A week in Generative AI: GPT-5.5, DeepSeek & Images 2.0
News for the week ending 26th April 2026
Big week this week with the release of GPT-5.5, DeepSeek v4, and ChatGPT Images 2.0. Thereās also a great video of a robot playing table tennis because why not?!
In Web 4.0 news, OpenAI has started testing cost-per-click ads and Anthropic is connecting Claude to lots more tools.
In Ethics news, thereās more commentary on Anthropicās Mythos model, and the Pentagon is looking to pivot to AI-powered warfare.
A few great Long Reads to check out too - Nilay Patel writes about āSoftware Brainā on The Verge which explores why some people are bullish on AI whilst others hate it. Thereās also a great video from Gary Stevenson on whether AI will drive wages up or down.
OpenAI says its new GPT-5.5 model is more efficient and better at coding
The big release of the week was from OpenAI who made GPT-5.5 available to the world, hot off the back of Anthropic releasing Opus 4.7 last week.
The promotional video definitely positions GPT-5.5 around tool use and automation and shows off how the Codex app is a much more capable interface than the usual way of interacting with ChatGPT. I suspect weāre seeing a move towards the ChatGPT interface just being for āchattingā whereas the Codex app is starting to pivot away from just being a ācodingā app to more of a general purpose āworkā app for getting things done.
As Ethan Mollick says in his review, itās difficult to articulate how GPT-5.5 improves on previous models as AI still has a jagged frontier. But it does intuitively feel better, faster, and more efficient. Itās definitely progress, but itās hard to articulate on what exactly.
OpenAI | The Verge | Ethan Mollick | Simon Willison
DeepSeek previews new AI model that ācloses the gapā with frontier models
Itās been just over 15 months since DeepSeek released their v3.2 and R1 models which temporarily sent tech stocks in the US spiralling. This week the released both DeepSeek V4 Flash and V4 Pro which are mixture-of-experts models with context windows of 1 million tokens each.
DeepSeek says they have almost āclosed the gapā with current leading models, but both their models are text only. However, the models are much more affordable than any other frontier model available today and are available as open-weights models unlike any of their peers.
HuggingFace | TechCrunch | Simon Willison
OpenAIās updated image generator can now pull information from the web
On Tuesday, OpenAI started rolling out its second generation image generating model, citing ānew thinking capabilitiesā, higher resolution images and more consistency. It wouldnāt surprise me if weāre about to see (hear!) a new voice generating model too, as Iām sure thatās what they used for the narration of this video.
ChatGPT Images 2.0 certainly looks very impressive, but we havenāt seen the same Ghibli-style hype we saw with the first version a year ago. To put it to the test I gave it the last 6 weeks of newsletters Iāve written and a couple of the more recent longer reads Iāve published and it came up with the hero image for this weekās newsletter. You can judge for yourselves!
Simon Willison had a fun test of the new model too, getting it to generate Whereās Wally/Waldo style images, which I think he quickly regretted!
OpenAI | The Verge | Simon Willison
AI-powered robot beats elite table tennis players
Sony AI in Zurich have developed a table tennis playing robot, called Ace, that was able to win three out of five matches against elite players, but lost the two it played against professional players. This might be the Garry Kasparov moment for robotics.
Back in 1996, IBMās Deep Blue, a specialised chess playing supercomputer, became the first machine to win a chess game against a reigning world champion. Similarly, Ace is a specialised table tennis playing robot - it uses multiple camera, specific ball-tracking and, isnāt bi-pedal, using an eight-jointed arm on a movable base.
It even pulled off one rapid backspin shot that a professional had thought impossible:
When Ace played an unusual shot, intercepting the ball early and imparting backspin, the former Olympic table tennis player Kinjiro Nakamura, said it had not thought it possible, but now believed that humans could learn the shot.
So, weāre still a waySteam way off a general humanoid robot being able to reliably beat table tennis professionals. It took about 20 years to get from Deep Blue to Alpha Zero, which was more of a general purpose game-learning system. I wonder how long it will take us to get to a general purpose game-playing humanoid robot?
Web 4.0
Claude is connecting directly to your personal apps like Spotify, Uber Eats, and TurboTax
OpenAI now lets teams make custom bots that can do work on their own
AI Ethics News
Some Unknown Group Is Reportedly Using Claude Mythos Without Permission
AI Companies Think Destroying the Planet Is an Acceptable Trade-Off for Unlimited Profits
Now Meta will track what employees do on their computers to train its AI agents
Officials hugely underestimated impact of AI datacentres on UK carbon emissions
AI failure could trigger the next financial crisis, warns Elizabeth Warren
Microsoft and Meta announce large staff reductions as they spend big on AI
Long Reads
The Verge - Beware Software Brain
Garys Economics - Will AI drive wages downā¦ or up?
The Ezra Klein Show - Stewart Brand on LSD, AI Black Boxes and the Beauty of Care
āThe future is already here, itās just not evenly distributed.ā
William Gibson