Sophos cut threat investigation time 96% with OpenAI Daybreak
Sophos reduced cyber-threat investigation time by 96 percent and automated 52 percent of cases using OpenAI Daybreak.
OpenAIRead the briefOpenAI
Top stories
Daily briefing on generative AI
AI briefingSaturday, 10 October
Listen to today's briefing3 min · AI voice by ElevenLabs
Updated · 12 stories · 8 sources · 2 min read
OpenAI's GPT-6 Astra cut model costs by 76 times in Asana's browser tests, while llama.cpp released build 11535 and Anthropic shipped Python SDK version 1.13.0 with new analytics types.
Step 1 · 30 seconds
Summary
- Asana reports GPT-6 Astra made browser agent 76 times cheaper, 5 times faster
- OpenAI's ttok tool switched default tokenizer from GPT-4 to GPT-5/GPT-6
- llama.cpp build 11535 ships with benchmark and attestation improvements
- Anthropic Python SDK 1.13.0 adds Chat analytics and Managed Agents filtering
- Book publishers using Claude and ChatGPT for publicity copy without disclosure
Step 2 · 3 min
Listen
Voiced by ElevenLabs from today's briefing.
Read the transcript
Good morning. This is dailybrief, and here are today's top stories in generative AI, for Saturday, October 10. Asana has cut model costs dramatically using OpenAI's GPT-6 Astra. OpenAI said the work management company made its browser agent 76 times cheaper and 5 times faster in tests using the model in Codex. The results suggest GPT-6 Astra offers major cost savings for automation tasks compared to earlier models. In other news, the ttok token-counting tool has shipped version 1.0. Simon Willison said the release switches the default tokenizer from GPT-4 to GPT-5 and GPT-6. OpenAI has not confirmed that GPT-6 uses the same tokenizer as GPT-5, but testing by William Liu found seven GPT models report identical token counts of 44,794 on the same text. Turning to infrastructure, llama.cpp released build 11535. The update includes changes to benchmark behavior and new attestation links. The open-source inference engine supports macOS, Linux, Windows, Android and openEuler platforms. Meanwhile, Anthropic released Python SDK version 1.13.0. The update adds types for Chat and Cowork unified analytics metrics. Managed Agents also gain workflows and thread status filtering, along with multiagent configuration types. Also today, Sophos reported major efficiency gains from OpenAI's Daybreak model. OpenAI said the cybersecurity company reduced cyber-threat investigation time by 96 percent and automated 52 percent of MDR cases while preserving human oversight. Simon Willison built a Django blog feature using voice commands to GPT-6 Astra High in Codex. He said the model built database models, views and templates over 30 minutes while he cooked dinner. The feature required additional typing for private repository API work, showing voice coding works for simple features but still needs a keyboard for details. The Allen Institute for AI has replaced its priority-based GPU scheduler. Hugging Face reported the lab now uses GPU time budgets and hierarchical fair-share allocation. The institute manages thousands of H100, B200 and B300 GPUs, with demand running two to three times available capacity. And finally, three major book publishers are quietly using AI for writing tasks. WIRED reported that HarperCollins, Simon & Schuster and Hachette are using Claude and ChatGPT for publicity copy and marketing without public disclosure, according to interviews with more than two dozen workers. HarperCollins purchased Claude, ChatGPT and Jasper licenses and assigned staff as AI Champions in brainstorming sessions. That's dailybrief for today. Every story is linked to its original source on our website.
Podcast Two-host conversation with Jessica & Eric · 6 min
Read the podcast transcript
Eric: Okay, so the big one today... Sophos cut threat investigation time 96% with OpenAI Daybreak.
Jessica: Yeah, we'll get right into that one.
Jessica: This is dailybrief, I'm Jessica.
Eric: And I'm Eric. It's Saturday, 10 October 2026, and we've got 12 stories.
Eric: So what actually changed?
Jessica: Sophos reduced cyber-threat investigation time by 96 percent and automated 52 percent of cases using OpenAI Daybreak.
Eric: Right. So for developers, demonstrates enterprise gains from applying frontier models to security operations.
Jessica: Next up. Hugging Face has this one: Ai2 replaces priority scheduler with GPU time budgets and fair-share.
Eric: Wait, what's the detail there?
Jessica: The Allen Institute for AI replaced priority-based scheduling with GPU budgets and hierarchical fair-share allocation.
Eric: Okay, that's interesting. So for developers, offers a model for labs managing scarce GPU resources among competing research teams.
Jessica: Moving on. Simon Willison's Weblog has this one: Quoting Matthew Green.
Eric: Okay, walk me through it.
Jessica: Everyone is very concerned about being respectable, so I’m going to be the goofball who raises worst-case possibilities.
Jessica: Next up. OpenAI has this one: Asana cut model costs 76 times with GPT-6 Astra in browser agent tests.
Eric: Huh. What do we know so far?
Jessica: Asana made its browser agent 76 times cheaper and 5 times faster using GPT-6 Astra in Codex, OpenAI said.
Eric: Makes sense. So for developers, shows GPT-6 Astra offers major cost savings for automation tasks compared to earlier models.
Jessica: Moving on. OpenAI has this one: How Oracle turns days of work into minutes with ChatGPT and Codex.
Eric: And is that out now, or just announced?
Jessica: Across recruiting, engineering, and operations, Oracle turns specialist knowledge into fast, repeatable workflows with ChatGPT Work and Codex.
Jessica: Next up. Simon Willison's Weblog has this one: Simon Willison builds blog feature using GPT-6 Astra voice mode in Codex.
Eric: So what actually changed?
Jessica: Simon Willison built a Django blog feature using voice commands to GPT-6 Astra High in Codex while cooking.
Eric: Right. So for developers, shows voice coding works for simple features but still needs keyboard for details.
Jessica: Moving on. Simon Willison's Weblog has this one: ttok 1.0 switches default tokenizer from GPT-4 to GPT-5/GPT-6.
Eric: Wait, what's the detail there?
Jessica: The ttok token-counting tool shipped version 1.0 with GPT-5/GPT-6 as the default tokenizer, Simon Willison said.
Eric: Okay, that's interesting. So for developers, developers need to ensure token counts match the model they're using for accurate pricing.
Jessica: Next up. MIT Technology Review has this one: MIT Technology Review to host Samuel King talk on AI-designed viruses.
Eric: Okay, walk me through it.
Jessica: MIT Technology Review will host a talk with Stanford PhD student Samuel King on October 16 about AI-designed viruses.
Eric: Huh. So for developers, highlights AI capability to design biological structures, raising both opportunity and safety questions.
Jessica: Moving on. WIRED has this one: HarperCollins, Simon & Schuster and Hachette use AI for publicity copy.
Eric: Huh. What do we know so far?
Jessica: Three major publishers are using Claude and ChatGPT for publicity copy and marketing without public disclosure, WIRED reported.
Eric: Makes sense. So for developers, shows enterprise adoption of AI for writing tasks despite staff and ethical concerns.
Jessica: Next up. llama.cpp releases has this one: llama.cpp build 11535 released with benchmark and attestation updates.
Eric: And is that out now, or just announced?
Jessica: The llama.cpp project released build 11535 with changes to benchmark behavior and new attestation links.
Eric: Yeah, fair. So for developers, keeps the popular local inference engine current for developers running models on their hardware.
Jessica: Moving on. TechCrunch has this one: Andreessen Horowitz sees consumer AI opportunity beyond subscriptions.
Eric: So what actually changed?
Jessica: Olivia Moore of Andreessen Horowitz said consumer AI needs revenue streams beyond subscriptions, according to TechCrunch.
Eric: Right. So for developers, highlights the challenge of making consumer AI profitable with current economics.
Jessica: And finally. Anthropic Python SDK releases has this one: Anthropic Python SDK 1.13.0 adds Chat analytics and agent filtering.
Eric: Wait, what's the detail there?
Jessica: Anthropic released Python SDK version 1.13.0 with Chat analytics types and Managed Agents thread filtering.
Eric: Okay, that's interesting. So for developers, gives developers new metrics and controls for building and monitoring agent applications.
Eric: That's the show for today.
Jessica: Every story is linked to its source on the dailybrief website. See you tomorrow.
Step 3 · 2 minutes
Top 3 stories
Sophos reduced cyber-threat investigation time by 96 percent and automated 52 percent of cases using OpenAI Daybreak.
- Investigation time decreased 96 percent using Daybreak
- 52 percent of MDR cases now automated
- Human oversight preserved according to OpenAI
Explain it simply
In plain words A cybersecurity company used OpenAI's Daybreak model to speed up threat analysis and handle more than half of cases automatically.
Term to know MDR — Managed Detection and Response, outsourced security monitoring service
What it means for developers: Demonstrates enterprise gains from applying frontier models to security operations.
The Allen Institute for AI replaced priority-based scheduling with GPU budgets and hierarchical fair-share allocation.
- Manages thousands of H100, B200 and B300 GPUs
- Demand runs two to three times available GPU capacity
- New system shifts allocation from case-by-case to transparent budgeting
Explain it simply
In plain words A research lab changed how it assigns computing time so decisions are based on transparent budgets instead of urgency levels.
Term to know fair-share allocation — system that divides resources based on agreed quotas, not requests
What it means for developers: Offers a model for labs managing scarce GPU resources among competing research teams.
Everyone is very concerned about being respectable, so I’m going to be the goofball who raises worst-case possibilities.
- — Matthew Green, on Twitter.
Today's newsmakers
Who's making AI news
The companies that feature most in today's stories, ranked by how often and how prominently they appear.
#1OpenAI
7 stories today
Steady- Sophos cut threat investigation time 96% with OpenAI Daybreak
- Asana cut model costs 76 times with GPT-6 Astra in browser agent tests
- How Oracle turns days of work into minutes with ChatGPT and Codex
+4 more stories
#2Anthropic
2 stories today
▲ 1+1 more story
#3Oracle
1 story today
▲ 2
#4Hugging Face
1 story today
▼ 2
#5GitHub
1 story today
New today
Also in the news
Company names and logos are trademarks of their owners, shown only to identify them.
Open source · for developers
Trending AI repos on GitHub
New repositories gaining stars fastest over the last 30 days. Source: GitHub.
- KKKKhazix/AIHOT
AIA framework that automatically finds trending topics and writes daily reports, customizable with your own sources and curation standards.
- rehan-remade/universal-modder
AIPython tools that let Claude analyze and modify PC games through reconnaissance and reverse engineering techniques.
- yetone/magpie
AIA menu bar application that lets you run different AI coding agents on various models from one interface.
- LosaLosSantos/aurelio-finance
AIPersonal finance application with AI advisor for tracking net worth, investments, ETFs, cash and debts locally on your computer.
- jev-chat/jev-chat-jarvis
AIAn Android assistant that reads chat messages, assesses intent and risk, then suggests replies you can send with one tap.
- Louis-CFM/coucou
AIA Mac and iPhone companion app that monitors your AI coding agents including Claude Code, Codex, Cursor, and others.
- shader-effects-inc/shaders
AIWebGPU shader components compatible with React, Vue, Svelte, Solid, vanilla JavaScript, and Framer.
- Ebony-Vinyl/dsh-our-free-model
AIA plugin for dsh that provides free access to AI models including DeepSeek and Kimi without login or API keys.
Optional · if you have more time
More stories
9 more stories
OpenAI
Asana cut model costs 76 times with GPT-6 Astra in browser agent tests
Asana made its browser agent 76 times cheaper and 5 times faster using GPT-6 Astra in Codex, OpenAI said.
- GPT-6.1
- GPT-6
- Browser agent became 76 times cheaper in tests using GPT-6 Astra
- Performance improved 5 times faster according to OpenAI
- Tests conducted using OpenAI Codex environment
Explain it simply
In plain words Asana tested OpenAI's GPT-6 Astra model in a browser automation tool and found it cut costs by 76 times while running 5 times faster than before.
Term to know browser agent — software that automates tasks in a web browser
What it means for developers: Shows GPT-6 Astra offers major cost savings for automation tasks compared to earlier models.
New Models
OpenAI
How Oracle turns days of work into minutes with ChatGPT and Codex
Across recruiting, engineering, and operations, Oracle turns specialist knowledge into fast, repeatable workflows with ChatGPT Work and Codex.
New Models
Simon Willison's Weblog
Simon Willison builds blog feature using GPT-6 Astra voice mode in Codex
Simon Willison built a Django blog feature using voice commands to GPT-6 Astra High in Codex while cooking.
- GPT-6
- Used voice conversation mode with ChatGPT desktop Codex tab
- GPT-6 Astra High built models, views, templates over 30 minutes
- Required additional typing for private repository API work
Explain it simply
In plain words A developer spoke his feature requirements aloud and the AI wrote the database, views and templates code while he cooked dinner.
Term to know Django — Python web framework for building database-backed websites
What it means for developers: Shows voice coding works for simple features but still needs keyboard for details.
Developer Tools
Simon Willison's Weblog
ttok 1.0 switches default tokenizer from GPT-4 to GPT-5/GPT-6
The ttok token-counting tool shipped version 1.0 with GPT-5/GPT-6 as the default tokenizer, Simon Willison said.
- GPT-4
- GPT-5
- GPT-6
- Version 1.0 changes default from GPT-4 to GPT-5/GPT-6 tokenizer
- OpenAI has not confirmed GPT-6 uses same tokenizer as GPT-5
- Seven GPT models report identical 44,794 tokens in tests
Explain it simply
In plain words A token-counting tool updated to assume users want GPT-5 or GPT-6 tokenization by default instead of the older GPT-4 method.
Term to know tokenizer — tool that splits text into chunks that language models process
What it means for developers: Developers need to ensure token counts match the model they're using for accurate pricing.
New Models
MIT Technology Review
MIT Technology Review to host Samuel King talk on AI-designed viruses
MIT Technology Review will host a talk with Stanford PhD student Samuel King on October 16 about AI-designed viruses.
- King used generative AI to propose virus genetic blueprints
- Event scheduled for October 16 at 18:30 BST
- Named one of MIT Technology Review Innovators Under 35
Explain it simply
In plain words A Stanford researcher will discuss how he used AI to design microscopic viruses in a live online talk.
Term to know genetic blueprint — DNA or RNA sequence that defines an organism's structure
What it means for developers: Highlights AI capability to design biological structures, raising both opportunity and safety questions.
New Models
WIRED
HarperCollins, Simon & Schuster and Hachette use AI for publicity copy
Three major publishers are using Claude and ChatGPT for publicity copy and marketing without public disclosure, WIRED reported.
- HarperCollins purchased Claude, ChatGPT and Jasper licenses
- Staff were assigned as AI Champions in brainstorming sessions
- Multiple books canceled for blatant AI use never made news
Explain it simply
In plain words Book publishers are quietly using language models to write book descriptions and marketing materials, according to anonymous staff interviews.
Term to know publicity copy — text used to describe and promote books to readers and media
What it means for developers: Shows enterprise adoption of AI for writing tasks despite staff and ethical concerns.
New Models
llama.cpp releases
llama.cpp build 11535 released with benchmark and attestation updates
The llama.cpp project released build 11535 with changes to benchmark behavior and new attestation links.
- Qwen3.6
- Qwen3.8
- Build 11535 includes llama-bench benchmark size adjustments
- Release provides attestations at GitHub
- Supports macOS, Linux, Windows, Android and openEuler platforms
Explain it simply
In plain words The open-source llama.cpp inference engine released a new build with updates to its testing tool and support across many platforms.
Term to know attestation — cryptographic proof that software has not been tampered with
What it means for developers: Keeps the popular local inference engine current for developers running models on their hardware.
New Models
TechCrunch
Andreessen Horowitz sees consumer AI opportunity beyond subscriptions
Olivia Moore of Andreessen Horowitz said consumer AI needs revenue streams beyond subscriptions, according to TechCrunch.
- ChatGPT dominates, but Suno and ElevenLabs show staying power
- Only 2.2 percent of US households pay for AI
- Moore prefers monetization that does not require consumer subscription dollars
Explain it simply
In plain words A venture capital partner says consumer AI companies must find ways to make money without charging users monthly fees.
Term to know prosumer — professional consumer, user between casual and enterprise
What it means for developers: Highlights the challenge of making consumer AI profitable with current economics.
Pricing & Tokens
Anthropic Python SDK releases
Anthropic Python SDK 1.13.0 adds Chat analytics and agent filtering
Anthropic released Python SDK version 1.13.0 with Chat analytics types and Managed Agents thread filtering.
- Version 1.13.0 adds Chat and Cowork unified analytics metrics
- Managed Agents gain workflows and thread status filtering
- Includes multiagent configuration types
Explain it simply
In plain words Anthropic updated its Python library with new ways to measure Chat performance and filter agent workflows.
Term to know Managed Agents — Anthropic feature for deploying and monitoring autonomous AI agents
What it means for developers: Gives developers new metrics and controls for building and monitoring agent applications.
Developer Tools
Last step · 1 minute
Daily quiz
5 questions on today's stories
Browse other days
How we work
Editorial standards
Primary sources first
Lab announcements, release notes and SDK changelogs come before commentary. Every story links to the original.
Only what the source says
Figures, model names and prices are taken from the published text. Where a detail is not stated, we leave it out.
Written for developers
Each story states what changed, by how much, and what it means in practice. No hype, no speculation.
Quick answers
What happened in AI on Saturday, 10 October 2026?
OpenAI's GPT-6 Astra cut model costs by 76 times in Asana's browser tests, while llama.cpp released build 11535 and Anthropic shipped Python SDK version 1.13.0 with new analytics types. 12 stories from 8 sources are summarised on this page.
What are the top AI stories today?
1. Sophos cuts threat investigation time by 96% with OpenAI Daybreak. 2. Impactful scheduling for GPU clusters. 3. Quoting Matthew Green.
Which AI topics does today's briefing cover?
New Models (8), Pricing & Tokens (1), Developer Tools (3).
Where does the information come from?
Headlines and short summaries come from the public feeds of 8 publishers, including Anthropic Python SDK releases, Hugging Face, MIT Technology Review, OpenAI, Simon Willison's Weblog, TechCrunch. Every story links to the original article.
Is there an audio version of this briefing?
Yes. A voice recording covering every story is at the top of this page and is published each day.
Sources
Every headline links to the original. How it works
- OpenAI 3
- Simon Willison's Weblog 3
- Anthropic Python SDK releases 1
- Hugging Face 1
- MIT Technology Review 1
- TechCrunch 1
- WIRED 1
- llama.cpp releases 1