Parameter Update: 2026-26

Loads of new model launches this week. It feels like every lab is trying their hardest to catch up to Mythos at a fraction of the price - exciting!

OpenAI

GPT-Live

A few weeks ago, Thinking Machines showed off their "Interaction Model" architecture where the model could combine input and output streams in realtime. This week, OpenAI announced something similar: GPT-Live, the successor to the Advanced Voice Mode in ChatGPT, follows the same "full duplex" idea, where incoming inputs no longer necessarily interrupt the models output. Remember how amazing the AVM demo felt, where you could interrupt the models as it was speaking? You can still do that now, but the model is now also capable of interrupting you - or can choose to keep talking. It is also much better at keeping quiet if you stop talking, so the days of "let me know if there's anything else you wish to talk about" are finally over. Add the ability to perform tool calls and reasoning in the background and it makes for a really nice experience.

I can't say that I'm the biggest fan of talking to my computer, but the ceiling for high-quality voice demos just increased massively, and this is one of the coolest products they shipped in a while. Try it out!

GPT-5.6 launch + ChatGPT Work

On Thursday, OpenAI launched GPT-5.6 globally. The model is good, very close to Fable in some tests, but more importantly, it's comparatively cheap and will continue to be included in Codex usage limits.

They combined this with the launch of "ChatGPT Work", which is their "superapp" that has been rumored for a while now. On web, it integrates some new generative UI artifacts and makes the agent loop more focussed on doing things rather than answering questions. In effect, it seems they're trying to bring the Codex agent loop to all users. On Desktop, it replaces the Codex app (or rather, it is the Codex app - you can even switch back to the old branding if you want!) with some enhancements to make it more useable for non-programming use cases.

Apple Lawsuit

Apple has filed a lawsuit against OpenAI, alleging coordinated effort to steal trade secrets and poach employees. At the center of the lawsuit is engineer Chang Liu, who spent 24 years at Apple before joining Jony Ive when he left Apple to found io Products (which was then acquired by OpenAI for $6.5 Billion). The core claims include:

• Liu kept an Apple-owned laptop and used it to keep access to internal files and data after joining OpenAI.
• OpenAI systematically poached over 400 engineers and executives from Apple. When inviting Apple engineers to interview, they told them to bring prototypes from the office.
• OpenAI approached one of Apple's suppliers and and had them show off a proprietary metal-finishing technique used to produce the distinctive look and feel of Apple products. When asked, OpenAI told the partner that Apple had given permission (when it obviously hadn't).

The allegations are not pretty, and it no matter how this plays out, it looks like this might slow down or kill OpenAI's hardware aspirations.

Grok 4.5

I was wondering how long it would be until we'd see the proper effects of the Cursor team joining xAI (which rebranded as SpaceXAI recently, lol). This week that point finally came in the form of Grok 4.5, which is a massive upgrade in all the areas you would expect it to be (i.e., all programming/agentic domains).

Overall, it roughly matches Opus 4.8 in most benchmarks, which would have been an extremely impressive feat two weeks ago, but already feels like playing catch-up now.

As you would expect, new Grok model also means new Elon drama.

Meta

Muse Spark 1.1

In the second model launch of the week that would've been much more impressive two weeks ago, Meta announced Muse Spark 1.1, which also reaches rough parity with Opus 4.8.

The big news here:

• The model is available through their new model API, and I expect availability to broaden in the future.
• Pricing is better than GLM-5.2 (!).
• The fact that this launch comes only three months after the initial Muse Spark launch signals to me that Meta's scaling recipies are working, so expect them to potentially move quickly going forward.

As far as PR victories go, on the other hand, it appears Meta still has some way to go.

As a reminder, with all these new "Opus 4.8 level" models, here's a pricing overview. Keep in mind that, in practice, pricing will also be influenced by token efficiency, tokenizer, caching, and other factors, though these are probably pretty good as a proxy.

Muse Image

Meta's first image model from the "new" MSL org, Muse Image, is both technically impressive and launched with systematic backlash. Technically, it does many of the things other SOTA image models do - i.e., autoregressive generation including reasoning and self-refinement, consistency across generations, strong ability to blend images). This was enough to earn them the #3 spot in the image generation arena.

Unfortunately, they decided to launch the model immediately inside Instagram, opting every public account into a new feature that allowed anyone to generate arbitrary images using their face.

They also teased Muse Video, which is part of the same model family but not yet publicly available.