If you pay for Claude Max or Ultra, there’s a line in your billing settings most people skip: free API credit every month, $100 on Max and $200 on Ultra. I read it as extra Claude at first, more room before I hit my limits. That turned out to be wrong, so I measured what the credit does buy.
📺 Watch the video: Claude Max Free API Credit: What It Pays For (It’s Not Claude Code)
What the credit is
You claim it in claude.ai under Settings › Billing › API credits, and it lands in a Claude Console organization linked to your plan. It tops up every billing cycle, and whatever you don’t use is gone when the cycle ends. In the Console it pays for API calls, Managed Agents, the Agent SDK and the Playground.
I think of it as a prepaid fuel card that only works at one station.
Test 1: Claude Code
I pointed Claude Code at an API key from the same account. It didn’t touch the credit. My paid balance went from $4.72 to $4.46, so I stopped it and logged back in with my plan.
The docs on API credits say the same thing: the credit doesn’t cover Claude Code, and it doesn’t change your plan’s usage limits. Claude Code keeps exactly the usage it had before. The credit is a separate tank for things you build.
The job
So I built something a small business would use: a support inbox for a little online shop, with 20 emails. There are refund requests, two bug reports, some questions and some spam. For each email the agent picks a label, finds the order number, looks the order up with a tool, decides the refund from the shop’s written policy and drafts a reply with the amount.
The policy has traps on purpose. Change your mind within 30 days and you get the item price back, but not the shipping. If it arrived damaged you get both, and final sale items only come back if they’re broken.
A hidden check marks four things on every email: the label, the order, the refund decision and the amount in the reply. An email only counts as correct if all four are right.
Test 2: which model
I ran the same inbox through my Agent SDK app on all three 5.5 models, two runs each.
| Model | Correct | Cost per inbox | Inboxes per $200 |
|---|---|---|---|
| Opus 5.5 | 40 of 40 | $0.134 | about 1,490 |
| Sonnet 5.5 | 99 of 100 (the race runs) | $0.061 | about 3,260 |
| Haiku 5.5 | 39 of 40 | $0.0067 | about 29,800 |
Haiku’s one miss was a tea towel set that didn’t match the photos. I’d counted that as a change of mind, so no shipping back. Haiku called it a wrong item and refunded the shipping too, and you could argue that one either way. Across the three models, the model was the biggest number on the whole bill.
Test 3: Batch
The Batch API charges half price if you can wait for the answers. Managed Agents can’t use it and a batch request can’t call tools, so there’s no agent here. Each email went in as its own request, with the whole orders file cached for an hour.
Batch got 19 of 20 (one email came back empty) at $0.040 an inbox, which is 5,000 inboxes per $200. It took about 5 minutes. The same requests without Batch got 20 of 20 at $0.082, more than any of the agents.
The race
Three routes ran with the same model (Claude Sonnet 5.5 at effort high), the same prompt, the same order tool and the same 20 emails. There were five rounds, one route at a time, with the starting order rotated every round.
- n8n 2.45.7, self-hosted free on my Mac: a Webhook, an AI Agent node with the Anthropic chat model, and a code tool for the order lookup.
- A Managed Agent: Anthropic runs the agent loop, and my script only answers its calls to the order tool. On top of the tokens you pay 8 cents for every hour a session runs.
- My own app on the Agent SDK, about 65 lines, with the same tool, on my machine.
| Route | Correct | Median time | Cost per correct email | Inboxes per $200 |
|---|---|---|---|---|
| n8n | 100 of 100 | 25 s | $0.0028 | about 3,600 |
| Managed Agent | 100 of 100 | 33 s | $0.0026 | about 3,850 |
| My Agent SDK app | 99 of 100 | 28 s | $0.0031 | about 3,260 |
Cost per correct email is everything a route spent divided by the emails that passed every check.
n8n surprised me in round 1. It looked up all 16 orders at once and finished the whole inbox in two calls to Claude. My app dropped one point in round 1: on a bug report it left the refund fields blank instead of marking them as not applicable. After that it was clean.
The Managed Agent won by a hair, and its hourly fee came to less than a tenth of a cent a run. My app came last because the SDK makes one extra small call every run to give the session a title.
Everything in the video cost $1.43 of the credit. The Console agreed to the cent, and my paid balance never moved.
What I took from it
How far the credit goes depends much more on the model you pick than on the route. On this job all three routes landed within a cent of each other, so pick the route for how you like to work: n8n if you’d rather not write code, a Managed Agent if the job should run while your computer is off, your own app if you want full control.
Then measure your own job once. This was one job on my Mac, and yours will use a different number of tokens.
Work out your own credit
I built credit-calc for that. It’s free and open source, and it needs no API key and no network. Give it the tokens one run of your job uses, and it prints the cost on each model with prompt caching and with Batch, how many runs your credit buys and the day it runs out at your pace.
git clone https://github.com/thecoderlog/credit-calc
cd credit-calc
node credit-calc.mjs --in 6600 --out 4400 --cacheable 6300 --runs-per-day 40 --credit 200
Use --credit 100 on Max. If you log one JSON line per run, node credit-calc.mjs --log your-runs.jsonl prints cost per run, cost per correct answer and runs per credit. No terminal? Open index.html in your browser; it has the same maths and prices.
Run your own job through it, then tell me in the comments how many runs your credit buys.
Code & video
- 📺 Full video: Claude Max Free API Credit: What It Pays For (It’s Not Claude Code)
- ⚡ The Short: youtube.com/shorts/qMTbSC-miNA
- 💻 credit-calc: github.com/thecoderlog/credit-calc
- 📚 Docs: API credits for Max, Ultra and Team plans, Pricing, Managed Agents, Agent SDK, Batch processing
- 🔁 Related: the prompt caching race (why cached runs cost less) and Claude Agent SDK vs Claude Code: same engine, who wins the race?
If this was useful, subscribe to The Coderlog on YouTube for the next race.