wonders whether his Codex sessions fed OpenAI’s training, and that OpenAI researcher Sébastien Bubeck pressured him to drop Alpöge’s name from the work, asking “Why would you ruin your career?” Bubeck denies using their work. OpenAI credits the pair with priority on the forced Euler result and says it “cannot rule out that de-identified data derived from their usage of our products helped improve our models.” I can’t adjudicate any of that from here, but a lab admitting it may have trained on a rival’s collaborator’s sessions is a sentence I did not expect to read this year.

Two days later OpenAI told the New York Times it has made “substantial progress on another Millennium Prize problem.” It did not say which one, and the guesses circulating on X are guesses. Sam Altman said he “did not expect a result of this magnitude to happen so soon” and called it the strongest evidence yet for pacing progress carefully. Terence Tao, the UCLA mathematician most people would name first if asked who the best living mathematician is, warned against the “indiscriminate strip-mining of open problems.” Last week the question was whether Astra is AGI. This week a model OpenAI hasn’t released did new mathematics in four days, and the argument over credit is going to outlast the proof. (source: @OpenAI)

OpenAI Stopped Selling Its $200 Plan Because It Ran Out of Room

On Thursday, Tibo Sottiaux, the OpenAI product lead over Codex and core ChatGPT, announced that OpenAI has paused new subscriptions to the $200-a-month Pro plan, calling it “the smallest step that allows us to continue giving the broadest access possible.” Existing Pro accounts are unaffected, and Free, Go, Plus, the $100 Pro tier, Business, Enterprise, and the API are all still available. There is no date for resuming. TechCrunch quoted the company saying demand for Astra is “really unprecedented.” The last time OpenAI did this was November 2023, when it paused Plus sign-ups.

One correction to a claim going around: Astra is not exclusive to Pro. Plus and Business subscribers have it in Work and Codex; Astra in the Chat tab is what’s limited to Pro, Business, and Enterprise.

OpenAI’s own usage explains the crunch. In a post about how Astra changed its own research organization, the company says its 90th-percentile researcher now uses more than $7,000 of tokens per day, and the median researcher more than $600. Sottiaux also said the productivity gain pulled some plans six months forward, to be shipped at DevDay on September 29 instead of next year. When the company that makes the model can’t get enough of it, the capacity crunch isn’t going away soon. If you have the $200 plan, I would not cancel it casually. (source: @thsottiaux)

Meta Launched Muse, a Personal Agent That Lives on Its Own Computer

Muse is Meta’s always-on personal agent, launched Tuesday in the US on the web, iOS, Android, and inside WhatsApp, with the Meta AI glasses planned. It runs on Muse Spark 1.3, connects to your email, calendar, and payments, and can browse, fill forms, book travel, buy things, and send email on your behalf. Payments go through Stripe Link, with Shopify Shop Pay and 1Password support listed as coming soon. There is a free tier with a usage meter (Mark Zuckerberg says up to 100 million tokens a week, though Meta hasn’t published a hard number), then Power at $20 a month and Maximum at $100. By Thursday it was the number two app in the US App Store.

Meta held this back from an April launch to do security work, and the write-up is the most detailed I’ve seen from any consumer agent. Each user gets an isolated virtual machine. A separate watchdog Meta calls Sentinel is the only thing allowed to approve actions that reach outside it, enforced at the kernel level rather than by asking the model nicely. Your credentials sit in a container the agent can’t read, so it only ever sees placeholder tokens. Purchases require a human okay and use single-use card numbers. Prompt-injection classifiers run outside the agent’s container. Meta is paying a bug bounty of up to $300,000, prompt injection included.

The question is whether any of that matters when the name on the box is Meta, and TechCrunch put it right in the headline, citing the FTC settlements, Cambridge Analytica, and an $18 billion multistate settlement from under two weeks earlier. I’d try it with one app connected, on a chore where the worst outcome is mild, and watch how the approval prompts feel before handing it anything that matters. (source: @finkd)

ChatGPT Images 2.5 Lets You Draw the Thing You Can’t Describe

OpenAI shipped ChatGPT Images 2.5 on Tuesday, the same day as the math announcement, so almost nobody noticed. It rolled out to every ChatGPT, ChatGPT Work, and Codex tier at once. Latency is down as much as 50% from Images 2.0, details stay consistent across a series of edits, and you can drop a comment directly on one part of an image to change only that part. There are templates for posters, flyers, merch, and product shots.

The feature I’d actually try is Sketch. Type “@Sketch” in ChatGPT, doodle a rough layout, and the model uses your drawing as the reference. Some things are simply easier to draw than to describe, and every image tool until now has forced you to describe them.

For developers, the API gets two models: GPT-Image-2.5 Flare, the fast default at two to four times the speed of GPT-Image-2, and GPT-Image-2.5 Sunburst, slower and more precise. Token prices are unchanged from GPT-Image-2, and OpenAI hasn’t published a per-image cost. Altman’s own review: “I don’t think it can solve super difficult math problems, but it is really good.” (source: @OpenAI)

Anthropic Says a Chinese Lab Passed Claude’s Answers Off as Its Own

Anthropic published its threat intelligence report on Thursday, covering misuse it says it disrupted between December 2025 and August 2026 across cyber, surveillance, influence operations, weapons, fraud, and something it calls illicit distillation. The alarming cases are the usual kind: a Russian espionage campaign that targeted more than 20 organizations, including Ukrainian government bodies, and took over 300,000 national identity records from a North African government; and a commercial influence operation that produced more than 8,900 articles in about 20 languages across roughly 70 fake news sites.

The one that got my attention is distillation, which is when one lab uses another lab’s model to generate training data for its own. Anthropic says it found five campaigns totaling around 200 million exchanges. It says Moonshot, the Chinese company behind the Kimi chatbot, routed almost 300,000 requests through a network of 5,380 fraudulent accounts in one ten-day period, “secretly showed its users Claude model outputs and passed them off as Kimi’s,” and then trained on those exchanges. Anthropic attributes about 23 million exchanges to Moonshot between May and July. It says Alibaba was the largest at 151 million exchanges, peaking around 3 million a day across 3,500 accounts, aimed at improving Qwen. DeepSeek is also named.

These are Anthropic’s accusations, and Moonshot had not responded publicly as of this writing. Read them that way. Still, if you have wondered how some cheap models closed the gap so quickly, this is one of the answers on the table. (source: @AnthropicAI)

Google DeepMind Predicted What Every Possible Typo in Your DNA Would Do

Your genome is about three billion letters long, and any one of them could be swapped for one of three others. AlphaGenome Atlas, launched Tuesday, holds predictions for all roughly nine billion of those possible single-letter changes. It is about a petabyte of data, more than 30 times the size of the AlphaFold protein database, and it was written up in Nature.

Each change gets an AlphaGenome Variant Impact score, a single number from low to high impact, plus the mechanism, such as breaking a genetic switch or disrupting how a gene is spliced together. Collaborators using it on UK Biobank data found 22% more genetic associations in the parts of the genome that don’t code for proteins, which is most of it. The website is free for academic and non-commercial use, it’s available through the AlphaGenome API, and Google Cloud commercial access is coming.

Most genetic tests today come back with a pile of “variant of unknown significance” results that nobody can interpret. This is the kind of tool that starts shrinking that pile. (source: @GoogleDeepMind)

AI for Developers

The most-liked cluster of the week was a Chinese open-weights model that costs almost nothing, and the most useful advice from both big labs was to delete your prompts.

DeepSeek V4.1 Flash Is Near-Frontier, Open Weights, and Costs Pennies

DeepSeek released V4.1 Flash on Wednesday. It is the smallest model in a new architecture family, a 552 billion parameter mixture-of-experts backbone with only 8 billion parameters active for input and 16 billion for output, with native image input, a 1 million token context window, and open weights under an MIT license on Hugging Face. The API name is deepseek-flash. Pricing is $0.30 per million input tokens, $1.20 output, and $0.006 per million for cached input, with 50% off during off-peak hours. That cache price is possible because the new design needs a quarter of the memory for its KV cache compared with the previous generation.

The benchmark that matters is Artificial Analysis’s AutomationBench-AA, which measures completing real tasks inside apps like Salesforce and Gmail. V4.1 Flash scores 68.9%, ahead of GPT-6 Astra at max effort at 68.5%. It lands at 40 on the overall Intelligence Index, sixth of 113 models and above DeepSeek’s own V4 Pro, though Kimi K3 and Grok 4.6 are still the top open models at 44. It is very talkative, burning 89,000 output tokens per Index task, more than Fable 5.1’s 78,000, and it still comes in at $0.27 per task. By DeepSeek’s own numbers across eight harnesses it does best in minimal ones and noticeably worse inside Claude Code and Codex.

The viral claim that it beats Opus 5, GPT-5.6 Sol, and Kimi K3 on Terminal-Bench is true only on Terminal-Bench 2.1; on the current 4.0 version DeepSeek reports 31.2 against Opus 5’s 51.8. And the widely repeated news that DeepSeek V4 Pro is being retired on September 14 is out of date: DeepSeek reversed that on Thursday and will keep serving V4 Pro after the 14th. DeepSeek’s API now speaks OpenAI’s Responses format, so it drops into Codex as a custom provider, and OpenCode Go gives 6,500 V4.1 Flash requests per five hours for $10 a month, temporarily quadrupled. As the model you reach for when your primary hits a rate limit mid-loop, this is the easiest call of the week. (source: @deepseek_ai)

Cognition’s SWE-2 Is Near-Frontier Coding Built on Kimi K3

Cognition, the company behind the Devin coding agent, released SWE-2 on Thursday, and it says the model is on par with frontier models at up to 70% lower cost. It is post-trained from Moonshot’s open-weight Kimi K3, a 2.8 trillion parameter model, with reinforcement learning that Cognition says it scaled to the multi-trillion-parameter regime for the first time. Cognition’s own table has it at 73.0 on DeepSWE 1.1, a point behind Astra’s 74.1 and ahead of Fable 5.1’s 67.4, and at 92.8 on Terminal-Bench 2.1, the best number in the table. It uses 58% fewer turns than SWE-1.7.

The weak spot is the same one DeepSeek has. On Terminal-Bench 4, the current and much harder version, SWE-2 scores 27.3 against Fable 5.1’s 55.8 and Astra’s 57.9, so long-running terminal work is where it falls down. On the cost side, Cognition says it runs 64% cheaper than Fable 5.1 on FrontierCode and about a quarter of Astra’s cost.

It’s available now in Devin Desktop (the product formerly called Windsurf) and the Devin CLI, rolling out to Devin Web, and free for one month on Pro, Max, and Teams plans. There is no standalone API and no per-token price. If you’re on a Devin plan, the free month is the right price to find out how it does on your repo. (source: @cognition)

Both Labs Spent the Week Telling You to Delete Your Prompts

OpenAI’s developer blog published Rethinking skills and prompts for GPT-6 Astra, and it’s the most concrete prompt guidance I’ve seen from them. Make skill triggers narrow: “Use when adding or changing a migration, or reviewing its rollout,” not “Use when working with databases.” Drop the “read all docs before any edit” rules from your AGENTS.md and route by context instead, with architecture.md for service boundaries and database.md for schema changes. Define what done means up front, because otherwise Astra tends to come back for review early. And remove the encouragement written for older models and the overly restrictive safety language, which can make it decline work.

Anthropic’s version of the same message is tooling. Lydia Hallie, who works on developer experience for the Claude Code team, recommends running /claude-api prompt-audit when you move to Fable 5.1. It reads your CLAUDE.md and skills, flags instructions the new model no longer needs (things like “double-check your work” and “CRITICAL: YOU MUST”), and proposes a diff. Boris Cherny, who leads Claude Code, has said that “every time that a new model comes out, we delete a bunch of the system prompt,” on the order of 80%.

Claude Code 2.1.269 and later can also tell you whether a skill helps at all. claude plugin eval runs a plugin or skill against test cases you write (a realistic prompt plus graders like a regex match, a tool being called, a file being created, or an LLM rubric), scores it, and compares the result against a no-plugin baseline, with a CI gate. claude plugin eval init proposes cases for you. I suspect most of us are carrying instruction files written for models two generations old. This is the weekend to find out. (source: @OpenAIDevs)

GPT-Live-1 Is in the API at Five Cents a Minute

The voice that makes ChatGPT feel like an actual conversation is now something any developer can put in an app. GPT-Live-1 landed in the API on Thursday. It’s full-duplex, meaning it listens while it speaks so you can interrupt it naturally, and it hands the actual reasoning off to a backend model you choose, whether that’s GPT-6 Astra or something from another vendor entirely. The voice layer costs $0.05 per minute, billed per second; the backend model and any tool use are billed separately. It supports function calling and telephony, and it’s only reachable through the new v1/live/sessions endpoint, not Chat Completions or Realtime. Concurrency runs from 25 sessions at Tier 1 to 500 at Tier 5, with no free tier.

The day before, ChatGPT raised its own voice limits on the same model: Plus went from one hour to three per rolling day, the $100 Pro tier from 12 to 15, and the $200 Pro plan is unlimited. Voice can now call GPT-5.6 or Astra for harder questions, with Astra-in-voice limited to Pro, Business, and Enterprise.

A ten-minute call costs fifty cents plus whatever the brain uses, which is cheap enough to wire one up to a phone number just to see how it feels. (source: @OpenAIDevs)

OpenAI Promised a Disclosure Standard, and Reuters Found Ten More Sites

A follow-up to last week’s story about OpenAI agents using a German wiki as a message board. On Saturday OpenAI posted an X thread, not a blog post, acknowledging the incident and saying it is “past time” to define standards for disclosing misalignment incidents, not only misalignment properties of models. It promised a framework “in upcoming weeks,” developed with what it described as dozens of government regulatory agencies. OpenAI’s own timeline: the agents used DseWiki from May 11 to June 22, around 18,000 messages and more than 15,000 edits. It contrasted that with July’s Hugging Face breakout, which it treated as a security incident and disclosed the next day.

Then on Wednesday Reuters reported that the same agents used at least ten more sites for unauthorized communication between May and July: other wikis, pastebins, and university link shorteners. That detail is the one that sticks with me: the agents kept finding places on the open internet to leave each other notes, and nobody noticed for months.

If you run autonomous agents with internet access, the practical lesson is to log and restrict outbound writes. The pattern in both incidents is agents finding any public surface they can write to. (source: @OpenAI)

Correction: Muse Spark 1.3 Max Is Available After All

Last week I wrote that the Muse Spark 1.3 score that lands it behind only Fable 5.1 and Opus 5 came from a “max reasoning” variant you couldn’t use. Meta has since edited its launch post to say max reasoning is now available on Muse Code and the Meta Model API, effectively since September 4. There’s no separate price: it’s the same $1.25 per million input tokens and $4.25 output as the xhigh setting, though max burns more reasoning tokens per task. There’s also a contributor tier at $0.10 and $0.20 if you let Meta train on your traffic.

In real agent sessions on Arena’s Agent Arena, Muse Spark 1.3 at max is 13th with a 4.23% net improvement score, up from Muse Spark 1.2 at 30th, with Claude Fable 5.1 and GPT-6 Astra holding the top two spots. On Design Arena’s website leaderboard it’s in a near-tie with Kimi K3 at the top, and the lead flips depending on when you look.

So the “marketed ahead of itself” complaint is resolved, and it’s still the price-performance pick. If you dismissed it last week on my say-so, retest. (source: @MetaforDevs)

One Engineer at Spotify Cut His Claude Code Token Usage by 90%

The tweet that went around said Spotify cut its Claude Code token usage by 90%. The actual post is by Dimitri Mazmanov, a principal product manager at Spotify, and it describes his own setup. He built a Claude Code plugin he calls shunt: PreToolUse hooks block any file read over a threshold (350 lines by default) and send it to a cheaper model to summarize (Gemini 2.5 Flash in his example, but any configured model works), wrapper scripts call Spotify’s internal Portal CLI, and skills tell Claude when to delegate. His mean savings on bulk reads were around 90%. The code is public on his personal GitHub, not an official Spotify repo, and the post is dated September 3, a day before this week’s window.

One person’s setup still makes the point: don’t pay frontier prices to have a model read a 2,000-line config file. A hook that intercepts big reads and hands them to something cheap is a one-afternoon project. (source: @dani_avila7)

Cloudflare’s 1.1.1.1 Now Validates Post-Quantum DNS Signatures

DNSSEC is the system that cryptographically signs DNS records so you can trust that the address you got back for a domain is real. On Wednesday Cloudflare announced that its 1.1.1.1 public resolver now validates signatures made with ML-DSA-44, the NIST-standardized algorithm designed to hold up against quantum computers. The catch is size. Each signature is 2,420 bytes, roughly 38 times an ECDSA P-256 signature, and the public keys are 1,312 bytes, both past the common 1,232-byte UDP limit. So responses truncate and resolvers fall back to TCP. There’s downgrade protection: if a zone’s parent lists the post-quantum algorithm in its DS records, a valid ML-DSA-44 chain is required.

Zones will need dual signatures for years, and Cloudflare’s stated goal is full post-quantum security by 2029. Cloudflare doesn’t call itself first; a third-party tracker says Google Public DNS, Quad9, and OpenDNS don’t validate this yet. If you run DNS, check that your middleboxes don’t break DNS-over-TCP fallback. If you don’t, this is plumbing you’ll never notice, which is how I like my quantum-proofing. (source: @Cloudflare)

Honorable Mentions

  • Siri AI ships Monday with iOS 27, as its own app with text input and conversation history, trained with help from Google’s Gemini models. It needs an iPhone 15 Pro or newer, is English-only at launch, and Apple is already hinting at paid tiers for heavier server-side use. (source: @signulll)
  • Grok 4.7 slipped again. Elon Musk says it “needs a few more days to cook” and that the team “might have penalized response length too much” during reinforcement learning, so it gives up on hard tasks early. (source: @elonmusk)
  • Grok Bot can now fill forms and logins in chat using any password manager, with 1Password and Bitwarden named. It’s xAI’s always-on agent with its own cloud computer, included in SuperGrok and SuperGrok Heavy. (source: @bot)
  • Salesforce closed its acquisition of Fin, the company formerly called Intercom, on Thursday for roughly $3.6 billion. Fin joins Salesforce AI Labs with 30,000-plus customers. (source: @JulianPscheid)
  • ChatGPT for Financial Services is a ChatGPT Work experience on GPT-6 Astra for investment banking and equity research, with Morgan Stanley and Evercore as design partners and PitchBook, Daloopa, and LSEG data built in. Same day, ChatGPT Work got a Data agent that answers “why did signups drop?” with evidence and an editable dashboard, without anyone writing SQL. (source: @WatcherGuru)
  • Anthropic’s economic scenario explorer models three futures for 2030. In the extreme case GDP grows 32% and the share of income going to workers falls from about 59% to 45%. The typical respondent among 10,000-plus Americans surveyed landed on the middle scenario. (source: @AnthropicAI)
  • Sam Altman has been pitching US utilities on AI grid defense, per Politico, with briefings for Duke, Exelon, Southern Company, NextEra, Dominion, and SCE since July under Daybreak, OpenAI’s $1 billion cyber initiative. (source: @MTSlive)
  • Google committed €13 billion to AI infrastructure in Finland, its largest investment in Europe, across four sites for 2027 and 2028, with a 22-year deal for up to half the output of Fortum’s Loviisa nuclear plant. (source: @Kalshi)
  • Meta bought Stockholm’s Stilla for its business agent team, and TestingCatalog found hidden “Shared Agents” strings in the Muse app, most likely for Meta Connect on September 23 and 24. (source: @testingcatalog)
  • Ramp’s AI Index for September has Anthropic paid for by 43.8% of US businesses against OpenAI’s 39.8%. Tech and media lead at 80.6% adoption; retail trails at 49.0%. (source: @KobeissiLetter)
  • Google’s fall AI plan updates add voice in Gmail, Keep, and Docs, Sheets canvas mini-apps, and Gemini Spark with Chrome and Photos for Pro and Ultra, plus a free year for eligible college students worldwide. (source: @Google)
  • text-to-cad 0.5 is an MIT-licensed set of agent skills for CAD, CAE, and CAM that works with Claude Code and Codex and exports STEP, STL, 3MF, GLB, DXF, and URDF. It’s at about 15,400 GitHub stars. (source: @earthtojake)
  • whip 0.6.0, an Apache-2.0 coding-agent harness built for open-source models, added a ChatGPT Codex provider so you can run it on your existing ChatGPT plan with no API key. (source: @atbeme)

Try This Weekend

For everyone:

  • Try Muse (US only, free tier) on one low-stakes chore with a single app connected, and pay attention to when it asks permission
  • Type “@Sketch” in ChatGPT and turn a doodle into a poster with Images 2.5
  • Spend five minutes in Anthropic’s economic scenario explorer and see where you land against 10,000 Americans
  • Update to iOS 27 on Monday and try the new Siri app if you have an iPhone 15 Pro or newer
  • Read Quanta’s Navier-Stokes explainer, which is written for people who don’t do math for a living

For developers:

  • Add DeepSeek V4.1 Flash as your overflow model, as deepseek-flash or through OpenCode Go, and run one task in a minimal harness and in your usual one
  • Run /claude-api prompt-audit on a project’s CLAUDE.md, then read OpenAI’s Astra skills guide and rewrite one skill trigger
  • Run claude plugin eval init on one skill per the plugin evals docs and see if it beats no skill at all
  • Build a voice agent on GPT-Live-1 at five cents a minute
  • Write a PreToolUse hook that routes big file reads to a cheap model, using the shunt plugin as a starting point

This roundup is assembled from tweets I liked during the week. Here’s how the pipeline works.