A week in Generative AI: GPT-5, Claude & Helix

News for the week ending 17th August 2025

Aug 17, 2025

There’s been a bit of fallout from the launch of GPT-5 this week with users complaining about the sudden deprecation of GPT-4o, the real-time router not working well, and GPT-5’s formal tone of voice. All things OpenAI have quickly fixed, but it’s always interesting to see what users are vocal about following a big new model launch as it tells you what they really care about. Users had become very attached to GPT-4o!

The Blueprint’s Generative AI Explorer is now updated with all the recent new model launches and feature releases. You can compare models to each other and see all the benchmark scores that are publicly available to find the best model for your use case. It’s also a great way of tracking progress of each of the major AI companies and exploring the products and features that they offer.

There were also some quality of life updates to Claude this week and an interesting update from Figure AI showing the progress they’ve made with the dexterity of their humanoid robotics platform.

In an interview with The Verge Sam Altman gave a great sound bite about how they’ll soon be the third largest digital platform in the world. In other Web 4.0 news there’s a long interview with the Head of ChatGPT, Nick Turley, who’s not convinced about taking advertising, but doesn’t completely rule it out.

In Ethics News it was reported that Meta’s chatbots were allowed to have romantic conversations with children and Sam Altman believes AI is in a bubble right now.

GPT-5 launch fallout

There has been lots of reactions to the launch of GPT-5 last week with complaints from users about the sudden deprecation of GPT-4o, issues with the model picker being taken away, and users not happy with GPT-5’s formal tone of voice. With every big new model release from OpenAI I suspect we will always see bumps going forwards - users have become very attached to the model they use in ChatGPT!

There’s a good interview with Sam Altman from The Verge about the challenges with GPT-5’s launch that gives a summary of all the user feedback, and OpenAI’s response. I think they realise that suddenly deprecating GPT-4o was a bad, and strange thing to do and so removing older models without any notice won’t be happening again.

ChatGPT’s model picker is back, which I think is more of an admission that the real-time router that they launched with GPT-5 just isn’t good enough. The idea is right, and I’m sure it will get better over time (especially as according to the GPT-5 system card it’s ‘continuously trained’). But for now it doesn’t work well enough. In my own testing its difficult to tell how much of a problem it is (as you can’t tell where your question is being routed to), but I’ve definitely been getting responses that aren’t factually correct, which is likely down to being routed to the ‘minimal’ GPT-5 model. The ‘medium’ model wouldn’t make those kinds of basic mistakes, especially as OpenAI claim to have reduced hallucinations with GPT-5 by 80%.

Sam Altman also updated on Friday that GPT-5 will soon be ‘warmer’ and that they’re working on allowing users to personalise their experience more, something I wrote about previously in my Beyond Chatbots series. One thing I hadn’t considered in my piece was the challenge of people becoming overly attached to their personalised AI model, which its great to hear is something OpenAI are considering.

There are lots of learnings with every major model launch, and I think OpenAI, with the exception of the sudden deprecation of GPT-4o, have made all the right decisions. GPT-5 is still a great model and a big step forwards, its just having a few teething problems!

Pick up where you left off with Claude

Claude has had some nice quality of life improvements made to it over the last week. It can now remember your past conversations and it can also now handle longer prompts.

Remembering past conversations isn’t like ChatGPT’s memory features - it will only reference them when you ask it to and it’s not a persistent profile that you can manage. The content window is now a whopping 1m tokens, up from 200k tokens, meaning Claude can now take in an entire codebase, or the entire Lord of The Rings trilogy. That’s a ridiculous amount of context!

Source

Latest update from Figure AI

Figure AI have made some more progress with the dexterity of their Figure 02 humanoid robot as you can see in the video above. It’s still not great at sorting laundry by human standards - it does a slow, but decent job. Remember that this is the worse it will ever be though, so I expect these robotics platforms will be very capable compared to humans when it comes to dexterous jobs in a couple of years.

Source