ChatGPT Dots & Space: Tested

Hello,

OpenAI gave us a lot to get our heads around this week (an unprecedented 22 announcements at Tuesday's Dev Day 2026 event in San Francisco!). I had early access to its new ‘dots’ and Space features, so there’s a longer hands-on review below.

It's funny how all these AI assistant products are settling on a similar visual language of cute blobs - essentially tamagotchis. Is this how AI tries to make itself less scary?

In this issue

Claude Sonnet 5.5: why care?
Unless you’re really nerdy about AI, another model probably doesn’t excite you. But now both Anthropic and OpenAI have their latest, most advanced models out in leaner, more efficient incarnations: Sonnet and Opus 5.5 in Claude, and GPT 6.1 Sol in ChatGPT and Codex.

Aside from the boring benchmark stuff, new use cases are emerging. These models are particularly good now at graphic design, PowerPoint presentations, 3D modelling, video editing, and motion graphics.

With the right guidance, these models are producing genuinely decent work. Worth trying again if you had crappy results with previous models. Anthropic.

Google’s Gemini now has skills, plus a major new model is coming
Gemini 4 Argon just got announced, without public access. I’m working with Google to try to get early access; I’ll keep you posted. More immediately useful though is skills, which are rolling out to replace Gems (much like ChatGPT now has skills and is replacing custom GPTs with them).

So now, Gemini, Claude and ChatGPT all support reusable instructions, reducing switching costs (you can download and import skills between them). Skills · Argon.

Dots & Space: my hands-on review
I’ve been using dots and Space for a week. They were the biggest mainstream-relevant announcements from OpenAI’s DevDay on Tuesday.

With dots, I can see this becoming the main way people interact with AI by next year. But I also think OpenAI rushed it out. With Grokbot and Meta Muse getting decent reviews these last few weeks, there’s a scramble to win in this new phase of AI.

What is a dot?

A dot is an always-on assistant with one ongoing conversation, its own cloud computer and access to the apps you connect. You can also let it use your computer.

“But I already chat to ChatGPT”, I hear you say!

The real difference here is that it is a persistent and proactive assistant 24/7: it can keep responsibility for a big job between chats. But it will also message you throughout the day to remind you of things you need to do based on its access to your email, calendar, meeting notes, systems, etc. I personally find those nudges very useful.

This week for instance I’ve asked mine to find creator economy podcasts and get me on them, organising outreach and keeping me updated on progress.

And it can take on that job, spin up new chats and tasks to get it done, but also keep an eye on my inbox and keep me updated over the coming weeks on progress. How dots work.

It improved during my early-access week, but there are rough edges. Calling my dot leaves the phone ringing for almost ten seconds (while ordinary voice mode feels instant 🤷‍♂️). It sometimes just gives up or ignores me. Again, I think this is just rushed out to not miss this current wave of all-in-one AI assistants. And there are some good features coming, so I think it'll be a pretty smooth experience by the end of the year.

Space

Space is the update that makes the most sense. Think Google Docs meets Notion, inside ChatGPT: one single place for all the documents, etc., that you've created that otherwise gets buried across hundreds of conversations and are a nightmare to find.

My dot created a page for the podcast outreach project, giving me somewhere to check progress. You can share pages and work on them with other people and their dots. It replaces the Library. It also has Google Docs integration, so you can work on those there too.

Collaborative Slides and Sheets, mobile editing and a few other features aren’t available at launch, but again, I feel like this will become more polished over the coming weeks. Space details.

Other OpenAI Updates

GPT-6.1 Sol has become my default model during early access: close enough to Astra for most of my work, while feeling quicker and using less allowance. It’s available in Work, Codex and the API, rather than ordinary Chat. There’s also a Mac Meetings plugin for notes and follow-ups, expanded Canva and Shopify plugins, and a way to authorise participating apps to use your ChatGPT subscription allowance. Sol · DevDay recap.

The pricing changes are not popular. Pro now comes at $100, $200 and $500 a month. The reopened $200 plan has a lower allowance, with eligible existing subscribers keeping their previous allowance through 29 October. The $500 tier includes Ultrafast, which I’ve found impressively quick, but Astra consumes subscription allowance at eight times the standard rate. Conversation with your dot doesn’t count towards normal ChatGPT limits; Work and Codex tasks it starts do. Plans · Speed.

Access is gradual. Personal Pro dots exclude the UK, EEA and Switzerland at launch; Business Premium covers supported regions, and Enterprise has an admin-enabled beta. Space is rolling out to Pro, Business and Enterprise. Also worth preparing for: custom GPTs are scheduled to retire on 11 December, subject to account exceptions. Their instructions migrate into skills inside plugins; custom actions need separate rebuilding. Check your own account’s notice before switching. Dots access · GPT migration.

AI insights led marketers to bad decisions (shock! 😳)
In GWI’s survey of 150 UK marketers, 55% said their business had made a significant decision using AI insight that proved wrong or misleading. Small survey, big eye-opener.

I see this in others’ work and my own: we delegate tedious analysis, then have little appetite to check it thoroughly afterwards. Once the client deck looks great, who wants to inspect every number, assumption and resulting recommendation? That’s the job we were getting out of. This is a major workplace teething problem for AI, that can probably only be fixed by time and better training. Read more.

Shopify agents can now reach the checkout
Shopify has added tools that let browser agents read and update a checkout, then submit it after the buyer confirms. Authentication and other steps needing a human hand control back to you. Merchants don’t need to configure anything new. The significance is straightforward: an assistant can increasingly help with the whole purchase, from finding a product to placing the order. For retailers, that makes clear product information, availability and a usable checkout part of being discoverable through AI. Read more.

Some early numbers on ChatGPT advertising. Mixed!
Graphite’s analysis of Similarweb panel data puts ChatGPT ad click-through at 0.74% in the UK and 1.91% in Brazil, with considerable movement among leading advertisers. Useful early evidence, but don’t turn those figures into your campaign targets. The markets cover different time periods, and these are panel estimates per ad-bearing exchange, not audited advertiser results. It’s a channel to test with a clear conversion measure. A click alone won’t tell you whether the conversation produced a customer. Read more.

Instagram’s latest Reels advice, straight from the platform
It’s always useful to hear best practice directly from the platform. Instagram’s updated guide covers hooks, text overlays, captions and editing controls. A useful distinction: overlays emphasise chosen moments; captions make the whole spoken message accessible. Give people a reason to keep watching immediately, use original footage and choose a length that suits the idea. Video remains, for me, the most efficient way to reach audiences at scale organically. This is a handy refresher, rather than evidence of another mysterious algorithm change. Grab the guide here.

Edits gets an AI assistant
Meta is adding an assistant to Edits that uses your content, performance and audience information to help explain what’s working and suggest what to make next. That could be more useful than simply generating another video: turning results into a better creative decision. Access and prompt limits vary as it rolls out. And after the survey above, you know the question to ask: which actual evidence supports its recommendation? Read more.

Reddit is switching off RSS
Reddit plans to end RSS feeds on 13 November as it tightens access to its data. If you follow subreddits through a feed reader, or use feeds for research, monitoring or an automated workflow, check what will break. It's also a blow to AI bots, scrapers and automations. Read more.

LinkedIn wants to help brands find the right creators
LinkedIn is testing creator-discovery filters covering subjects such as industry, job function, content type and audience demographics. Wider availability is expected in early 2027. The useful part for B2B marketing is finding someone whose expertise and audience fit the job, rather than starting with follower count. A respected specialist in the right niche may be a much better partner than somebody with a much bigger, less relevant audience. Read more.

Instagram collabs are worth a closer look
An Emplifi study published earlier in September found that accepted Instagram collaborations outperformed ordinary posts across brand sizes in its first-half data. Creator-led collaborations were particularly strong; the smallest brand profiles saw more than five times the interactions. This is observational data, not a promised uplift. Still, it’s a useful reason to test a genuinely relevant creator partnership and a shared post, rather than simply paying someone to mention you. Reels showed particularly strong gains. Read more.

Jean Phil and the very believable AI oddballs
Seen any of these guys below in your feeds?

Why is this happening now? AI video generation has gotten so good that it's really hard to spot. When you move away from perfect model-like people and cast quirky, eccentric ones, it looks even more real.

Jean Phil, the fictional French shadowboxer, has racked up millions of views and hundreds of thousands of followers in just a few weeks.

It won't surprise you that the links in the bios of these characters lead to a hastily spun-up cryptocurrency that you can invest in! Read more.

Google is testing payments for AI answers
Amidst fury from online publishers that their traffic is being stolen from Google, with content instead being served in AI answers, Google is piloting a programme with around 100 publishers paying them when using their content in AI answers, in search, and in Gemini. They've not confirmed if this is going to become a broad rollout yet, but in many ways, it follows the model of social media, where creators do get a share of ad revenue. I think that would be a good thing. Read more.

Thomson Reuters wins another round on AI copyright
A US appeals court has upheld Thomson Reuters’ copyright win against Ross Intelligence, which used Westlaw’s legal summaries to build a competing research product. The decision matters, but its scope matters too: Ross’s system retrieved existing text and wasn’t a generative language model. This doesn’t settle every dispute over training ChatGPT or Claude. It does reinforce the risk of copying a publisher’s valuable material to build a direct substitute for its product. Read more.

The VMAs was a hit
The VMAs drew 8.43 million US viewers across CBS and MTV, their biggest audience since 2015, according to Paramount’s preliminary figures. TV is so back!

What struck me though was the attention around Gener8ion and Yung Lean’s “Storm” performance, recreating Damien Jalet’s choreography from the video. With all the talk of AI ruining video and media, we still love watching humans move in interesting ways. That fascination has been with us for thousands of years. Figures · Watch.

An 85-year-old Enigma message finally gives up its secret
A German Enigma message from 1941, published as a challenge in 2005, has been cracked with help from GPT-6 Astra. The researchers say it wrote a simulator and used a clue from another message to recover the key. Crypto Cellar’s cryptanalyst checked the result. The solution was submitted on 15 September and received fresh coverage this week. It’s a fascinating meeting of historical detective work and modern computing, with a concrete result that somebody else could check. The account.

A solar system to get lost in
This browser-based explorer lets you move around the solar system, follow spacecraft and satellites, and see a catalogue of roughly 526,000 asteroids. The creator uses JPL and CelesTrak data and says it refreshes daily. These are calculated positions, with enlarged markers so you can actually see things; older satellite positions are estimates. It’s a wonderfully nerdy way to turn an abstract sense of space into something you can explore. Open it when you have a few minutes to spare. Explore · Creator’s notes.

What Apple has lined up for October
MacRumors has collected twelve things to expect from Apple this month. The confirmed list includes the foldable iPhone Duo on 23 October and a Mac Studio configuration with 512GB of memory later in the month. Rumoured additions include an OLED iPad mini, refreshed Macs, Apple TV and HomePod mini, plus the long-awaited home hub. Those latter products remain rumours for now. Worth a look if you’re weighing up a purchase. The preview.

Sony is skipping CES
Sony has confirmed it won’t attend CES in 2027, pointing to its growing focus on entertainment, intellectual property and technology for creators. For a company so closely associated with televisions, audio equipment and gadgets, that’s quite a marker. Games, films, music and the tools used to make them increasingly define the business. That shift changes what Sony has to show at a consumer-electronics fair, and who it most wants to reach. Engadget · Confirmation.