Exactly one week ago, I had to release the bug-fix version 2.5.1 of ComicStripBrowser because Marmaduke was broken. The culprit was a typo that actually meant I inadvertently pasted a date in a URL when editing the description for the comic. I literally don’t know how this could happen, but it did happen.

All I knew is that Marmaduke worked when I tested it, but then it didn’t. I’m not a big fan of it, and I didn’t access it after the release of version 2.5.0. But I also knew that Gemini CLI could help me, so what’s a lazy guy to do if not fire up Gemini CLI and ask it to find the reason for this malfunction?

Enter Antigravity CLI

Well, Gemini CLI had a message for me:

It read:

Gemini CLI is transitioning to the new Antigravity CLI for Google One and unpaid tier (Gemini Code Assist for individuals) users.

What’s changing: We are unifying our tools into a single, multi-agent platform called Antigravity, with Antigravity CLI now available. Gemini CLI will stop serving requests for Google One and unpaid tiers starting June 18th. Please migrate to Antigravity CLI before this date to avoid disruption to your workflow.

To learn more visit: https://goo.gle/gemini-cli-migration

So I visited that page, dated May 19: An important update: Transitioning Gemini CLI to Antigravity CLI.

This is what I understood, in a nutshell:

  • Antigravity CLI was made available to anyone starting May 19.
  • Gemini CLI will stop being available starting June 18, except for existing Enterprise customers.
  • Gemini CLI and Gemini Code Assist IDE extensions will stop working for:
    • Free users who use Gemini Code Assist for individuals.
    • Users of Google AI Pro and Ultra via a Google One subscription.
  • Users of Google Cloud Projects (GCP) via API keys seem to be unaffected by this change.

As usual, the communication is based on wrong documentation. Corporations don’t know shit about their own products.

Free users also had (and still have until June 18) access not only to Gemini Code Assist but also to Gemini CLI. Regardless of what the documentation says, Gemini CLI worked (and still works until June 18) via OAuth, but the daily limit is 250 API calls or less. No need for an API key!

Google One subscriptions are basically storage subscriptions (for Drive, Photos, and Gmail) that come with higher limits for the Gemini chatbot app (in a browser and in the mobile apps). You don’t have to use an API key if you don’t want to, so you won’t be billed when usage limits are exceeded. As a side note, I have no fucking idea how to create a Gemini API Key without importing a project into Google AI Studio. Since I don’t want to import anything into fucking Google, I’ll probably never use a Gemini API Key. More important, though, is that the announcement didn’t specify anything about the cheaper Google AI Plus subscription.

Corporations are dumb and shitty by definition. They have tens of thousands of employees, yet they all seem clueless.

That said, my previous experiences with Gemini CLI were satisfactory, so I decided to upgrade to Antigravity CLI, especially as I wanted to test the limits of Google AI Pro (via Google One).

Antigravity CLI made it clear that I can only use an API key with a Google Cloud project. I suppose the planet is populated by idiots who import their projects to Google AI Studio, or whatever else can be used for a Google Cloud Project.

OAuth still works, so who needs an API key? What I don’t understand is how those enterprise customers host their code as a Google Cloud project. I suppose they trust Google.

The color schemes were all shitty (meaning, with low contrast), so I stayed with the default one, which is also shitty:

But, wow, what a choice of LLMs! 😲

Claude Sonnet and Opus (!), plus the quantized, open-source GPT bastard, gpt-oss-120b!

The limits are obscure, and Gemini-the-chatbot said they’re all bound to the famous 5-hr limits, plus a weekly cap. It couldn’t find any reliable limits other than the old ones, but it acknowledged that using a non-Gemini model would probably hit the limits “significantly faster” than when using Gemini. Thank you, Captain Obvious, for your detailed, confident answer!

By the way, those old limits are wrongly saying that a free user has 250 requests via an API key, and 1,000 Gemini Code Assist requests via OAuth, when reality proved that about 250 requests can also be obtained via OAuth in Gemini CLI! But those 1,500 requests per day for Google AI Pro plans most likely don’t apply as described, what with the new 5-hr limits for the chatbot.

I set the model to “Gemini 3.1 Pro (High),” and I noticed that the quotas are non-descriptive. 100% of what? Of the daily quota, of the weekly quota, of the 5-hr quota, of the monthly quota?

It identified the broken URL in data_models.py, and the most detailed usage report it could issue was this very nice and detailed, but also extremely useless piece of crap:

Well, I suppose I’ll keep using Antigravity CLI when in need of help. 🤷‍♂️

I was happy with Gemini CLI. Why the fuck did they need to discontinue it? (Except for Enterprise users.)

Alternatives? What alternatives?

I’m not sure whether I hate ChatGPT or I despise it, but I suppose both are true. So, Codex is not for me, and it will never be.

Amazon Kiro is still suspending or blocking or closing or whatever they randomly do to unlucky accounts.

Anthropic also randomly “finds” that Claude was “used by children,” so it also blocks those accounts (what else?).

I was quite happy with Qwen Code while they still had a free tier for it. Paying doesn’t seem a good choice after having such complaints on Reddit: I finally figured out how to sign up for Qwen’s new subscription plan… and… it was a real disappointment…

Maybe that guy was wrong. But the comments still don’t look promising. So, if “it lasts 1.5 days if you work 8 hours a day, and it costs $30,” what more could be said?

Kimi has its own detractors, too. One of them: Kimi K2.6 is not worth the hype — my real usage stats. Apparently, he didn’t purchase from Moonshot. (Here’s a K2 Vendor Verifier.) But what’s worse is this: kimi-code: Access terminated. Review our Community Guidelines (bug?). Et tu, Brute?

Idiocracy at its best.

Oh, but wait: there is more!

Le Chat Mistral is now Vibe!

In the LLMs field, France represented Europe. And now they’ve just proven themselves retarded. How in the fucking world to replace the pun on the word “cat” with the most detested concept, “vibe coding”? Oh, the vibe of using Mistral, eh? Not vibe coding…

But you do need “Vibe for Code”!

Oh, but the vibe, man!

Still, there’s a thing called Codestral (“Use Codestral via your favorite Code completion tool for free.”) and the new Mistral Vibe CLI (duh). Could the latter be of some use?

I mean, the free tier.

Scale as you go plan:

Pay for what you use. Billed monthly or drawn from your credits.

  • Pay-per-use; see pricing
  • Access to frontier models
  • Access to creation and deployment of agents
  • Access to the fine-tuning API
  • Max. 6 requests per second, scalable on demand
  • Max. 2M input tokens per minute, scalable on demand
  • Max. 10B input tokens per month, scalable on demand
  • 20M tokens per minute, and 200M tokens per month for Mistral Embed
  • Access to all features, including previews and beta features

The next invoice will be triggered after €10 of usage.

However, the “see pricing” link doesn’t show any price. They’re French. Maybe that explains something. Nice colors, but not much more.

On the other hand, the plans page says that Pro offers

10x more usage of Vibe CLI
10x more API usage

than the Free plan, for €14.99 + VAT (the VAT-inclusive price will be displayed).

Well, if that’s ten times more than something, let’s find out what that “something” means!

That “something” means virtually nothing, I’m afraid.

The 3 models are 2, or actually one worth trying:

Notice that it says I should have 200k tokens, right?

Oh, there’s also an adjustable thinking level:

Some nice commands:

Now, for a test. I didn’t really need anything, but I had an idea. Every time I asked an AI (Amazon Kiro, Qwen Code, Gemini CLI, Antigravity CLI) to do something, the AI had to analyze all that Python code (which is not that much, really, but still) to figure out the architecture and why it does what it does. So I wanted to ask Vibe CLI to document the app’s architecture in a Markdown file so I could reuse it and save tokens.

Vibe CLI started crunching files, but after a few minutes it said, “Rate limits exceeded. Please wait a moment before trying again, or upgrade to Pro for higher rate limits and uninterrupted access.”

I waited a minute and tried again. A couple of minutes later…

So, 86,015 prompt tokens, and 86,411 total LLM tokens? Not much!

Let’s try again:

For fuck’s sake! It literally didn’t do anything more! Not a single extra token!

Also, $0.1320 cost in the free plan and without having a payment added?! Then, what does it mean that Pro is “10x more than Free”?! WTF are they smoking in France?!

Not only is Vibe CLI useless in its free tier, but how can one pay to get ten times more than nothing? It cannot do anything for free!

Mistral is a horrendously bad joke. They even forgot to replace their cat with an image of… something that could stand for “vibe.” Maybe a leaf of cannabis.