More from this creator
Other episodes by Kitty Cat.
More like this
If you liked this, try these.
Transcript
The full episode, in writing.
If you’ve been waiting for one benchmark score to tell you who’s winning AI, recent updates offer a different picture. As of October 14, 2026, frontier labs are competing not just on what a model can answer, but on what it can do for you—and how cheaply it can do it. Here’s a tour of five major providers, starting with the most recent dated product update.
On October 13, Google said Gemini 4 Argon was now available in the Gemini app for Google AI Pro and Ultra subscribers. Google describes stronger reasoning and coding, but the bigger product story is the move from chat toward tasks. A toggle lets users ask Gemini to work across connected apps, including services for shopping, reservations and design. Google also says its Antigravity coding interface is available inside the Gemini app on macOS. Picture the shift: instead of asking an assistant how to book a table, you ask it to help get the booking done. That’s Google’s pitch; how smoothly those multi-step tasks work in practice is a question users will have to test.
Meta’s October 12 announcement points in a similar direction. Its new Muse Wonder and Muse Wonder Full models are built for agentic work, such as interpreting a lease or building an app. Meta says they have Active Memory, intended to carry relevant details between conversations, and Active Focus, intended to help them recover when a task goes wrong. At the announcement, Muse Wonder was available through the Meta Model API and Muse Code, while Wonder Full was available to Muse Code subscribers. Meta said a fuller safety evaluation report was still to come. That matters because an assistant working across messages and documents needs more than persistence: it needs clear boundaries around private information and permission to act.
Anthropic is making a different part of the agent equation cheaper. On October 7, it introduced Claude Haiku 5.5, its small model for fast, high-volume jobs: summaries, classification, support interactions and narrowly scoped work delegated by a larger model. Anthropic estimates that typical use costs about seventy-five percent less than Haiku 4.5. It also cut Sonnet 5.5 cache-read prices in half. Why care about a cache discount? An agent may revisit the same instructions and documents over and over; reducing the cost of those repeated reads can change what’s practical to build. Anthropic still positions its larger Sonnet and Opus models as better choices for complex coding tasks. The point isn’t that the smallest model wins everything. It’s that a capable assistant may use different models for different steps.
Mistral and OpenAI show two more ways providers are competing. Mistral’s documentation lists Mistral Large 4 as an October public preview: an open-weight, multimodal mixture-of-experts model with a one-million-token context window. Open weights give developers another option for how they deploy a model; that is a different proposition from simply accessing a provider’s API. Meanwhile, OpenAI’s developer documentation describes GPT-6 Luna as an efficient model for focused, high-volume tasks. It lists a context window of just over one million tokens and support for tools, including web search and computer use through the Responses API. That documentation does not establish a launch date, so the update here is what OpenAI currently lists—not a claim that Luna launched today.
Put these updates together, and the frontier looks less like a single race for the smartest chatbot. Google and Meta are pushing assistants toward sustained tasks. Anthropic is lowering the cost of the small, frequent calls those tasks require. Mistral is offering an open-weight route, while OpenAI documents an efficient model with a broad toolset. The question to watch next isn’t only, ‘Can it reason?’ It’s, ‘Can it finish the job, at a sensible cost, without crossing a boundary it shouldn’t?’