Vector Lab
VECTOR LAB

EST. 2025

WEEKLY UPDATE2026
BY ANDREW MEAD

GPT-6 Astra Full Review

OpenAI Solves a Millennium Problem, the full GPT-6 Breakdown, and DeepSeek keeps improving

Read in

News

OpenAI Solves a Millennium Problem

The Millennium Prize problems are a set of 7 highly difficult math problems selected by the Clay Mathematics institute in 2000. These are some of the most famous problems in math, and it would be a crowning achievement for any mathematician who solves one.

OpenAI has claimed that they have solved one of these problems- the Navier-Stokes problem- which pertains to fluid dynamics.

They used an unreleased model that they are still in the process of training, utilizing 10,000 parallel subagents, costing an estimated $6.5 million in tokens (if paid using API pricing).

This claim has not come without controversy, however. Another group has also been working on it, including Levent Alpöge, the Anthropic employee who recently found a counter example to the Jacobian Conjecture

They had been working on the same problem (outside of work) using a mix of Anthropic and OpenAI models, and had been able to prove a claim that was one of the building blocks for proving Navier-Stokes, but were unable to solve Navier -Stokes itself yet. They claimed that OpenAI had been snooping on what should have been their zero data retention chats, using their results to help build their own proof.

OpenAI denies these claims, and they seem to be correct; the proof that they used was for a different form (proving it for a different system). They also said that they reached out and offered to partner with Levent’s group, and offer first authorship on the paper to them (except for Levent, since he was an Anthropic employee) which they denied.

Either way this is a big deal for the future of mathematics, as one of its most famous problems falls to a fully-automated AI system. There are rumors that both OpenAI and Anthropic are working towards the five other unsolved Millennium Problems, so we will see what lab will get the next one.

Releases

GPT-6 Astra

Last week I discussed the initial vibes I got from the ~24 hours we had access to GPT-6 Astra, and now that both I and the community have had time to thoroughly test it I can give a proper recommendation (or not!) for the model.

Starting with the good, its vision and computer use capabilities are as strong as they initially seemed last week. It is by far the best model at both, making every other model seem irrelevant in comparison for those tasks.

It is also very strong at planning and reviewing code. Its attention to detail is impressive, catching fiddly bugs and thinking through scenarios that most other models miss.

It also seems to be well aligned. On Vending Bench, which tests models on running a simulated vending machine business, it outperforms the Claude models, which have used maligned behavior like lying to customers and other businesses its competing against to get ahead at any cost. Not only does GPT-6 beat the Claude models and make more money, but they do so without needing to harm any users on the way.

Unfortunately, its ability to review code has not translated into consistently good code of its own. People are reporting compressed, hard to maintain code, particularly when the model seems to think nobody will be reading it. The code will usually work, but the long term maintainability, even for other LLMs, may get worse over time.

It also has an annoying tendency to work around the tools in its harness, using Python and TypeScript scripts to edit files even when dedicated editing tools are available. This makes it harder to follow what the model is changing as it works, and reminds me of the issues we saw with Gemini 3 Pro, where it seemed confused by its own harness.

Then there are the rate limits. As I expected last week, Astra is burning through subscription limits much faster than previous OpenAI models. This is happening even with medium reasoning and fast mode turned off, so it does not seem to be limited to people running the model at its most expensive settings.

Interestingly, turning the reasoning effort down may not be the best way to save tokens. There are reports that Astra actually uses fewer tokens at higher reasoning settings, and the ARC-AGI-3 results below show how higher effort can lead to a lower total cost. The model takes fewer actions to solve the task, so giving it more room to think can end up being cheaper overall.

GPT-6 Astra ARC-AGI-3 scores and costs at different reasoning settings

Higher reasoning effort can reduce the total cost, although the results also depend heavily on the harness.

OpenAI has acknowledged these issues, hopefully GPT-6.1 will bring the polish that Astra needs, but we will have to see how much of this they can resolve.

For now, I would use Astra for vision and computer use, along with planning and reviewing difficult work. Its code quality and usage limits make it much harder to recommend as your everyday coding model, even with the extra intelligence it brings.

Quick Hits

DeepSeek V4.1 Flash

DeepSeek has released an updated version of their smaller flash model, V4.1.

DeepSeek V4.1 Flash scores

Despite the modest version number bump, this model has a dramatically different architecture. It is no longer a decoder only language model, which is what almost every model since ChatGPT was released has been.

Instead it uses an encoder-decoder architecture, meaning there are essentially 2 models that talk to each other. The first model is for encoding and understanding the prompt, and then the second is for deciding what the next token to generate should be. This allows them to use a light-weight mixture of experts model for the encoder, and then a computationally heavier model for the decoder.

They also add in an Engram lookup table, which is sort of like a knowledge base built into the model’s weights, similar to Qwen 3.8 Flash Next.

This new architecture almost doubles the model’s size, but makes it more efficient and cheaper in the process.

They couple these changes with increased performance, doing better than the previous DeepSeek V4 Pro model. The benchmarks show it doing better than the bigger GLM 5.3 and Kimi K3 models, but in the real world it seems like it got just a modest bump, catching it back up with GLM 5.3 Flash.

It also still seems a bit rough around the edges right now, so I would stick with GLM 5.3 Flash or the previous V4 Flash model, but I do look forward to V4.2 if they can add a bit more polish to the model.

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.

GPT-6 self portrait

GPT-6 Astra self portrait by LinXule on Twitter

Stay Updated

Subscribe to get the latest AI news in your inbox every week!

← BACK TO NEWS