The End of Claude Code Subscriptions

Claude Code subscriptions had an insane run. For a time the deal was simple: pay for Max, point Claude Code at a repo, and stop worrying about API bills. It was close to a flat rate for serious coding-agent work.

That bargain is running out.

On February 24, 2025, Claude Code was announced as a limited research preview, and the early product really was something special. It could read the repo, edit files, run tests, and run a task through the terminal. Anthropic had developer goodwill because Claude Code felt like the tool serious people used.

The subscription worked because the limits were blurry but generous. There’s five-hour windows and weekly ones. It does not sell a definite token entitlement. When the plan felt spacious enough for everyday work, no one cared much.

People know when the same plan starts to feel hollow.

Opus became the model people wanted for hard coding work. That was already a problem because Opus is expensive. Then Anthropic released Opus 4.7, stating that the revised tokenizer means the same input can map to about 1.0-1.35x as many tokens, depending on content type. It also said that higher effort runs can produce more output tokens. In simple English: the same type of coding session can take more of your allowance even before it feels like you did more work.

My logs reflect that feeling. That’s how I work, not some universal law, but suddenly my Claude subscription seemed to be about 5x more expensive.

Sonnet 5 was supposed to be the escape route. Make the default model cheap, good enough, and plentiful, and the subscription still works.

That's not what happened.

Anthropic says that Sonnet 5 can rival Opus 4.8 capability in some cases, at standard pricing of $3/$15, when you push it into the right effort range. But at this effort it will use so many tokens that pricing is at or above Opus for the same task. That is not a strong cheap-model story. If you try hard enough it might get close to Opus, but it's not the model I want for the hardest coding jobs and it's also not cheap at all, so why use it at all?

Next up is Fable 5.

Fable 5 is the true premium Anthropic model. Anthropic says it’s included for Pro, Max, Team, and some Enterprise plans through July 7, 2026, and only up to 50% of weekly usage limits. Then it goes to usage credits. The API costs $10 per million input tokens and $50 per million output tokens.

That changes the subscription equation.

If the best model is metered, the best way to buy the best Anthropic coding experience is no longer the subscription. If Sonnet 5 is the default for the subscription, then the subscription is no longer the best way to buy frontier coding quality. The subscription is no longer the best way to buy predictable work with an opaque meter.

Codex is now the clear default for what I do on the coding-agent side of things. OpenAI’s Codex subscription stack has GPT-5.5 (which for me feels at or above Opus 4.8 quality) and in my experience it gives you way more actual work per paid account. It’s quicker, too. If I want cheap everyday automation, I'd rather push more of that towards open models like GLM-5.2, Kimi-2.7-Code or DeepSeek-v4-pro. If I need the best Anthropic model for a task that justifies the spend, I can pay for Fable 5 explicitly via credits or API. And if i need a close to SOTA model on a cheap subscription I use Codex with GPT-5.5.

Which makes Claude Code subscriptions the absolute worst middle ground.

Not the cheapest one.

Not the most transparent.

Not the best-model plan.

Soon, not even the place where Anthropic’s best model really lives.

This looks like the probable outcome of Anthropic turning toward enterprise and API economics just as it prepares for public markets. On June 1, 2026, Anthropic submitted a confidential S-1. Good. It can create that business. But the goodwill that got enterprise buyers to accept Claude in the first place was from developers happily using Claude Code on subscriptions.

If you burn that goodwill, you weaken the enterprise story as well.

The old Claude Code subscription was great because it was like a flat rate for real work. When the best model is metered, the cheap model is not obviously cheap in practice, and the meter is still a black box, the subscription is no longer justified.

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com