My annual Claude Pro plan renewed in late July. Opus 5 underperformed with a greater token cost, and by late August I asked Fin, Anthropic’s support chatbot, for a refund over the performance problems. It replied that my purchase “doesn’t meet our refund eligibility criteria due to falling outside of the timeframe for our policy”. Time for a creative solution.
Tagged Claude
Is GPT Losing High Ground?
OpenAI released GPT-6.1 Sol on September 29, only a week after GPT-6 Sol. OpenAI says it “delivers near-Astra performance at a lower cost”, at a fifth of Astra’s list price, and some on Reddit even go so far as to call it Astra minor. Cost-effective Astra-level performance is great news. But GPT’s sweet spot is increasingly the lower bound, and OpenAI is no longer competing at the flagship level.
Big Surprise Week from Claude and Codex: Opus 5.5, GPT-6 Sol, and Luna
Anthropic released Claude Opus 5.5 on September 22, the same day OpenAI released GPT-6 Sol and Luna. OpenAI is making its lower-end models cheaper, while Anthropic is still pushing upward. On price to performance, Opus 5.5 now beats Fable 5.1, which makes one wonder who would still use Fable.
Claude Talks Too Much
Claude’s answers have become incredibly verbose and full of tics, and the community has a name for it: Claudism. This is not a taste problem. Anthropic never calls Sonnet 5, Opus 5, and Fable 5 a family, but the condition seems rather hereditary. Claude is wordier than it was a generation ago, the extra words make the work worse, and Anthropic’s advice is to just prompt better.
How to Redirect Claude Code to AGENTS.md
I run Codex and Claude Code side by side, and each wants its own instruction file at the root of a repository. Codex reads AGENTS.md, the shared format that Gemini CLI and two dozen other agents read as well. Claude Code reads CLAUDE.md. If I had the luxury of writing AI usage off as a business expense, I would have settled on one. Since I can’t, I keep both, and the two files drift as the projects progress. A rule would get corrected in one file and not the other, and the two agents would then follow different instructions for…
App to Track Claude Code Limit Resets
Anthropic replaced its temporary 50 percent boost to Claude Code weekly limits with a permanent 25 percent one, leaving subscribers with 17 percent less in total. Everybody wants a reset, now more than ever.
Streamlining Analytics, Migrating with AI Pt. 5
Post-migration, I knew some jobs were left untouched. Not because of the amount of work involved — again, I don’t think the workload has meaningfully increased. For the first project afterward, I wanted to focus on a part of the website I hadn’t touched in a while: analytics.
Fable 5.1 Released
Anthropic released Fable 5.1 while Opus 5 was still new, making the company look on fire. Its recent commercial opened with an actual house on fire, so the joke is on us for not taking it seriously. Once more, the company claims its new model is outrageously exceptional. Yet vendor-drawn benchmark charts cannot tell us the true performance out in the field. They merely support the existing adage in the community: let the frontier model design the work, then let cheaper models grind.
Claude Max 20x Revealed as 8.5x
Claude Max 20x’s fine print has become the talk of the town, thanks to a viral tweet. Anthropic’s $200 plan does provide 20 times the Pro usage — but only inside each five-hour session. For the separate weekly limit, a proposed class-action complaint puts the upgrade over $100 Max 5x at 1.7 times, not four. Apply that multiplier to the advertised 5x baseline and Max 20x becomes 8.5x. Calling it “20x” is a euphemism.
Claude to Cut Claude Code Limit 17 Percent
Anthropic will cut weekly Claude Code limits by 17 percent on September 14. In the official tweet, the company announced it is “permanently raising standard weekly limits in Claude Code by 25%”, only to downplay the fact that it will replace the current 50% promotional increase. Existing subscribers will lose 16.7% of today’s capacity. In its defense, that 50% boost was temporary. Only, my workflow would not survive the upcoming cut, and community reports suggest many others would not either.
To Self-Host or Not To Self-Host, Migrating with AI Pt. 4
No doubt the idea of complete control of a blog at $10 a year would be appealing for many. And no doubt many would share the same concerns as I do. During the planning phase, some of the definitive downgrades I was able to come to terms with, some of the regained controls I was able to make use of, and the saved cost I was ready to file under an existing list of subscriptions. I believe there are three major concerns most bloggers may have: uptime, support, and technical debt to AI.
The Part Where I Pay, Migrating with AI Pt. 3
The plan finished baking. Reviewed, revised, the whole nine yards. I wasn’t going to execute the plan myself; I wanted to automate it as well. The only decision left was how.
App to Organize Coding Agent Output
As much as I rely on Claude Code while managing home lab and writing code for DIY projects, coding agents never meet the media’s perception of AI: “slaves to logic” as I, Robot has put it. Agents hallucinate, make mistakes and take them back, and more often than not, their output is a wall of text that is completely out of order. For Claude, the recent models print more than their predecessors.
Fable 5 Now Included in Max Subscription
Fable 5 saga is finally coming to an end with Anthropic deciding to include the flagship model in Max. Since I am on Max tier for this month temporarily, the announcement is disappointing in two ways: a. Fable won’t be available on a similar price tier as GPT 5.6 Sol, and b. Fable is simply too inefficient when it comes to token counts compared to Sol. Whichever way you cut it, you can get a better deal from Codex.
1Password Adds Claude Support
If you have been on a vibe coding frenzy thanks to Fable 5, 1Password has something to offer for you today. The company has announced a new feature that fills website credentials on Claude’s behalf. By the developer’s claim, the feature is designed to have zero-exposure to the agent — tasks are handled by AI, and the credentials do not enter the model or its memory. Each request still needs to be approved on 1Password. It is currently only available on Mac and Google Chrome. The detailed announcement and support document is in the link. Give it a read if…