Simon Smith

I think this will all largely be solved by AI agents becoming more like employees than apps. We see this happening now with Claude Tag. Claude has become a colleague who works in Slack with you. People delegate tasks to Claude like they would to any other colleague. People see others do it and learn by example. It's not perfect, but it feels like a key part of the AI future.

Not all enterprises are billed the same by OpenAI. Some have more baked into their subscription that have allowed them to push off managing usage-based billing. This changes on renewals, but patterns have been established and it makes the transition feel more painful. I’m also seeing more people hit the limits when given access to Codex (will be the same for Work). So much so that three of my clients won’t open up Codex to the full team. This may be due to more conservative credit allocations, which is a mix of the first issue and not understanding the ROI on agentic work over one-off chats. But the costs can be quite absurd. The challenge is that we’re turning everyone into managers. People have to know how to manage the tools somewhat like we manage people. What to delegate, to which models, in which type of workflow, and how to measure performance across various setups. When I read that most people don’t hit their limits I wonder if it’s because they’re too intimidated to really dive in.

something I think no one talked about? the @ChatGPTapp router is back and better than ever in chat mode. instant for my simple questions. extended thinking when I ask for lookups and research seems to perfectly handle my casual chats with no model or reasoning setting

@jdjohnson @albfresco @ChatGPTapp It shouldn’t be a big transition because under credit-based enterprise billing you already consume credits for things like Pro models. We’re used to usage-based billing. And most users don’t even hit their monthly credit allocation now.

@_simonsmith @albfresco @ChatGPTapp I’ve struggled to answer this question for our entperise clients. Like you say, not an issue for most, but a problem for power users who blow through Codex usage. As they also move from ChatGPT to Work, I anticipate token problems. Interested to read how this unfolds for you.

@deredleritt3r Labor force participation is down, though. Fewer people in the labor force, not counted in unemployment rate.

In late November 2025, AI agentic harnesses became highly useful, and enterprise adoption exploded. For example, based on BTOS data, the percentage of large (250+ employees) U.S. enterprises using AI grew from ~25% to 37% between November 2025 and May 2026. And the unemployment rate... hasn't moved. For workers aged 20+ it was: September 2024 (first reasoning model): 3.6% November 2025: 4.1% July 2026: 3.8% We see the same thing for workers aged 20-24: September 2024: 7.0% November 2025: 8.3% July 2026: 7.1% This is still very early data, but perhaps some aggressive assumptions regarding AI's impact on the labor market should be tempered a bit?

@jdjohnson @albfresco @ChatGPTapp Yes, I believe Work uses credits the way Codex did before. As for determining, we don’t at this point. We encourage experimentation and monitor usage and intervene only if needed. Most users aren’t big token burners. Only a small percentage of power users, typically.

@_simonsmith @albfresco @ChatGPTapp Does Work use Codex limits now? How do you determine when people should use Chat vs Work in enterprise?

For what it's worth I've been using GPT-5.6 Sol Medium to do lots of data analysis today and am nowhere near my 5h or weekly limits in my Plus plan (which is separate from my Enterprise plan, which is effectively unlimited as it pulls from a credit pool).

There's an odd gap in ChatGPT's Library on iOS, @JustinBleuel. There's no way to create a note, as far as I can tell. There's no button to create a new note in the library, nor can Chat do it upon request, nor can I turn a writing block into a new note. Am I missing something?

My X feed follows a power law. The first 20% of "for you" posts contain 80% of the signal. And after about 30% you get into rage bait and disinformation and end up worse off for having read anything at all.

I don't think we got two new "frontier" model labs this week in SpaceXAI and Meta. We got cost-efficient tier-two models from established labs that had fallen behind. Only two labs, OpenAI and Anthropic, have tier-one models, and if we define "frontier" as top-tier capabilities then nothing changed in that area. It may now look more possible for SpaceXAI (that is a really awkward name) and Meta to challenge at the frontier, but right now they're more on par with tier-two models like Opus 4.8 and GPT-5.5 in terms of performance. Nothing bad about that, competition is good, but it's also a stretch to call this "frontier" or else that term loses all meaning.

Has anyone benchmarked models on the accuracy of their estimates for how long a task will take them? They estimate like a human, like “this task should take 1-2 weeks,” but execute like machines, often in minutes. Can’t we just add a reward in the RL pipeline for accurately predicting how long a task will take them? Or, at the very least, how many tokens, which we can then convert to time depending on the tokens/second for the model in whatever hardware it’s running on? Like, imagine you ask GPT-6 to estimate how long a task will take it, and it’s like: “This should be about 47 million output tokens. Given the information I have that on this hardware I’m outputting 1,000 tokens/second, I estimate 13 hours.”

Absolutely. If it would just stay pinned after I scroll up, that would immediately solve this use case for me. Right now, the issue is that while I’m translating, if I ask GPT a question, the transcript for its response pulls me down to the bottom of the thread and I lose my place.

@_simonsmith great use case! for now, would it be enough if we fixed the bug where it auto-scrolls when you’ve scrolled up yourself? I do like the idea of the model showing you a “text widget”

So, @athyuttamre, I have a very specific feature request for GPT-Live and I’m not sure I have the right solution but do think I’ve identified an opportunity. I’ve been using GPT-Live to learn Spanish, and it has been the most exciting and effective language learning experience I’ve had in years of practice. I’m combining listening, reading, translating, and speaking into single practice sessions. I get GPT-Live to speak a few paragraphs about a topic of interest to me. For example, I’m planning to go to Tamarindo in Costa Rica in December and asked it tonight to tell me a bit about it, in Costa Rican Spanish. I turn on transcription while it’s talking so I can see and read the words it’s saying to learn how the words are pronounced. Then, when it’s done, I go through everything line-by-line and try to translate it back to GPT-Live, asking questions when I get stuck, just as I would with a real Spanish tutor. At the end of the session, GPT-Live can then also test my comprehension and pronunciation through a Q&A session. This is such a powerful and enjoyable learning experience, and feels incredibly natural. I also find it more practical than language lessons that have nothing to do with my interests and don’t reflect local idiomatic usage for places I plan to visit. The one annoying thing right now, though, is that every time transcription starts, it moves the thread down to the latest transcribed text. Normally this is great, but it’s not great when I’m trying to read through an earlier response! My first thought was: can you let me pin a transcription output so it stays on screen while I’m talking to GPT-Live? But maybe another option is to give GPT-Live the ability to create a note or other UI element that can stay on screen while we’re talking about it, and have this note be populated from its transcript so I can see the words as it’s saying them and learn the pronunciation. I know it’s a very specific ask, but this could be one of the best tools for language learning and so easy for anyone to use with just a few additional features.

Learning things with AI using voice is very effective and feels quite natural. Ask questions, get explanations, debate topics. Perhaps we’ll transition back to a more oral learning culture, but with the twist that AIs mediate much of the learning.

GPT-Live discovery: If you drag the circle to the bottom of the screen or turn your phone sideways, you get live transcription. This is fantastic for language learning because you can hear the words and see them simultaneously.

After about two weeks, Claude Tag has ~16% our usage of Claude web and ~11% our usage of Claude Cowork. It's now in 99 channels, with messages to Claude up ~303% week over week. It still feels like an early adopter tool, but one that spreads fast through ease of use and example.

Maybe we all just take a breath and be grateful for the bounty of models and harnesses and apps and plugins and people working on them for a few minutes. Yeah, rough edges, but this new world isn't even four years old and we're all trying to figure it out.

Oh, interesting. GPT-5.6 is particularly good at presentations based on this benchmark. I'm going to run my typical model presentation benchmark now...

Confirmed by Arena voters: With GPT-5.6, OpenAI has overcome its frontend woes.

It feels like we're really on the cusp of rapidly accelerating STEM with AI. Like, just a rapid fire series of advances. This took less than an hour thanks to a model advance and parallel compute. Models will get even smarter and parallelization even broader. Plus, 750 tokens/second and up. What will it feel like when we're just getting a continuous stream of significant STEM advances, seeing the next before we can even absorb the current?

Another one here, for training on chords. The sound actually works in the browser too.

So, ChatGPT Work doesn't just build inline interactive explainers. You can build full inline applications. For example, I built an inline Spanish tense training app. This is great. Lots of ideas for things I can build inline.