Gemini finally fixed, Google and Anthropic shot themselves in the foot, and OpenAI is farming karma - weekly news.

Gemini finally fixed tool calling

I just wrote in my last post about OpenClaw that Gemini is a terrible assistant model that can’t handle anything - and then Google went and released Gemini 3.1 Pro, a model specifically trained on tool calling. They finally fixed Gemini’s biggest weakness.

I’m already using it in OpenClaw, as much as the limits allow. Google still has stupid limits - 250 requests per day - but even so, you can clearly see Gemini performing well: it actually calls tools now. It still hallucinates way more than GPT, by orders of magnitude, but I like the model’s tone and openness more. It’s gotten better at following instructions, and now the Markdown identity files actually influence the output - roughly on par with Sonnet.

GPT-5.3 is coming

Word is it drops on Thursday, but I wouldn’t trust the rumors - they’ve been floating around for weeks now. Looks like the model release was delayed so GPT-5.3 could be better trained on non-coding management tasks: OpenAI clearly saw Gemini and Anthropic focusing on that direction and noticed OpenClaw’s success.

Hopefully they also fix the identity and communication style - OpenAI is currently the last provider that allows subscriptions for things other than coding, and it’d be great to use the subscription with a personal assistant too.

Along with the GPT-5.3 release, there’ll most likely be a Pro Light subscription for around $100 - something a lot of people have been waiting for. I’m definitely going to stick with it: looks like the best ROI option.

Google and Anthropic shot themselves in the foot

Google and Anthropic started aggressively banning subscription use anywhere outside their own tools, and for anything other than coding.

Google is just silently banning accounts, cutting off access to Gemini CLI/ Antigravity. Anthropic updated their privacy policy and terms of use, prohibiting any token usage outside Claude Code and strictly for coding purposes.

Against the backdrop of OpenAI actively supporting OpenClaw, buying it, and hiring Peter Steinberger, this looks like complete nonsense. OpenAI has been farming karma hard for the past six months - and it’s working.

Twitter shutting down the free API

Twitter cut the free API subscription, which already had a pretty low request limit. Now you have to pay for every request, and in March the free credits they gave to people who used that subscription expire. My blog post to tweet automation will probably stop working, unless the cost turns out to be totally negligible.

Starting work on my own AI assistant

Seeing all of OpenClaw’s flaws and all its potential, I’ve started working on my own AI assistant app. I’m going to rework the chat in Respawn into a standalone personal assistant app with a new architecture that I’ve mostly already figured out. I want to turn this into an experiment.