TRD An inside look at Anthropic’s AI speedup
Posted: Fri Nov 06, 2026 11:00 am
.bh__table, .bh__table_header, .bh__table_cell { border: 1px solid #C0C0C0; } .bh__table_cell { padding: 5px; background-color: #FFFFFF; } .bh__table_cell p { color: #2D2D2D; font-family: 'Helvetica',Arial,sans-serif !important; overflow-wrap: break-word; } .bh__table_header { padding: 5px; background-color:#F1F1F1; } .bh__table_header p { color: #2A2A2A; font-family:'Trebuchet MS','Lucida Grande',Tahoma,sans-serif !important; overflow-wrap: break-word; } Read Online | Sign Up | AdvertiseGood morning, {{ first_name | AI enthusiasts }}, and welcome to our 4,687 new readers. Anthropic has made “pacing the frontier” part of its pitch, holding back on what it hands the public. Inside the lab, the pace appears to be a different story.After 11 weeks inside the lab’s data, METR estimates Anthropic’s progress has roughly doubled, with its own AI driving most of the jump and no clear bottlenecks to slow it down. When the public gets to see the models behind it is another question.In today’s AI rundown:
METR’s inside look at Anthropic’s AI speedup
Image source: METRThe Rundown: AI-evaluation nonprofit METR just published a new report on its findings from 11 weeks of inside access to Anthropic’s internal data and interviews with staff, concluding the company’s AI progress is likely running about 2x as fast as before, mostly due to its own AI — though Anthropic redacted most of the strongest evidence.The details: 
Your favorite notetaker now does the typing too
The Rundown: Granola, the AI notepad for people with back-to-back meetings, just launched Voice Typing. Hold Fn and talk, and clean text appears in any app, whether you’re replying to an email or writing a prompt.With Voice Typing, you can:
The ‘Three Times’ rule
Image source: The Rundown @ Google I/ORowan: I have a rule before I’m allowed to automate anything: I have to do it manually at least three times first.Every automation I’ve shared here has followed it. I ran my Friday “factory inspections” by hand for months before handing them to an agent. I sorted my contacts and wrote my own follow-ups before my agents took those over. The manual reps were the training data for my own judgment.When you do a task yourself a few times, you learn what good looks like, which steps actually matter, and where things tend to go wrong (that’s what also lets you check an agent’s work later instead of blindly trusting it).It also fixes the most common mistake I see when people start building with AI, which is going into the tools looking for problems to solve. Flip it. Find the real problems in your life and work first, do them manually three times, then (and only then) automate them. The time you free up will surface the next problem, and then you repeat.That loop is how real AI productivity happens. It’s never one tool automating everything at once; it’s an iterative process that stacks, and six months in you’re a different operator.So before you build the agent, do the task yourself on Monday, again on Wednesday, and again on Friday. Then automate the version you actually understand.AI TRAINING
Use AI to find a job and prepare to get hiredThe Rundown: In this guide, you’ll use AI to find real job openings, compare them with your experience, tailor your resume, and practice for an interview. One fictional applicant carries through the full process, so you can follow along before using your own details.Step-by-step:
Why most AI investments fail
The Rundown: Agentic AI investment is up 31.8%, but only 11% of organizations turn that into measurable business value. The 2027 CIO Agenda identifies four critical disconnects and what leading organizations do to bridge the gap to value.What you’ll learn:
OpenAI’s Images 2.6 adds GIFs to ChatGPT
Image source: OpenAIThe Rundown: OpenAI just released Images 2.6, an upgraded, faster image model family rolling out across its ChatGPT and Codex platforms with new built-in GIF capabilities that can animate a typed idea, photo, or sketch. The details: 
Trending AI Tools
Everything else in AI todayAtlassian’s webinar on the State of the Developer Experience in 2026 features insights from industry experts on moving beyond AI adoption to operationalization. Register today.*OpenAI released GPT-6 Cyber, an AI built for hunting and exploiting software flaws that beats OpenAI’s main GPT-6 Astra model on most cyber tasks, also revealing the model found 81 serious new bugs in Firefox’s browser in testing. Google rolled out Gemini Omni 1.2 Flash, cutting the price of its video model by 30% to $0.07 per second at 720p and extending clips from 10 seconds to 15.Anthropic introduced a new Personal Finance tool for Claude, allowing its AI assistant to connect to a user’s financial accounts, answer questions, and provide tailored info.Perplexity announced Teammates, new AI agents on its Perplexity Computer platform that each own an ongoing project, with CEO Aravind Srinivas saying he anticipates “every org having hundreds of teammates working for each person.” Arizona Rep. Yassamin Ansari said a pause on advanced AI and a ban on superintelligence should be the “top of the Democratic agenda on day 1” of the new House majority.*Sponsored Listing
Highlights: News, Guides & Events
Source: https://therundownai.beehiiv.com/p/an-i ... ai-speedup
- METR’s inside look at Anthropic’s AI speedup
- Rowan’s Corner: The ‘Three Times’ rule
- Use AI to find a job and prepare to get hired
- OpenAI’s Images 2.6 adds GIFs to ChatGPT
METR’s inside look at Anthropic’s AI speedupImage source: METRThe Rundown: AI-evaluation nonprofit METR just published a new report on its findings from 11 weeks of inside access to Anthropic’s internal data and interviews with staff, concluding the company’s AI progress is likely running about 2x as fast as before, mostly due to its own AI — though Anthropic redacted most of the strongest evidence.The details: - METR’s estimated range of speedup over Anthropic’s 2024–25 pace runs from 1.3 to 3x, with METR crediting AI for around 80% of the company’s faster progress.
- Anthropic co-founder and safety lead Sam McCandlish said the lab’s heavy redactions were to avoid “handing out a roadmap for others to go faster.”
- The report also revealed two internal models ahead of its top-ranked Claude Opus 5.5, including a “Model 4” that cleared 50% on all 15 tested benchmarks.
- Anthropic’s IPO filing defined a danger threshold as “dramatically accelerated progress attributable to AI,” with Monday’s update saying that point may be close.
Your favorite notetaker now does the typing tooThe Rundown: Granola, the AI notepad for people with back-to-back meetings, just launched Voice Typing. Hold Fn and talk, and clean text appears in any app, whether you’re replying to an email or writing a prompt.With Voice Typing, you can:- Speak naturally and get clean text in Gmail, Slack, or Docs
- Dictate in any app, outside Granola
- Use it now on Mac and Windows, iOS and Android coming soon
The ‘Three Times’ ruleImage source: The Rundown @ Google I/ORowan: I have a rule before I’m allowed to automate anything: I have to do it manually at least three times first.Every automation I’ve shared here has followed it. I ran my Friday “factory inspections” by hand for months before handing them to an agent. I sorted my contacts and wrote my own follow-ups before my agents took those over. The manual reps were the training data for my own judgment.When you do a task yourself a few times, you learn what good looks like, which steps actually matter, and where things tend to go wrong (that’s what also lets you check an agent’s work later instead of blindly trusting it).It also fixes the most common mistake I see when people start building with AI, which is going into the tools looking for problems to solve. Flip it. Find the real problems in your life and work first, do them manually three times, then (and only then) automate them. The time you free up will surface the next problem, and then you repeat.That loop is how real AI productivity happens. It’s never one tool automating everything at once; it’s an iterative process that stacks, and six months in you’re a different operator.So before you build the agent, do the task yourself on Monday, again on Wednesday, and again on Friday. Then automate the version you actually understand.AI TRAINING
Use AI to find a job and prepare to get hiredThe Rundown: In this guide, you’ll use AI to find real job openings, compare them with your experience, tailor your resume, and practice for an interview. One fictional applicant carries through the full process, so you can follow along before using your own details.Step-by-step:- Choose one target job title, two related titles, and your limits for location, work setup, salary, and experience.
- Ask ChatGPT to find five current openings. Check each role on the employer’s own careers page before adding it to your list.
- Compare one verified job description with a redacted copy of your resume. Ask AI to show strong matches, partial matches, and missing requirements.
- Tailor the resume and write a short cover letter using only facts you can support. Review every new claim before saving or applying.
- Practice eight interview questions, one at a time, then add the application and next follow-up date to your tracker.
Why most AI investments failThe Rundown: Agentic AI investment is up 31.8%, but only 11% of organizations turn that into measurable business value. The 2027 CIO Agenda identifies four critical disconnects and what leading organizations do to bridge the gap to value.What you’ll learn:- Why ambition outpaces budgets
- How to govern AI economics
- How to align talent with value
OpenAI’s Images 2.6 adds GIFs to ChatGPTImage source: OpenAIThe Rundown: OpenAI just released Images 2.6, an upgraded, faster image model family rolling out across its ChatGPT and Codex platforms with new built-in GIF capabilities that can animate a typed idea, photo, or sketch. The details: - OAI already held the top spot across image benchmarks with its previous models, and 2.6 edges just past its predecessor for No. 1 on Design Arena.
- Two versions are available for developers via API, with Sunburst for higher-quality work and Flare for lighter, faster jobs.
- 2.6 is the first OpenAI image model to add motion, able to convert existing images or ideas into moving GIFs or create entirely new animations from scratch.
- Users can change a GIF’s speed or rework single frames, with the company saying the model also generates images 50% faster than previous versions.
Trending AI Tools
CData Connect AI Gateway - One governed gateway for AI agents, models, & data, adding business context to every request for accurate, cheaper answers*
Images 2.6 - OpenAI’s faster image model that also makes GIFs
Gemini Omni 1.2 Flash - Google’s new cheaper, SOTA video model- Atmee v2 - Atmanity’s real-time avatars built from one image and voice clip
Everything else in AI todayAtlassian’s webinar on the State of the Developer Experience in 2026 features insights from industry experts on moving beyond AI adoption to operationalization. Register today.*OpenAI released GPT-6 Cyber, an AI built for hunting and exploiting software flaws that beats OpenAI’s main GPT-6 Astra model on most cyber tasks, also revealing the model found 81 serious new bugs in Firefox’s browser in testing. Google rolled out Gemini Omni 1.2 Flash, cutting the price of its video model by 30% to $0.07 per second at 720p and extending clips from 10 seconds to 15.Anthropic introduced a new Personal Finance tool for Claude, allowing its AI assistant to connect to a user’s financial accounts, answer questions, and provide tailored info.Perplexity announced Teammates, new AI agents on its Perplexity Computer platform that each own an ongoing project, with CEO Aravind Srinivas saying he anticipates “every org having hundreds of teammates working for each person.” Arizona Rep. Yassamin Ansari said a pause on advanced AI and a ban on superintelligence should be the “top of the Democratic agenda on day 1” of the new House majority.*Sponsored Listing
Highlights: News, Guides & Events- Read our last AI newsletter: AI agents can now make your phone calls
- Read our last Tech newsletter: Apple can’t build enough iPhone Duos
- Read our last Robotics newsletter: Agility’s $2.5B humanoid hits Nasdaq
- Today’s AI tool guide: Use AI to find a job and prepare to get hired
- RSVP to our next workshop on Nov. 12: Use Claude to its full potential
Source: https://therundownai.beehiiv.com/p/an-i ... ai-speedup