Parameter Update: 2026-07

Pretty slow week, all things considered, so post is also pretty short. Happy to hear OpenClaw is moving to a more stable envionment though.

OpenAI

Hiring Peter Steinberger

Ever since the OpenClaw hype cycle really got of the ground, there has been speculation about it's creator, Peter Steinberger, potentially selling the project or aligning himself more directly with one of the big labs. This week, we finally received confirmation of what the plan looks like: Steinberger joins OpenAI, and OpenClaw transitions into an independent foundation. Feels like this is not the worst way things could've gone!

GPT-5.3-Codex-Spark

Over the last weeks, we've got a fair share of model speed improvements - at various costs to intelligence or pricing. This week, OpenAI launched the fastest one yet - GPT-5.3-Codex-Spark, delivering over 1K tokens/second. While I still see few current use cases that require this speed, it feels like a very sensible way to move if your goal is to push the test-time compute scaling graphs further. For now, it comes with reduced intelligence and is positioned for "interactive, targeted edits" instead of long-running tasks. From an infrastrcture POV, this one is particlarly interesting as it's the first time we're seeing OpenAI serving a model on non-Nvidia hardware at any kind of scale (-Spark is running on Cerebras HW). Currently, access remains limited to Pro users, so I'm very interested in seeing if OpenAI is able to scale this further (and, ideally, deploy a more intelligent model at scale).

Deep Research Upgrade

Historically, I've not been a heavy user of OpenAI's Deep Research - the tool failed regularly, results often felt incomplete and the entire thing took forever. This week, OpenAI is trying to solve at least some of these pain points. For one, the tool is finally upgraded to GPT-5.2 (turns out it was still using o3?). On the other hand, you can now specify specific websites you want searched and/or connect to other external tools through MCP.

This feels like a pretty natural direction for the tool to evolve - having a compute-heavy agent that can aggregate info across specific sources is obviously more valuable than just aggregating a few hundred websites.

Google

Gemini 3 Deep Think Upgrade

After the two model launches of the past week, it seems Google was not content to let Anthropic and OpenAI have the stage all to themselves. "Gemini 3 Deep Think" is a much worse branding than the leaked "Gemini 3.1 Pro" internal naming, but the underlying model still seems great - it beats Opus 4.6 and GPT-5.2 in all benchmarks we currently have (still no GPT-5.3 to compare against yet), sometimes by a solid margin, reaching SOTA across a series of tests.

In typical Google launch fashion, the model is currently limited to "AI Ultra" subscribers and API access is in limited early access. The two open questions, therefore: What will the thing cost and how slow will it be?