Anthropic split third-party agent tools onto a separate credit meter on May 14, 2026. Where a Claude Code wrapper actually lands.
Two Anthropic policy changes hit the Claude ecosystem inside eight days. On May 6 the 5-hour Claude Code limits on Pro, Max, Team, and seat-based Enterprise plans were permanently doubled. On May 14 Axios reported Anthropic put outside agent tools on a separate monthly credit allowance instead of letting them spend off the main plan. The two pull opposite directions for users of Cursor, Windsurf, Zed, and OAuth-based Claude Code wrappers. This page is the bucket map.
What changed, what stayed the same
Two policy moves inside eight days, pulling opposite directions for different buckets of tools. The list below is the difference, not a summary of either announcement.
Anthropic changes, May 6 to May 14, 2026
- May 6, 2026: Claude Code 5-hour limits doubled on Pro, Max, Team, seat-based Enterprise. Peak-hour throttle removed on Pro and Max.
- May 14, 2026: outside agent tools moved to a separate monthly credit allowance, distinct from the main Claude subscription pool.
- Unchanged: claude.ai and the official Claude Desktop app fund off the main plan as before.
- Unchanged: direct Anthropic API key billing is per-token, separate invoices, separate ceiling.
- Unchanged: the eight-field rate_limit_event SDK payload. Same fields, different next-call meter.
- Unchanged: Pro and Max subscription prices. The May 14 move is a meter split, not a sticker-price change.
The wedge, one sentence
A Claude Code wrapper that signs in as your Claude account over OAuth runs the same agent loop the bare CLI does, so it sits in the bucket Anthropic doubled on May 6, not the bucket Anthropic separated on May 14.
The May 14 third-party harness credit meter is for outside agent tools that fund themselves off a user's Claude subscription as their own caller. An OAuth wrapper does not run its own caller. It relays the Claude Code agent loop, the loop is the real one, and that loop bills against the user's Pro or Max plan exactly the way the bare Claude Code CLI does.
So on the same day a Cursor or Windsurf user reading the Axios story was figuring out whether their outside-harness budget had just been carved off from their main plan, a Fazm user was seeing a 5-hour cap that filled at twice the previous rate, with no change in their meter at all.
That is the wedge. It is unrelated to clever code on the wrapper side. It is structural: OAuth as your Claude account puts the wrapper on the same meter as the CLI, and Anthropic doubled that meter.
Before the split and after
One bucket became two. The user-visible side is which pool funds the next request once the warning fires. The cap math is different on each side of May 14, and different again from May 6.
<div class="space-y-4"> <p><strong>Before May 14, 2026:</strong> Anthropic's main plan budget covered everything that authenticated against your Claude account, regardless of whether the caller was claude.ai, the Claude Desktop app, the Claude Code CLI, an OAuth-based wrapper, or an outside agent harness asking to fund itself off your subscription. One bucket, one meter, one set of plan limits.</p> <p>Practical consequence: a heavy outside-tool day could eat the same budget your interactive claude.ai chat was supposed to draw from. Engineers running both noticed the cap fill earlier than their typing suggested.</p> </div>
The wire payload did not change. Only the next-call meter did.
Anthropic's Agent SDK still emits a rate_limit_event with eight fields regardless of bucket. What changed in May is which pool funds the next request once one of those fields flips. Fazm forwards every field through ACP and renders the banner from the live values, not a hard-coded string.
One request, four actors, OAuth-as-your-account end to end
The wrapper does not interpose between you and Anthropic as a separate billed caller. It relays the agent loop. The OAuth identity on the wire is yours, the plan limits are yours, and after May 6 the bucket is twice as deep.
Fazm request path - OAuth as your Claude account
The 30-second test for which bucket your tool is in
Open the tool. Look at the auth panel. The answer is one of three shapes, each of which puts the tool in a different bucket. The buckets do not overlap.
The Twitter wave that rode this news cycle
Two posts in 48 hours pushed roughly 50 thousand combined views toward fazm.ai. Both linked to the homepage with no dedicated landing for the news context. This page is that landing.
@svpino
May 17, 2026
views
28,133
Mention thread about Fazm. The cluster ran the same week as the Anthropic May 14 announcement on third-party agent credits. The interpretation question (which bucket the OAuth wrapper lands in) is what readers arriving from the post needed answered.
https://x.com/svpino/status/2055759187825893461 →
@eastweb3eth
May 16, 2026
views
23,092
Companion thread the day before, with 175 reposts. Same news cycle. Same downstream question for anyone running an OAuth-based Claude Code wrapper.
https://x.com/eastweb3eth/status/2055482080134586787 →
Cluster context
May 14-17, 2026
views
50,000+
The two threads sit on top of Axios's May 14 piece on Anthropic moving outside agent tools to a separate credit meter. They drove a wave of inbound traffic to fazm.ai's homepage. This page is the dedicated landing they should have hit instead.
https://www.axios.com/ →
The post bodies are short mentions, not in-depth analysis. The discussion they kicked off is the part that matters: which tools fall in the doubled bucket, which fall in the separated bucket, and which sit in neither. The bucket map above is the answer this page exists to give.
What every May 14 roundup covered
The first-page coverage for the May 14 announcement is a short list and it points at the policy. Nothing wrong with it. What it leaves out is the bucket question for any given tool, which is the only question a user actually has.
What those roundups do not answer
The downstream gap
- Most coverage treats May 6 and May 14 as separate stories instead of two halves of one policy split.
- Few sources state which bucket each named tool (Cursor, Windsurf, Zed, OAuth wrappers) actually lands in.
- The auth-flow test (OAuth vs API key vs own subscription) is rarely written down anywhere a user can find it.
- The rate_limit_event protocol shape is documented, but not the consequence of which bucket the next request lands on after a warning.
- Direct API customers occasionally get told the May 14 meter applies to them. It does not.
The eight-day window, in order
Five moments between Anthropic's May 6 limit doubling and the May 17 viral wave on Twitter. The viral wave is the symptom, the bucket split is the cause, and the bucket map is what every reader arriving from a tweet needed to see.
May 6, 2026 - Claude Code 5-hour limits doubled
Anthropic permanently doubled the 5-hour rate limits on Pro, Max, Team, and seat-based Enterprise plans, and removed the peak-hour throttle on Pro and Max. ClaudeDevs announced it on X. The change applies to the real Claude Code agent loop, including the same loop when it runs inside an OAuth-based wrapper.
May 12, 2026 - Claude Platform on AWS reaches GA
Anthropic shipped general availability of the Claude Platform on AWS with native IAM and AWS billing. Direct API tracks, not relevant to the OAuth-Pro-or-Max flow most desktop wrappers use.
May 13, 2026 - Claude for Small Business launches
Anthropic dropped Claude into QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365 with ready-to-run workflows. Cloud-API integrations against SaaS surfaces. Different category from a desktop agent that drives the Mac through accessibility APIs.
May 14, 2026 - Third-party agent tools split onto a separate credit meter
Axios reported Anthropic put outside agent tools on a separate monthly credit allowance instead of letting them spend off the main plan. Applies to tools that fund themselves off a user's Claude subscription. Does not apply to OAuth wrappers that relay the Claude Code agent loop, and does not apply to tools with their own Anthropic API key billing.
May 16-17, 2026 - Viral Twitter wave
A Santiago Valdarrama post about Fazm reached 28 thousand views. A related thread by eastweb3eth posted the day before reached 23 thousand. Both ran in the same news cycle as the May 14 announcement. The interpretation question on the wave was identical: which tools sit in which bucket.
Three auth flows, three buckets, one fork at setup
The fork is at the moment you set up the tool. Each input lands on exactly one bucket and stays there for the life of that auth session.
Auth flow -> Anthropic billing bucket
The numbers from the announcements
Four hard facts, taken verbatim from Anthropic's May 6 and May 14 communications. Counts animate as they scroll into view. Nothing on this row is an estimate.
For scale, the eight days between the two policies (May 0 to May 0) are the shortest gap between two major Anthropic billing changes in 2026.
OAuth Claude Code wrapper vs everything else
The left column is a wrapper that signs in as your Claude account. The right column is every other category: direct API key, tool-own-subscription, or outside harness on your subscription. Six rows, plain comparison.
| Feature | Direct API / own pool / outside harness | OAuth Claude Code wrapper (Fazm) |
|---|---|---|
| Auth flow at setup | Paste sk-ant API key (direct API). Or pay a separate subscription (own pool). | OAuth to claude.ai. You type your Claude Pro or Max password. No API key entry. |
| Which Anthropic meter funds calls | Either: direct API per-token billing, the tool's own subscription budget, or the new May 14 third-party harness credit meter. | Your Claude Code 5-hour meter on your Pro or Max plan. Doubled May 6, 2026. |
| Effect of May 14 third-party meter split | Applies if the tool funds itself off your Claude subscription. Does not apply to direct API key flows or the tool's own subscription. | None. OAuth wrappers fund off your subscription's main Claude Code bucket, not the new third-party meter. |
| Effect of May 6 limit doubling | Does not apply to direct API key billing or to tools using their own pooled credits. | Applies. The same 5-hour bucket the bare Claude Code CLI uses is doubled, and a wrapper runs the same bucket. |
| Where the rate_limit_event arrives | Direct API tools see the event but it is per-key per-request, not plan-scoped. Subscription tools generate their own internal usage UI off their own counters. | Forwarded through ACP with all eight fields (status, resetsAt, rateLimitType, utilization, overageStatus, overageDisabledReason, isUsingOverage, surpassedThreshold). Banner copy generated from the live values. |
| What "third-party app" means in Anthropic banner copy | Direct API users do not see Claude subscription banner copy. Tool-own-subscription users see only the tool's banner, not Anthropic's. | A tool authenticated against your Claude account that is not claude.ai or the official Claude Desktop app. OAuth wrappers like Fazm qualify when the plan cap fills and overage starts billing. |
Things the May 14 story did not actually say
A short list of paraphrases that picked up traction in replies and are not what either announcement contained. None of these are true.
What to do this week
Five plain steps. The first identifies your bucket. The next four cover what to do once you know which one.
Identify which bucket your tool is in
Open the tool's auth panel. OAuth to claude.ai means Claude Code bucket. API key paste means direct API. In-app subscription means tool's own pool. The bucket determines whether May 6 or May 14 changed your day.
If you are on the OAuth Claude Code bucket
Your 5-hour limit doubled on May 6. The May 14 third-party meter split does not apply. Long agent runs that used to hit the plan cap now run twice as long before the cap fires.
If your tool funds itself off your Claude subscription as a harness
You are in the May 14 bucket. Check the third-party credit balance separately at claude.ai/settings/usage. Heavy outside-harness days now hit a known monthly ceiling instead of competing with your interactive chat.
If you are on direct API key billing
Nothing changed. You bill per token against your Anthropic API account. The plan-side changes do not touch this track.
If your tool uses its own subscription budget
Nothing changed at Anthropic's layer that you can see. The tool's internal budget UI is what you watch. The May 14 split is between Anthropic and the user, not between you and the tool.
Plan-by-plan impact
Claude Pro
Claude Code 5-hour limit permanently doubled May 6, peak-hour throttle removed. Third-party agent tools that fund off the Pro subscription got their own monthly allowance on May 14. OAuth wrappers (Fazm) sit on the doubled bucket.
Claude Max
Same as Pro: 5-hour limit doubled May 6, peak-hour throttle removed, third-party harness meter separated May 14. Long agent runs on Max under an OAuth wrapper now go materially further before the cap fires.
Claude Team / seat-based Enterprise
5-hour Claude Code limit doubling applies. Peak-hour throttle removal is Pro and Max only. Enterprise-level bring-your-own-key paths sit on direct API billing, untouched by the May 14 meter split.
Direct Anthropic API
Pay per token against your API account. Neither the May 6 doubling nor the May 14 meter split applies. Your rate limits are per-key, your billing is per-invoice, your numbers do not move because of either policy change.
Honest caveats
A short list of edges where the simple bucket map breaks down. None of these change the core conclusion, but they are worth knowing if your setup is unusual.
- ▸The Axios reporting on May 14, 2026 is the public source on the third-party agent credit meter. Anthropic's own published copy describes the same allowance but in less direct language.
- ▸Tool vendor billing can change. The OAuth-vs-API-key test applies as of this writing. If a tool changes its auth flow, the bucket changes with it.
- ▸The May 14 split is a meter separation, not a price change. Subscribers were not charged a different amount on May 14 itself.
- ▸The May 6 doubling is for the 5-hour Claude Code limit. The 7-day weekly caps did not double on the same day.
- ▸If you run multiple tools on the same Claude account (one OAuth, one outside harness), each lands in its own bucket. The buckets do not interact.
The synthesised takeaway
Anthropic gave the Claude Code bucket twice the headroom on May 6 and put outside agent harnesses in their own lane on May 14. An OAuth wrapper that runs the real Claude Code loop benefits from the first move and is untouched by the second.
That is the entire story for a Mac wrapper. The eight-field rate_limit_event still ships, the banner still renders, the session keeps going. What moved is the budget on the bucket the wrapper already lived on, in the right direction. There is no migration to do, no settings page to visit, no new credit card to add. Sign in as your Claude account, the same way you did in April, and a long session lasts twice as far before the cap fires.
See the doubled bucket on a real Mac
Fazm runs the Claude Code agent loop via ACP using your Claude Pro or Max account. Persistent sessions, one-click fork, no auto-compacting. Free to start, fully open source, runs locally.
macOS 14.0+. Bring your Claude Pro or Max account. Voice-first, fully local.
Claude third-party tool credit meter, May 2026 - FAQ
What changed on May 14, 2026 with Anthropic and third-party tools?
Per Axios on May 14, 2026, Anthropic put outside agent tools on a separate credit meter. Subscribers on Pro, Max, Team, and seat-based Enterprise plans now get a monthly third-party-harness credit allowance instead of letting those calls draw from the main plan. The change affects tools that run as their own agent loop against the Anthropic API and ask the user to pay for their own credits. It does not change how the Claude Code agent loop bills against your Pro or Max plan when a tool signs in as your Claude account over OAuth and runs that loop on your behalf.
What changed on May 6, 2026 with Claude Code limits?
Anthropic permanently doubled Claude Code's 5-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans, and removed the peak-hour limit reduction on Pro and Max. ClaudeDevs announced the change on X on May 6. The doubling applies to the Claude Code agent loop itself, including the same loop when it runs inside a wrapper that authenticates as the user's Claude account.
Are those two changes contradicting each other?
No, they apply to two different buckets. The May 6 doubling is for the Claude Code agent loop on your Pro or Max plan. The May 14 separate credit meter is for third-party agent tools that operate against the Anthropic API as their own caller. A given product can be in exactly one of the two buckets at a time. The simple test: if the tool asked you to log into claude.ai during setup, it is almost certainly in the Claude Code bucket and got the May 6 doubling. If the tool asked you for an Anthropic API key or set up a billing relationship of its own, it is in the May 14 bucket.
Where does a Claude Code wrapper like Fazm sit?
In the Claude Code bucket. Fazm wraps the Claude Code agent loop via the Agent Client Protocol using your Claude Pro or Max account over OAuth. Calls hit your Claude Code 5-hour meter, which is the bucket Anthropic doubled on May 6, 2026. The May 14 third-party harness credit meter does not apply to Fazm because Fazm does not have an independent billing relationship with Anthropic and does not run its own API caller. The wrapper relays the agent loop, the loop is the real Claude Code, and it bills the user the same way the bare Claude Code CLI does.
Where does Cursor sit?
Cursor offers two billing tracks. With a Cursor subscription, you use Cursor's own pooled inference budget, which has been the case for a long time and is unrelated to the Anthropic plan changes. With Cursor's Bring Your Own Key configuration against an Anthropic API key, you are billed directly by Anthropic per token and the rate_limit events behave like raw API usage. Cursor's default flow is its own subscription, so most users sit outside both the May 6 doubling and the May 14 separate meter.
Where does Windsurf sit?
Windsurf operates against its own subscription on Cascade and on top of model providers via integrations. Windsurf users on the Cascade plan use Windsurf's own budget. Users routing calls through their own Anthropic API key are on the direct API billing track. Neither flow is affected by the May 6 Claude Code doubling, and the May 14 third-party agent credit meter applies to tools that ask for outside-harness credit on the user's Claude subscription, which is a narrower category than every product that touches the API.
Where does Zed sit?
Zed's AI features have a similar Bring Your Own Key flow. If a Zed user routes through their own Anthropic API key, billing is direct per token. If Zed offers an OAuth-based flow against the user's Claude subscription for any of its agent surfaces, that surface lands on the same OAuth-Claude-Code-account meter as Fazm does. The general rule holds: OAuth as your Claude account means the May 6 doubled meter. API key with a direct billing relationship means raw API billing, and the May 14 credit meter is irrelevant.
How do I tell which bucket my tool is actually in?
Open the tool's settings or auth panel. If it logged you in by redirecting to claude.ai and you typed your Claude subscription password, it is an OAuth flow against your Claude account, and you are on the Claude Code bucket Anthropic doubled on May 6. If it asked you to paste an sk-ant API key, you are on direct API billing, and your usage is per-token against your API account. If neither, the tool is using its own pooled credits and the Anthropic policy changes do not touch you directly. The Axios May 14 announcement is specifically about the third category being separated when a tool uses your Claude subscription to fund its own harness loop.
Is the May 14 change a price increase?
It is a meter split, not a sticker-price change. The Pro and Max plan prices did not move on May 14, 2026. What moved is the allocation: outside tool usage that previously drew from the main plan now draws from a separate monthly credit allowance. If your day was spent inside the Claude Code agent loop on your own subscription, you got more budget on May 6 (doubled limits) and the May 14 split does not touch you. If your day was spent inside an outside agent harness that funded itself off your Claude subscription, you now have a separate budget for that, distinct from your main plan budget.
Does this affect users on the Anthropic API directly?
No. Direct API billing on an Anthropic API key is a third, separate track. Pay per token, separate invoices, no plan limits and no extra usage. The May 14 third-party credit meter applies only to tools that fund themselves off a user's Claude subscription, not to tools that have their own billing relationship with Anthropic.
Do third-party tools still see rate_limit events from Anthropic?
Yes. The Anthropic Agent SDK still emits the rate_limit_event with the same eight fields (status, resetsAt, rateLimitType, utilization, overageStatus, overageDisabledReason, isUsingOverage, surpassedThreshold) regardless of which bucket the call is being billed against. The difference is which pool the next request lands on once the warning fires. Fazm forwards all eight fields through the ACP bridge so the UI can render the exact state, not a paraphrased banner.
Where is this discussed in Fazm's source today?
Two files. acp-bridge/src/index.ts at lines 4457 to 4480 destructures the eight rate_limit fields and forwards them with sendWithSession. Desktop/Sources/Providers/ChatProvider.swift at lines 865 to 898 maps the machine string "overage" to the human label "extra usage limit" and updates the published state the chat banner reads. The behaviour did not change with the May 6 doubling or the May 14 meter split, because Fazm sits on the OAuth Claude Code bucket and the events still arrive the same way.
I saw a viral tweet about Fazm around May 17, 2026. What was that about?
A Santiago Valdarrama post about Fazm picked up around 28 thousand views by May 17, 2026. A related thread by eastweb3eth ran the day before. Both rode the wave around Anthropic's May 14, 2026 third-party agent tool credit meter announcement. The story Fazm has from that wave is the wedge on this page: a wrapper that authenticates as your Claude account and runs the real Claude Code agent loop sits in the doubled May 6 meter, not in the separate May 14 third-party harness meter.
Related Fazm guides
Keep reading
"Third-party apps now draw from your extra usage, not your plan limits." What that Claude banner means
The eight-field rate_limit event behind the banner most third-party clients drop on the floor.
Anthropic Claude update, May 2026: every change, and what tighter limits do to a Mac wrapper
Every Anthropic announcement in May 2026 in one place, with each change mapped to a specific Fazm code path.
Anthropic news April 15, 2026: the Claude outage, and the 100-line error pipeline that kept Fazm's chat window honest
The April 15 outage from a downstream app's view, traced through acp-bridge's typed error pipeline.
Want to see the OAuth bucket math on your team's actual usage?
I will walk through the May 6 doubling and the May 14 meter split against your current setup on a 20-minute call.
Comments (••)
Leave a comment to see what others are saying.Public and anonymous. No signup.