tl;dr
- Sonnet 4.5 is out, comparable to Opus 4.1, still worse than GPT-5 for coding
- GLM 4.6 is better than Sonnet 4 while being only $3 a month
- OpenAI released Sora 2, and is the best video generation model (join the Vector Lab Discord for an invite code)
- DeepSeek 3.2 hints at the future of LLM architecture
- IBM releases a set of strong, small, and fast open source models
- Thinking Machines has revealed their first product
Releases
Sonnet 4.5
Major release from Anthropic this week, as they dropped their Sonnet 4.5 model, showing promising improvements in coding and safety benchmarks.
Straight to the real-world performance though. Having used it for the last week and also read a bunch about what others are saying, this is not the major performance increase we were expecting and hoping for. It is definitely an improvement. The model feels similar in quality to Opus 4.1, but it still does not have that raw intelligence and attention to detail that GPT-5 has.
In my testing this week, I wouldn’t say the model is necessarily smarter, but more that it is less dumb, meaning that it does not make some of the silly mistakes or have as many oversights about its implementation as Sonnet 4.
This is also somewhat corroborated by Anthropic themselves, as in their safety report for the model, they mention that Sonnet 4.5 does not reach the “notably more capable” threshold that would require a brand new comprehensive assessment of the model for its potential harmful capabilities.
They also have not changed the pricing from $15 per million output tokens, meaning that it’s 50% more expensive than GPT-5 still. This, combined with all the other factors above, make this a rather lackluster “upgrade”. If you were using Sonnet 4 previously, then expect a slight boost from what you’re used to, but it is not leaps and bounds better by any stretch of the imagination.
GLM 4.6
Speaking of pricing, the price-to-performance agentic coding kings, Z.ai have released an upgrade to their GLM 4.5 model. If you haven’t heard us talk about this model previously, the GLM 4.5 and now 4.6 models are available from Z.ai for only $3 a month, are comparable to Sonnet in quality, and has a four times larger rate limit than the $20 Anthropic subscription. It also plugs directly into Claude Code, allowing you to keep all of your existing agentic coding infrastructure in place.
GLM 4.6 shows an impressive bump over the previous 4.5 model, and when matched up head to head against Sonnet 4 and other open source models, comes out on top. I have been using it the past week as well along Sonnet 4.5, and there is very little difference between the two.
Because of this, my current coding stack recommendation is Codex-cli with GPT-5-codex for all of the hard tasks ($20/month plan), and the $3/month GLM coding plan for easy and medium tasks. This combo will give you the best bang for your buck in terms of model intelligence and raw output.
Sora 2
OpenAI has decided to release their Sora 2 model in the opposite way that they did the original Sora. This time, directly releasing a way for users to go and access the model and play around with it, instead of dropping a few examples from the model and then disappearing, with no real model release in sight.
Although it is not on any of the usual public benchmarks, Sora 2 is very clearly the best video generation model out there right now. OpenAI has forgone the lawyers and safety filters and are directly allowing users to generate copyrighted content from the likes of Family Guy and SpongeBob.
The model has a very strong real-world physics understanding and scene composition capabilities It has a level of clarity and cohesiveness that none of the other models on the market now seem to have.
Similar to VO3 from Google, it also does the audio generation for your videos as well. I will say on this front, it is a little bit lacking when compared to VO3, but still very usable.
They also released the ability to add yourself to the videos as well as use your voice allowing for a lot of creativity and use in real-world video production.
But on the flip side, you can now generate videos of almost anyone doing illegal things. For instance, Sam Altman has made his likeness available on the app by default for everyone, and so there have been numerous videos of him performing illegal acts like stealing GPUs from the store, fighting people, and other such crimes.
Quick Hits
DeepSeek 3.2 Exp
DeepSeek has released yet another version bump to their V3 model, this time calling it 3.2 Experimental. The main highlight of this release is their new DeepSpeak Sparse Attention (DSA) architecture, which is a linear attention based transformer that drastically reduces the computation needed for long sequences.
This architecture promises to be relatively straightforward to train into your existing model. Expect to see this or another variant of sparse attention in the release of DeepSeek V4.
Thinking Machines LoRa
Thinking Machines dropped a blog post this week showing how LoRa, when used correctly, is identical to full fine-tuning. And then building upon this, they also released a platform called Tinker to allow you to go and fine-tune LLMs using LoRa’s, abstracting away all the infrastructure code, while still leaving you in control of the data, loss function, and algorithms being used.
IBM Makes LLMs
IBM has quietly been releasing some small open source lms that are fairly decent over the last few months and this week they released another set in their Granite series of models which are competitive if not better than similar sized Qwen3 models while also being two to five times faster.
ChatGPT Instant Checkout
OpenAI has just announced Instant Checkout in ChatGPT in collaboration with Etsy and Spotify, allowing you to purchase products directly on the ChatGPT website. They also released the Agentic Commerce Protocol that they used to power it, which is built on top of Stripe.
I don’t have too much else to say about it, but I thought this meme was funny, which is why I wanted to highlight this topic.
Needless to say, I won’t be using this feature anytime soon.
Finish
I hope you enjoyed the news this week. If you want to get the news every week, be sure to join our mailing list below.
