Is Claude Max Worth It? I Left, Tried the Cheaper Roads, and Came Back

Max is an all-you-can-eat buffet with one rule: you eat with the house cutlery.

Claude Max is worth it if you code with frontier models for most of the day and Claude Code is a harness you can live in. It is the cheapest way I know to run Claude for hours. It is not worth it if your usage is light (Pro covers it), if your agents run unattended (an API key fits better), or if your work depends on a harness of your own, because the plan does not cover third-party harnesses.

I paid for Max 20x, left it in May 2026 for OpenCode and open models, moved those agents to a flat-rate provider in July, and came back. Nobody made me. The arithmetic did. Here is the arithmetic, and the part of it I still do not like.

What Max actually buys

Claude plans, October 2026

$20

Pro$17 a month billed yearly

$100

Max 5x5x Pro usage per session

$200

Max 20x20x Pro usage per session

$4 / $20

Opus 5.5 on the APIinput / output per million tokens

Read the plan for what it is: prepaid model capacity at a discount, with a usage limit instead of a meter. Part of why Anthropic can price it below the API is efficiency: Claude Code is written to hit the prompt cache, so the same work costs Anthropic less compute inside its own harness.

The question “is Max worth it” has a formula, and it is not about features.

The only math that matters

What Max saves you=your month at API rates−the plan price

You do not have to guess the first term. ccusage reads the session logs that Claude Code, OpenCode, pi, Codex and a dozen other agents keep on your machine, and prices them at API rates (npx ccusage). Run it over a normal month, not your busiest week, and compare.

My numbers, from this laptop

I ran ccusage on the machine I am writing this on. Claude Code keeps 30 days of logs by default, so my Max months before the switch are gone. The round trip is all there.

ccusage on my laptop, October 1, 2026

$6.23

all of AugustOpenCode and Hermes on open models, 138M tokens

$626

four days back on Claude CodeSeptember 28 to October 1, about 2B tokens

$415

September 30 aloneparallel sessions writing this cluster

$200

Max 20xfor the whole month

Read the two ends together. In August, on a flat plan for open models, my usage was worth $6.23 at API prices, and the cheapest plan on that provider cost more than ten times that. I was paying for capacity I never used. Back on Claude Code, four days of real work came to $626 at API rates, against $200 for a month of Max 20x.

That week was heavier than most. Even at a tenth of that pace, the plan pays for itself. So the formula above stops being theory, and it gives me a rule I did not have before: for open models at my volume, pay per token; for frontier models at my volume, buy the plan.

The condition: the house cutlery

Since April 4, 2026, a Claude plan covers Anthropic’s own products: Claude Code, Cowork and the Claude apps. A third-party harness that logs in with your plan, like OpenClaw or OpenCode, bills as extra usage at API rates. Anthropic started with OpenClaw and said the policy applies to all third-party harnesses. OpenCode’s own docs now say Anthropic explicitly prohibits using Claude Pro or Max through its plugins.

Be fair about the reason, because it is a real one. Boris Cherny, who leads Claude Code, said the subscriptions were not built for the usage patterns of these tools. A harness that ignores the cache burns far more compute for the same work, and a flat price cannot absorb that forever.

Apps built on Anthropic’s own Agent SDK are a separate case. Anthropic announced a separate monthly credit for them, then paused that change on June 15, so for now they still draw from your plan. Check the help center before you build on it.

✓

What you keep on Max

1. 01Skills, hooks, subagents and plugins
2. 02CLAUDE.md and path-scoped rules
3. 03MCP servers and your own CLI tools
4. 04Plan mode and the permission modes

✗

What you give up

1. 01Running another harness on the plan: pi, OpenCode, Hermes
2. 02Per-phase open models with their own permissions
3. 03A loop you can read and rewrite end to end
4. 04Moving the plan to a different harness when one ships

That right column is the part I do not like, and I will not pretend otherwise. Owning the harness is why I left. The heavy customization, one agent per phase with its own model and its own permissions, is the work I enjoy most.

What leaving taught me

The full migration is in OpenCode vs Claude Code. The short version has two lessons.

The first one is good news. Open models are good enough for a lot of the work when the harness is strict. My OpenCode pipeline ran explore, plan, build, review and fix on four open models, with permissions doing much of the work: explore could not edit, plan had no shell, review could not fix.

The second one is why I came back. The value of open models is their price. When I moved my agents to a flat-rate provider for open models, its plans landed close to the price of a Claude plan. Paying frontier money for non-frontier tokens is the worst of both. Take away the price gap and the frontier plan wins on plain arithmetic, even with the house cutlery.

The four ways to pay for coding tokens

Every option on the market is one of four shapes. Each gives up something different.

Four ways to pay, October 2026

For two years the second row did not exist in any official form. You picked two of frontier model, your own harness, and a flat bill.

Shape	Model	Harness	Bill
Frontier plan, vendor harness	Frontier	Theirs: Claude Code on Claude Pro or Max	Flat, with limits
Frontier plan, your harness	Frontier	Yours, from a partner list: ChatGPT Plus or Pro in OpenCode, pi, Hermes	Flat, with a weekly limit per app
Open-model plan	Open	Yours: OpenCode Go from $10, other open-model plans	Flat, with limits per model
API key	Any	Any	Metered per token

The second row is the news. On September 29, 2026, at DevDay, OpenAI started letting ChatGPT Plus and Pro subscribers use their plan inside participating third-party tools, and the launch list includes OpenCode, pi and Hermes Agent. You set a weekly limit per app in ChatGPT’s settings. I have run my Base25 agents on Hermes with a ChatGPT Plus login through Codex since June, and the official route is the provider type I am evaluating now for coding.

Two asterisks before you cancel anything. The allowance is whatever OpenAI says it is this month: on October 30, 2026, the $200 Pro plan drops from 20x to 10x the Plus allowance, next to a new $500 tier. And frontier models from different labs are not interchangeable in a harness tuned for one of them.

Is Max worth it for you?

Max pays off if

• ✓Required:
  Your month at API rates is well above the plan priceRun ccusage over a normal month. If it lands below $100, Pro or Max 5x is enough. If it lands far above $200, Max 20x is hard to beat.
• ✓Required:
  Claude Code is a harness you can live inSkills, hooks and subagents cover most customization. If your workflow needs another harness on the same plan, Max cannot give you that.
• ✓Required:
  A human is at the keyboardPlans are sized for interactive work. Agents that run unattended, in CI or around the clock, belong on an API key with a budget.
• •Optional:
  You hit the weekly limit most weeksThat is the signal to move from 5x to 20x, not the five-hour limit on one heavy afternoon.
• ✗Anti-pattern:
  You want to cut cost by switching to open models on a plan priced like MaxThat is the trade I made. Frontier prices for non-frontier tokens. Cut tokens in the harness instead.

That last line is where the real savings live. Before you change plans, change what the harness feeds the model. RTK compresses shell output 60 to 90 percent on everyday commands, and a lean system prompt plus memory in files does the rest. On a plan that means fewer limit hits. On an API key it means a smaller bill. The stack is in where your agent’s tokens die, and the larger map of harness, model and plan is in the best AI for coding.

Claude Max, quick answers

Is Claude Max worth it?+

Yes, if you code with Claude for most of the day inside Claude Code: it is the cheapest way to run frontier Claude models for hours. No, if your usage fits in Pro, if your agents run unattended, or if you need a third-party harness, since the plan only covers Anthropic's own products.

Should I pick Max 5x or Max 20x?+

Max 5x costs $100 a month and gives 5x the Pro usage per five-hour session. Max 20x costs $200 and gives 20x. Both have weekly limits on top. Start at 5x and move up if you hit the weekly limit most weeks, not because of one heavy afternoon.

Is Claude Max cheaper than the API?+

For heavy daily use, usually by a wide margin. In my case, four days of Claude Code came to $626 at API rates, against $200 for a month of Max 20x. Measure your own: ccusage reads Claude Code's local session logs and prices them at API rates. If a normal month costs less than the plan at API rates, the plan is not worth it for you.

Can I use Claude Max with OpenCode, pi or other harnesses?+

Not on the plan's usage. Since April 4, 2026, third-party harnesses that log in with a Claude plan bill as extra usage at API rates, and OpenCode's docs state that Anthropic explicitly prohibits using Claude Pro or Max through plugins. Use an API key for those harnesses.

Is ChatGPT Pro a better deal than Claude Max?+

They solve different problems. Since September 29, 2026, ChatGPT Plus and Pro can power participating third-party harnesses such as OpenCode, pi and Hermes Agent, which a Claude plan does not allow. But OpenAI cuts the $200 Pro allowance from 20x to 10x Plus on October 30, 2026, so compare your own usage after that date, not the old allowance.

Does Claude Max include Fable?+

Yes. As of October 1, 2026, Max 5x and Max 20x include Fable models for up to 50% of the weekly limits. Pro gets Fable through usage credits.

What happens when I hit the Max limit?+

You wait for the five-hour window or the weekly limit to reset, move to a higher plan, or turn on usage credits to keep working at standard API rates.