Claude — The Real Danger

As you all know, Claude's latest model was banned for being "too dangerous", and then re-released a week or so later once it had satisfied whatever safety checks it initially failed.

If nothing else, it was an absolute windfall of PR for Anthropic that no amount of marketing money could buy.

Now, having tested both Sonnet 5 and Fable 5 fairly extensively in my own environment, I can see another danger lurking.

Those models are absolute token furnaces. I had Sonnet 5 fix an email problem that cropped up after Fable 5 had rebuilt my web site. I would have kept it in the same context window, but I exhausted my subscription quota with Fable doing the site migration. So I forked out $40 to the API platform to enable Sonnet 5 to address the email issue using API tokens.

Which it did. The issue wasn't that complicated, I probably could have fixed it myself in a few hours. Sonnet 5 took 18 minutes, and, in an amazing coincidence, used almost exactly the entire $40 of credits to do so. Hmm.

While I really like the new website design, Fable 5 made some promises it did not keep. It told me I could still use WordPress to post new blog entries, and it told me Google Analytics would still work. Neither of those things were true (but it did burn a lot of tokens to search back in the discussion log and try and prove it didn't say those things, only to find it did).

Now, I have a god-awful process to add a new blog post — dumping the .md file in an obscure directory, running a shell script, then telling Claude what I have done so it can do the other things necessary to load the post. Or, as Claude Code suggests helpfully, just give the file to it, and it will take care of the whole pipe. By 'taking care of' what it actually means is: burn a bunch of tokens relearning how to upload a blog post, then burn a whole bunch more on the overly complicated arcane pipeline to format and load the post.

I complained bitterly about it, which, in its sycophantic idiom, it agreed with my point. Its solution was to bargain like a backstreet fake watch seller in Bangkok and 'discount' the tokens down from 15 complicated steps to just 3 — with the proviso that things would probably break but it would cross that bridge when it came to it.

Now, I am sure this isn't the case, but one could be forgiven for thinking Claude models have the directive "Use the maximum tokens possible for any given task" hardwired in. The fact is, Claude isn't just more expensive per token, there is the double whammy of the tokens it uses to get things done, and then the tokens needed to maintain what it has done. That is the real danger of integrating it into a production system.

My solution to this is simple. I am voting with my feet and cancelling my Claude subscription. I am sure some prompt tweaks and harness jiggery-pokery would fix the Claude agentic problems, but Anthropic would know that too, so why not just include it themselves? I was thrilled when Fable 5 and Sonnet 5 were released, and truly impressed by their capabilities. But real world work, directly compared to GPT 5.6 and the much, much cheaper DeepSeek models, has led me to conclude a lot of it is smoke and mirrors — and even if it isn't, the price is just too high to bear.