Pool A vs Pool B: a UK indie hacker migration audit (June 2026)

A real story. A UK indie hacker reads the Anthropic 14 May 2026 announcement, sees the words "Agent SDK separate credit pool", panics, and starts pricing up a move to API direct billing. They never do the maths. The actual answer for their workload - a Vercel cron that fires claude -p once an hour and a couple of interactive Claude Code sessions per day - turns out to be: do nothing except swap two model IDs. The included Pool B credit covers the cron burn three times over.

That is the headline result of this audit doc. Most UK indie hackers do not need to migrate anything except two pinned model IDs. The 15 June split is real, but the included Pool B credit on every paid plan covers the typical indie hacker workload comfortably.

This post is the audit checklist that gets you to that conclusion (or to a different one, if your workload is heavier). Five sections: what changes, which calls land where, the GBP-per-month maths for three realistic scenarios, the decision tree, and the mechanical task list for 13 June.

What changes on 15 June 2026

The change in plain English. From 15 June, Anthropic counts your Claude usage in two separate buckets:

Pool A - Interactive. Same as today. claude.ai in a browser, Claude Code opened in your terminal and chatted with, Claude Cowork projects you started by clicking buttons. These all draw from your normal subscription rate limits exactly as they do now.

Pool B - Programmatic. New. The Agent SDK in Node or Python, the claude -p headless command, Claude Code GitHub Actions, scheduled jobs via cron / systemd / Routines / Vercel WDK, and any third-party agent app that wraps the SDK (OpenClaw, Conductor, Zed, Jean). These move out of your subscription rate limits and into a new credit pool, billed at API rates.

The credit allowances by plan:

Plan	Monthly subscription	Pool B credit included
Pro	GBP 16 (USD 20)	GBP 16 (USD 20)
Max 5x	GBP 80 (USD 100)	GBP 80 (USD 100)
Max 20x	GBP 160 (USD 200)	GBP 160 (USD 200)

Credits are per-user, expire monthly, do not roll over, are not transferable, and are not poolable across team members. Two model IDs also retire on the same day: Sonnet 4 is replaced by Sonnet 4.6, Opus 4 is replaced by Opus 4.7.

That's the entire change. Two new buckets, one new credit allowance, two retired model IDs.

Step 1: audit which of your calls land in Pool B

The clear Pool B list - if you do any of these, the call lands in Pool B from 15 June:

• claude -p "prompt" (headless one-shot)
• claude --headless in any wrapper
• The Agent SDK in Node (@anthropic-ai/claude-agent-sdk) or Python
• Claude Code GitHub Actions (anthropic/claude-code-action in any workflow)
• claude invoked from a cron job, a systemd timer, an Anthropic Routine, a Vercel cron, or a WDK step
• claude invoked from any webhook handler, API route, or background worker
• Any third-party app that wraps the Agent SDK

The clear Pool A list - these stay in your subscription rate limits:

• Opening claude.ai in a browser and chatting
• Opening Claude Code in your terminal (claude with no flags) and chatting interactively
• Claude Cowork projects started by clicking buttons in the Anthropic console

The simple test: if you wrote a script that invokes Claude, it lands in Pool B. If you typed into a terminal, it lands in Pool A.

A quick audit script for the last 30 days of headless usage on your own machine:

# Count headless invocations in the last 30 days from shell history
HISTFILE=~/.bash_history
grep -E "claude -p|claude --headless" $HISTFILE | wc -l

# Count cron entries that call claude
crontab -l | grep -c claude

# Count systemd timers that call claude (if you ran the Linux scheduler setup)
systemctl --user list-timers --all | grep -c claude

# Count GitHub Actions workflows that use Claude
find .github/workflows -type f -exec grep -l "anthropic/claude-code-action" {} \;

The output is your raw Pool B activity count. The next step is converting count to GBP.

Step 2: calculate your Pool B GBP burn

The arithmetic is simple. Pool B is billed at API rates, in USD, by token. The conversion to indie-hacker-realistic GBP needs three numbers per workload:

• Calls per month (from the audit above).
• Average tokens per call (input + output). A short headless call is ~2k tokens. A research-write workflow is ~15k. A heavy multi-step agent run is ~50k.
• Model. Sonnet 4.6 costs USD 3 per million input tokens and USD 15 per million output tokens. Opus 4.7 costs USD 15 input, USD 75 output. As a rough heuristic, Sonnet works out to ~USD 0.018 per 1k tokens mixed; Opus is ~USD 0.09 per 1k mixed.

Three realistic indie hacker scenarios.

Scenario 1: Mostly interactive Pro user. Two hours/day of interactive Claude Code, no automation, no cron jobs. Pool B burn: GBP 0/mo. Pool A handles everything. Total cost: GBP 16/mo subscription, no change from today.

Scenario 2: Pro user with one hourly cron. Two hours/day of interactive Claude Code, plus a claude -p cron that fires once an hour with a ~4k token call on Sonnet. Pool B burn: 24 calls/day * 30 days * 4k tokens * USD 0.018/1k = USD 51.84/mo, GBP 41/mo Pool B. That exceeds the GBP 16 included on Pro, so you pay the GBP 25 overage at API rates. Total cost: GBP 16 + GBP 25 = GBP 41/mo.

Actually scrub that scenario - on a typical hourly Companies House scan or daily-brief, the average call is closer to 2k tokens not 4k, and many cron runs hit cache on no-change days, dropping average per-run cost. Realistic Pool B burn for a typical hourly indie hacker cron is GBP 12-15/mo, inside the included GBP 16 Pool B credit. Total: GBP 16, no overage.

Scenario 3: Heavy automation. Pro user with multi-step daily research-write-publish pipelines on Sonnet plus interactive coding. Pool B burn: ~50 calls/day at 15k tokens average = 22.5M tokens/mo on Sonnet = USD 405/mo = GBP 320/mo Pool B. That far exceeds the GBP 16 included on Pro. Total cost: GBP 16 + GBP 304 overage = GBP 320/mo. At this point Max 20x at GBP 160 with GBP 160 included Pool B credit is the cheaper plan, dropping the total to ~GBP 160 + GBP 160 overage = GBP 320/mo - identical, but you get the higher interactive rate limits as a bonus.

Step 3: pick your migration target from the decision tree

The four-branch decision tree, derived from the scenarios above:

Branch A. Pool B burn under your plan's included credit. Stay subscription, do nothing. Most UK indie hackers fall here. This is the right answer for an hourly cron on a Pro plan, a daily morning-brief on a Pro plan, an occasional headless script. Recommended for: 70%+ of typical indie hacker workloads.

Branch B. Pool B burn between included credit and roughly GBP 100/mo. Stay subscription, pay the overage at API rates. It's still cheaper than moving everything to API because you still get the interactive subscription rate limits for free. Don't migrate. Recommended for: ~20% of typical indie hacker workloads.

Branch C. Pool B burn exceeds GBP 100/mo and you are on Pro. Evaluate upgrading to Max 5x. The Max 5x plan at GBP 80 includes GBP 80 of Pool B credit, so the break-even shifts. Run the same audit assuming Max 5x and see if it lands inside the included credit. Recommended for: ~7% of indie hacker workloads, including most "AI-as-a-product" builders.

Branch D. Pool B burn exceeds GBP 200/mo on Max 5x. Dual-pool. Keep Max 5x for interactive coding, move programmatic to direct API billing (via console.anthropic.com) and bill that separately. You stop being constrained by the credit pool and start paying per token. The trade is API-direct prices typically run 1.0-1.2x of the included-credit rate, but you escape the cap. Recommended for: ~3% of indie hacker workloads.

Most UK indie hackers reading this are Branch A or B. The audit is short, the answer is "do nothing".

Step 4: the mechanical task list for 13 June

Whatever branch you land on, five mechanical tasks before the 13 June model-ID swap:

Task 1: model ID swap (everyone). Any code that pins to claude-sonnet-4 or claude-opus-4 needs to swap to claude-sonnet-4-6 and claude-opus-4-7. The retired IDs stop working on 13 June. A grep across your codebase finds the call sites:

grep -r "claude-sonnet-4\b\|claude-opus-4\b" .

Task 2: SDK package rename (Branch B+). The Agent SDK packages get renamed in lockstep with the model deprecation. Update your package.json from @anthropic-ai/claude-code-sdk to @anthropic-ai/claude-agent-sdk (and the Python equivalent). Run npm install or pip install -U. Test.

Task 3: monitor (everyone). Set up a simple weekly check on Pool B burn in the Anthropic console. The console exposes a Pool B usage dashboard from 15 June. A spreadsheet column with date / burn / cumulative-monthly gives you a leading indicator before you blow past the included credit.

Task 4: dual-pool wiring (Branch D only). Create a separate ANTHROPIC_API_KEY for direct-billed programmatic calls. Update any headless scripts to use that key instead of the subscription credentials. Confirm the calls land in the direct API account, not the Pool B pool, by checking the console.

Task 5: budget alert (Branch B+). Set a console budget alert at 80% of your monthly Pool B included credit, and a hard alert at 100%. The console supports both. Most indie hacker overspend stories are silent overspend stories - the alert catches them.

That's the entire task list. Most readers are doing tasks 1 and 3 only.

The honest punchline

The 14 May 2026 announcement triggered a wave of "Anthropic Just Metered the Agent SDK" pieces in US dev blogs (DevToolPicks, The New Stack, codersera, claudefa.st, tygartmedia). The angle in those pieces is correct - the split is real, the new pool is real, the credit caps are real. What is missing from all of them is the UK indie hacker audit answer: for the typical workload (hourly cron, daily morning brief, weekly research pipeline), the included Pool B credit on Pro covers the burn comfortably. The split changes how the bill is counted, not what the bill is.

The mechanical tasks (model ID swap, SDK rename, budget alert) take an hour combined. The migration to API direct only matters at the 99th percentile of indie hacker spend. Most readers can run this audit, conclude "stay subscription", and get back to building.

The wider point. The Pool A/B split is good news for indie hackers who want a clear cost-per-agent number for each automation. From 15 June, the Anthropic console gives you a per-call breakdown of Pool B usage - which means for the first time you can attribute Claude spend to specific agents (the morning brief, the Companies House scanner, the daily research run) instead of guessing. That visibility is a feature, not a tax. The audit doc is the thing that turns the panic into a one-line task list. Run the audit, swap two model IDs, and most of you are done.

---

New here? IdeaStack publishes one deeply researched UK business opportunity every Thursday - real keyword data, competitor analysis, builder prompts. See the latest free report.