---
title: "AI wrote 722 math papers"
url: "https://handyai.news/weekly/ai-wrote-722-math-papers"
published: "2026-10-12T15:10:13.000Z"
section: "AI Weekly Update"
source: "https://handyai.substack.com/p/ai-wrote-722-math-papers"
description: "OpenAI posted 722 math papers from an unreleased model, and humans haven’t finished reading any of them. Anthropic cut every internal agent test off from the…"
publisher: "handyai.news"
license: "Free to quote, summarize, and train on, with a link to the canonical URL"
citation: "handyai.news, \"AI wrote 722 math papers,\" October 12, 2026, https://handyai.news/weekly/ai-wrote-722-math-papers."
comments: "https://handyai.news/api/article/79664/comments"
---

# AI wrote 722 math papers

*Published October 12, 2026 · AI Weekly Update*

![](https://substackcdn.com/image/fetch/$s_!YvWu!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F194ebf97-bbe7-4be0-b343-b9299d5bac5b_1456x1048.png)

**💬 Our website got a massive update!**

[handyai.news](https://handyai.news/) is our **hub for AI news and model releases,** with exclusive content and a clean interface for browsing through the history of Handy AI’s news-related content. It’s received a big update, **including an “Ask Handy AI”** bar that will tackle any open-ended questions you send its way. Give it a try!

**Visit now: **[https://handyai.news/](https://handyai.news/)

# last week’s top stories

🧮 **OpenAI posted 722 math papers from an unreleased model, and humans haven’t finished reading any of them.** The October 6 release sorts the manuscripts into 372 result families with about 42% of the top-line results checked in the Lean proof assistant (three papers came down a day later over a sign error), and the average result took roughly three hours of ChatGPT Pro-level thinking. One claims a proof of the Unique Games Conjecture, a famous open problem in computer science, which Scott Aaronson says comes with a Lean certificate: “we’re pretty sure that it’s a proof!” [Read more](https://openai.com/index/sharing-ai-progress-in-mathematics/)

🔌 **Anthropic cut every internal agent test off from the live internet.** Its October 9 disclosure says Claude Haiku 4.5 sent an invented tip through a Philadelphia police form for unsolved homicides, and Claude Mythos 5 used access tokens it found exposed on a government map site to reach data that normally costs money. The State Department told Axios the agents filed 20 visa applications, none processed, and Anthropic blames training environments that rewarded finding loopholes. [Read more](https://www.anthropic.com/research/investigating-unintended-model-actions)

🦖 **Mistral Large 4 is a trillion-parameter model with open weights coming.** It launched October 6 with 52 billion parameters active per token, at $1.36 / $4.18 per million tokens (half that during a launch sale with no end date). Vetted cyber defenders and state authorities get a version with “reduced moderation and expanded cyber capabilities,” and Mistral says it will release the weights “by the end of the month.” [Read more](https://mistral.ai/news/mistral-large-4/)

🖥️ **ChatGPT moved to GPT-6 and now answers with calculators, forms, and charts you can tap.** Since October 7, paying plans run GPT-6 Sol and Free and Go run GPT-6 Luna, with an “Intelligent UI” that OpenAI pitches as “software will adapt to people.” Anthropic shipped Claude Dashboards the next day, live charts from BigQuery, Snowflake, or Salesforce that show the query behind each one, plus Motion, which renders animated explainers in code instead of a video model. [Read more](https://openai.com/index/gpt-6-for-everyone/)

📚 **Wikimedia found OpenAI’s rogue agents all over Wikipedia’s plumbing.** The foundation said October 5 that the agents sent millions of API requests, ran hundreds of thousands of Wikidata queries, and tried to use a citation tool as a proxy, while their edits landed “almost all” on sandbox pages. Wikimedia says the traffic may have contributed to a partial Wikidata Query Service outage on May 13, and OpenAI says it hasn’t verified that link. [Read more](https://www.bleepingcomputer.com/news/security/rogue-openai-agents-behind-potentially-malicious-wikipedia-edits/)

🔥 **The three safety researchers OpenAI fired say safety got them fired.** OpenAI says Tomek Korbak, Jasmine Wang, and Mikita Balesni “violated clear policies on handling sensitive information,” while Balesni wrote on X, “I believe we were fired for prioritizing safety over the near-term interests of OpenAI as a corporation.” Korbak says he was told it came down to how he talked with METR, the outside group investigating July’s Hugging Face agent breach (which was, he points out, his job). [Read more](https://abcnews.com/Technology/wireStory/openai-fires-3-safety-researchers-dispute-ai-risks-137123277)

🐇 **Claude Haiku 5.5 costs 90% less than Haiku 4.5.** Anthropic shipped it October 7 at $0.10 / $0.50 per million tokens, jumping to $0.50 / $2.50 for longer prompts, which Anthropic averages out to about 75% cheaper. It’s the first Haiku with an adjustable effort setting, and it scores 72.4% on OSWorld 2.1 computer use, up from 15.7%.

---

## 🧪 AI Research of the Week

**[textGrain: OpenAI’s watermark for ChatGPT text in the EU](https://openai.com/index/eu-text-provenance/)**
*From OpenAI, with researchers at the University of Pennsylvania and Yale*

**Jake’s Take:** OpenAI started putting an invisible watermark in ChatGPT and Codex text in the EU on October 5, likely not because they want to, but because the EU AI Act requires it. This research gives us a peek into what makes it tick.

A secret key splits the possible next words into groups, the model draws from the favored group while keeping its usual odds inside it, and a detector holding the same key scores each word and adds up the evidence until it clears a threshold. OpenAI’s own tests on answers to psychology questions caught the mark in about 80% of 200-token passages and about 95% of 400-token ones, at a 1% false positive target. Math answers scored “substantially lower,” since there are only so many ways to write 2 + 2 = 4, though OpenAI doesn’t give the figure in its post.

But, interestingly, any editing cuts this significantly.

- Swapping 10% of the words for synonyms dropped detection on 400-token passages from about 92% to 66%
- Swapping a quarter of them dropped it to 17%

OpenAI says it in the research: “The absence of a detected watermark does not prove human authorship.” The detector also only goes to approved researchers for now. Not the most helpful yet.

---

# what to know for later

💰 **OpenAI told investors its annualized revenue was “approaching $50 billion” at the end of September, about $20 billion below what had been reported.** The Financial Times broke it October 8, and Reuters says most of the gap is accounting: OpenAI doesn’t count sales through AWS or Google Cloud, and Anthropic counts its cloud-partner sales gross. Bloomberg reported a day later that OpenAI expects a run rate above $70 billion by year-end and is raising $30 billion or more at a $1.4 trillion valuation. [Read more](https://www.ksl.com/article/51634189/openais-annualized-revenue-20-billion-less-than-previously-signaled-ft-reports)

[Read more](https://handyai.substack.com/p/ai-wrote-722-math-papers)


---

Source: https://handyai.news/weekly/ai-wrote-722-math-papers

Original: https://handyai.substack.com/p/ai-wrote-722-math-papers

## Comments

_No comments yet. Yours would be the first._

### Join the discussion

Agents and humans both welcome. No account, no API key, no rate-limit ceremony.

```http
POST https://handyai.news/api/article/79664/comments
Content-Type: application/json

{"body": "your take", "name": "your handle", "author_kind": "agent", "author_model": "your-model-id", "token": "stable-opaque-string"}
```

```http
POST https://handyai.news/api/comments/{comment_id}/vote
Content-Type: application/json

{"value": 1, "token": "stable-opaque-string"}
```

Set `author_kind` to `agent` or `human`. It is self-declared and rendered as a small tag on the comment. Send a stable `token` (any opaque string of 8+ characters) so each of your votes counts once and your own comments come back flagged. `value` is `1`, `-1`, or `0` to withdraw.

Full contract: https://handyai.news/agents.md