William Craig retweeted
OpenAI just updated its dataset on how people use ChatGPT.
I went through every file and found a few surprises:
1. Classic Google query types are now about half of all messages
- 48% of all messages were Google-style questions in September 2026, up from 43% in July 2024.
- The 48% breaks down into 20% simple searches (facts, products, recipes) and 28% Practical Guidance (how-to, tutoring, health, creative ideas).
- OpenAI sorts every message into work or personal. Google-style questions make up 53% of personal messages and 38% of work messages (September 2026).
- Personal messages grew from 48% to 68% of all messages since July 2024.
- Most of the rise comes from personal messages taking a bigger share, because they're search-heavy.
2. Commercial research at work grew 75%
Among US work messages (Q1 2025 vs Q3 2026):
- "Obtain information about goods or services": 1.7% → 2.9%
- "Determine values or prices of goods or services": 0.3% → 0.7%
- "Gather information from physical or electronic sources": 1.2% → 1.8%
Surprised me: researching a subject for work fell by up to two thirds (technology -66%, history and social issues -65%, health -42%). Product and price research grew.
3. Countries that use ChatGPT more use it more for search
- The quarter of countries with the highest ChatGPT use per capita put 24% of messages into simple searches in Q3 2026, vs 16% for the lowest quarter.
- Across ~150 countries, the correlation grew from 0.30 to 0.74 (Q1 2025 vs. Q3 2026).
4. Personal use grows faster than work
- OpenAI's classifier tags every sampled message as work-related or not (consumer plans only).
- Work went from 52% of messages in July 2024 to 32% in September 2026. The data only shows shares, so work messages could still be growing in absolute numbers.
In September 2026, the share of messages that were work-related was:
- Free: 25%
- Go: 42%
- Plus: 51%
- Pro: 68%
Source: OpenAI Signals v3.0, announced in "ChatGPT's expanding world of work"
159 countries, 27 million sampled messages, July 2024 to September 2026
William Craig retweeted
INTERESTING!
Bing Webmaster Tools is once again leading the charge: you can now see topics you're being cited for, how often they're cited and related grounding queries.
This is the closest we have to classic search data.
Your turn, Google!
William Craig retweeted
Gemini 4 Argon is here.
It’s available globally today for developers and enterprises in the paid tier of the Gemini API and the Gemini platform, for Ultra subscribers in Google Antigravity, and Ultra and Pro subscribers in the Gemini app and Search.
Hope you all enjoy it!
William Craig retweeted
RIP Git
New on our Engineering blog: Creating in-house source control.
In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale.
Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: anthropic.com/engineering/re…
William Craig retweeted
Salesforce in Claude is now available in beta.
It brings your accounts, opportunities, and pipeline into Claude, with 37 pre-built sales skills. Prep a call, review a deal, create a pipeline dashboard, or send your forecast without leaving the conversation.
William Craig retweeted
GPT-Live-1 is now available in the API.
Bring ChatGPT’s natural back-and-forth to your app, with voice agents that listen while they speak and work with the models and harness you choose.
William Craig retweeted
Developers wanted to resume Claude Code sessions so badly they built their own tooling for it. Shell hooks that save the session ID on exit. A published Ruby gem for bookmarking sessions by name. Custom zsh functions passed around like recipes.
All of that engineering existed to preserve a single string.
The sessions were never lost, which is what makes this a great product lesson. Claude Code stores every conversation as a JSONL file on disk, and the CLI could always resume them. The desktop app just couldn't see them. A GitHub issue back in April called it mostly a surfacing problem.
So for months, users filled the gap themselves. A SessionEnd hook here, a bookmarking tool there, each one a tiny patch over the same hole.
That's the strongest demand signal in software. A feature request costs the user 30 seconds. A workaround costs them hours, and they maintain it. When people start publishing packages to route around a missing feature, they've stopped asking and started paying with their own time.
The best product teams read this signal deliberately. Upvotes tell you what sounds nice. Workaround tooling tells you what people already can't live without.
And the reason this feature earned that level of desperation is worth sitting with. An AI coding session carries hours of accumulated context. Decisions made, constraints explained, dead ends ruled out. Losing it means paying that cost again. The session is quietly becoming the durable unit of work, and it just learned to travel.
William Craig retweeted
We have a limited window to strengthen cyber defenses, and together with organizations including @AnthropicAI, @awscloud, @Google, @Microsoft, and @Oracle, we're calling for a global effort to give defenders the tools, resources, and support to protect the infrastructure we all depend on.
If we act decisively, we can turn today's AI advances into lasting improvements in security and make our digital world safer for everyone.
openai.com/collective-cyberd…
William Craig retweeted
ChatGPT Work can now use its computer and browser to sign in to websites on web and mobile, without ChatGPT ever seeing your username or password.
That means you can ask it to:
• Set up utilities for a new apartment
• Book a DMV or passport appointment
• Check reimbursement costs through your insurance
• Find and book an in-network doctor around your availability
• Check when your car registration expires and prepare the renewal paperwork
• Compare your rental insurance policy with an issue you’re emailing your landlord about
• Find and save apartment listings that match your criteria
• Restock something just by uploading a photo
• Schedule a package pickup for a return
• Cancel tickets for a rescheduled trip
• Book a vet appointment
• Submit reimbursement paperwork for medical treatments
• Check resale sites for new drops and save things you might like
• Find candidates with specific experience and draft outreach
• Take invoices from your email and submit them to your accounting software
• Draft replies to rental property inquiries
• Fill out permit applications for your small business
• Add action items to a vendor portal based on a recent client call
• Analyze the latest ad campaign for your small business
William Craig retweeted
More data than open-source AI is taking share from OpenAI and Anthropic. Open source has gone from 28% token share to 62% token share @vercel over the last 2 months. Chart from @rauchg
Super impressive given that the sum of OpenAI and Anthropic accelerated in July. So net token/AI infra demand accelerated even more than the acceleration we saw at the frontier. And suspect Grok growing even faster than open-source and we saw some of this in the @tryramp data.
Open-source AI taking share is positive for AI infrastructure demand as it lowers margins at the model layer and an open-source token costs just as much compute to produce as a frontier token. Nothing about open-source AI inference is “free.”
Most likely end state IMO is that closed, frontier tokens are 60-90% of economic value but only 15 to 25% of tokens.
William Craig retweeted
Claude Code can design now. The new /design skill (research preview) brings Claude Design's artboard workflow into the CLI and Desktop, built on artifacts.
Run /design to get editable artboards for your UI — pick one, tweak it, then have Claude implement it.
William Craig retweeted
GPT-5.6 Sol just hit 750 tokens per second in Ultrafast mode — 14x faster than the standard model. At that rate, it writes a full 90,000-word novel in under 3 minutes. Every 3 minutes. Forever. The economics of knowledge work just changed permanently. We're still in the early innings.
William Craig retweeted
K3 at half the price and comparable performance is crazy
Today we're sharing the Notion Knowledge Board.
We get to observe a wide range of real-world AI tasks at @NotionHQ: meeting follow-ups, support tickets, sales pipeline updates, market research, all the good stuff. Knowledge board is a live, ongoing eval of how models handle real knowledge work measured on live anonymized traffic and scored by an ensemble judge we built in collaboration with researchers at Anthropic and OpenAI.
The takeaway is the tradeoffs: cost per task variance is high, and some of the fastest models are also very capable. The right model genuinely depends on your work.
We will update as new models ship: labs.notion.com/knowledge-bo…
