Token Pricing
Getting hooked
A year ago, the available AI tooling was not good enough to be helpful in my research, but now, like most of my colleagues, I spend a large chunk of my workday talking to Claude.
Initially, Anthropic’s $100/month Max plan was sufficient, but then I started hitting limits, and it’s rough to have your Claude taken away abruptly when you hit your usage cap. So I had to upgrade to the $200/month Max plan. I upgraded without hesitation, but it was a bit scary because it made it clear just how reliant I had become on Anthropic.
But Anthropic’s Max plan alone isn’t sufficient. There’s an issue that when I want to script actions that use Claude, I need an API key, and the Max plan does not provide API access. You need a completely separate, independent Anthropic account to get API access. So I have one of those too. This is so bizarre that I can only assume that it’s the consequence of being a rather young startup whose main focus is model development rather than user experience.
Although I have an account on the API plan, most of the time I just use the Max plan through Claude code.
There are a few times when I use the API plan.
- Scripted calls: I have a workflow that takes an academic paper, extracts the citations to a list, then for each one, it calls an agent to check whether the citation is accurate. The script is python-based, and uses the API to spin up each independent agent. Yes, I know I can just throw the paper into Claude Code, and ask it to check all the citations, but the default Claude Code will not spin up agents aggressively enough. If I do it directly through Claude Code, when I have a paper with 50-100 citations, Claude Code will spin up five or so agents and check a bunch of citations, but it’s almost impossible to get it to give you full coverage. By contrast, with a script, you’re guaranteed to get full coverage (but you have to pay per token). Similarly, I have a script for checking mathematical proofs. For each proof in the paper, it extracts the proof and spins up an independent agent to check the proof and write a report. This still gives much better coverage than simply asking Claude Code to read the paper and check the proofs.
- Claws: When OpenClaw hit big in February 2026, I decided to try out IronClaw. I got it set up, and I connected it to my Matrix server so you could chat with it. This required an API key, but it wasn’t very helpful, so I didn’t burn too many tokens.
- Fable: When Fable came out (again) in July, I found that I was blowing through my Fable budget on the Max plan, so I started using the API plan to satisfy my Fable fix. And this is expensive.
Pricing
Reading people’s reports of overspending — one company reportedly ran up a $500M Claude bill in a single month, and Uber capped its AI coding tools at $1,500 per employee per month — I realize just how cost-effective the Max plan is. For $200/month I get almost as much (non-Fable) use as I need. But Anthropic’s recent price hike made me uncomfortable. When Anthropic made me jump from the $100/month plan to the $200/month plan I realized that I was hooked. There was no consideration about whether it was “worth it.” When I started hitting my limits, I had to upgrade: to keep working the same way, with the same workflows, I had no choice. If Anthropic had jacked the price to $2000/month, I still would have paid (assuming they could have gotten the other frontier model providers to agree to a similar price hike).
But this got me wondering, just how much better is the flat-rate $200/month deal than buying your tokens a la carte through the API plan?
What’s the deal?
It’s tough to figure out how good a deal the Max plan is, because Anthropic doesn’t actually provide strict token caps. In July 2025, when Anthropic rolled out the weekly-caps they gave cap estimates in hours of usage (not tokens).
| Plan | Price | Sonnet (est. hrs/week) | Opus (est. hrs/week) |
|---|---|---|---|
| Max 5x | $100/month | ~140–280 | ~15–35 |
| Max 20x | $200/month | ~240–480 | ~24–40 |
These limits are intentionally vague, and seem to have gone down over time. Anthropic even seems to acknowledge as much, conceding that the 5-hour session limits now deplete faster during weekday peak hours. Although these estimates are vague, they seem fairly reasonable. I rarely burn through my limits (on Opus) before the end of the week, but as the week ticks down, I’m always dangerously close to being locked out.
Although the flat-rate Max plans are very vague, the API token pricing is very straightforward:
| Model | Input $/1M | Output $/1M |
|---|---|---|
| Fable 5 | $10 | $50 |
| Opus 4.8 | $5 | $25 |
| Sonnet 5 | $3 ($2 intro) | $15 ($10 intro) |
| Haiku 4.5 | $1 | $5 |
I basically never use Sonnet or Haiku, so let’s focus on Opus. Thankfully input tokens are cheaper than output tokens. In my Claude Code use, I consume over 100x more input tokens than output tokens. So, dropping the output tokens completely from my estimates, my Max plan is cost-effective if I consume more than about 40M input tokens per month. I definitely do that, but if you’re only using Claude casually it might be worth considering the API plan. Really, though, I think for any serious users the Max plan is a much better deal than the API plan, and people only use the API plan when they need API access (which is not available through the Max plan).
Enshittification
People have often pointed out the Silicon-Valley playbook: get people hooked on a subsidized service then jack up the price (and drop the quality). Cory Doctorow’s term enshittification is so popular because it captures this phenomenon so nicely. That’s certainly happening here.
Anthropic’s APIs are noticeably slower than they were a few months ago, and I’m bumping into my limits more and more often. But I’m stuck, and it’s scary. Let’s hope the open source models catch up soon. I’ve been hearing good things about Z.AI’s GLM5.2. It’s currently 5x cheaper than Opus on DeepInfra, but I still don’t use it because Opus works better. And even running such a model yourself on US infrastructure isn’t a clean escape: NIST found that CCP-aligned censorship is baked directly into the weights of Chinese open models like DeepSeek. That means that even if you run the models locally (or on a trusted US provider) you will still likely get CCP-aligned responses (even though your queries will remain private).
This is the bind. We’re all hooked on Claude (or one of the other frontier models) and we’ll pay what it takes to keep access. The open models are the obvious way out, but they’re not there yet, and when they arrive they’ll come with values baked into the weights, same as the closed ones. There’s no version of this where you’re stuck with someone else’s model.