The Anthropic Claude update for May 2026, and what the tighter limits do to a wrapper

The May 2026 update log

Seven entries, in the order they landed. Each one carries a note on what it means for a developer running an agent on a Mac, not just for a press cycle.

1. May 6, 2026

   Claude Code 5-hour limits permanently doubled

   Anthropic 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. The change applies to the actual Claude Code agent loop, including the same loop when it runs inside a wrapper that authenticates as the user's Claude account over OAuth.

   Why it matters here: This is the favorable side of the month. A Fazm window authenticates as your Claude account and runs the real Claude Code agent loop via ACP, so the doubled bucket is the one your usage hits.

   Source: ClaudeDevs (X)

2. May 7, 2026

   Claude Managed Agents gain three new features

   Anthropic added capabilities to Claude Managed Agents, including a memory feature it calls dreaming, which reviews past sessions to find patterns, and multiagent orchestration, where a lead agent breaks a job into pieces and delegates each to a specialist.

   Why it matters here: The orchestration idea maps cleanly onto what a local agent already does with windows. A forked Fazm window is a specialist with the full prior context, running independently.

   Source: 9to5Mac

3. May 12, 2026

   Claude Platform on AWS reaches general availability

   AWS announced that the Claude Platform is generally available, exposing the full Claude API surface, Claude Managed Agents, code execution, web search, web fetch, prompt caching, batch processing, citations, Files API, Skills, and MCP connectors through native AWS billing and IAM. New Claude models and beta tools land on AWS at the same time as the native Claude API.

   Why it matters here: A clean enterprise path that does not change what a personal Mac agent does on your laptop. Useful framing for buyers who already live in AWS, but unrelated to a single developer running Claude Code in a Mac UI.

   Source: AWS

4. May 13, 2026

   Claude for Small Business launches

   Anthropic launched Claude for Small Business, connecting Claude to tools like QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365 with ready-to-run workflows for payroll, invoicing, sales, marketing, and month-end close.

   Why it matters here: It validates the small-business automation thesis at scale. The line a buyer has to draw is API reach versus desktop reach, which the comparison section below walks through.

   Source: SiliconANGLE

5. May 14, 2026

   Separate credit meter introduced for third-party agent harnesses

   Axios reported that Anthropic announced support for outside agent tools on paid Claude plans, but put that usage behind a separate credit meter, with subscribers getting a new monthly credit to use with third-party harnesses. The change does not touch the Claude Code bucket itself, it carves the rest of the agent ecosystem into a distinct allowance. Developer reaction on X was sharply negative, with replies calling the change gaslighting and some users saying they would switch to Codex.

   Why it matters here: This is the change that gets misread as a uniform tightening. It only bites tools that run their own agent loop against the Claude API and ask the user to pay with Claude credits. A tool like Fazm that authenticates as the user's Claude account and runs the actual Claude Code agent loop via ACP stays on the Claude Code meter that was doubled on May 6.

   Source: Axios

6. May 2026

   Microsoft 365 apps reach general availability

   Claude for Excel, PowerPoint, and Word became generally available, and Claude for Outlook entered public beta for paid plans. Alongside that, Anthropic released the Claude Platform on AWS, bringing the full Claude API feature set to AWS billing and authentication.

   Why it matters here: These extend Claude's reach inside specific cloud products. A desktop agent reaches the same apps a different way, through the operating system rather than each product's own surface.

   Source: Anthropic

7. May 2026

   Legal practice plugins and a Gates Foundation grant

   Anthropic introduced new AI plugins for specific legal practice areas, including corporate, regulatory, and employment law, and committed $200 million over four years in grant funding, Claude usage credits, and technical support to a Gates Foundation partnership.

   Why it matters here: Vertical depth and a research grant. Neither touches the agent loop directly, but both signal where Anthropic is putting weight: domain depth over raw model count.

   Source: Anthropic

The anchor fact: line 68 of ChatQueryLifecycle.swift

When a usage-limit error comes back from Anthropic, exactly one place in Fazm decides what the user sees. It is a seven-line branch in Desktop/Sources/FloatingControlBar/ChatQueryLifecycle.swift. Here is the part that runs on every errored turn:

} else if let errorText = provider.errorMessage {
    let isRateLimit = errorText.contains("usage limit")
                    || errorText.contains("rate limit")
    let isPersonalMode = provider.bridgeMode == "personal"

    if isRateLimit && isPersonalMode {
        state.showUpgradeClaudeButton = true
    }
}

That is the whole guard. Two substring tests against the error text, one equality check on bridge mode, one flag the floating bar reads to render an upgrade banner instead of a stack trace. The matching is on substrings, not exact strings, on purpose: Anthropic has reworded rate-limit and usage-limit errors more than once, and an exact-match check would break on the first rewording. After the May 6 doubling, this branch fires later in long sessions than it did in April, but the wording it matches did not have to change for the guard to keep working.

There is a second path in the same file, around lines 233 to 247, that closes the loop. It subscribes to the rate-limit status and the reset timestamp Anthropic returns, and once the reset time has passed and the status is back to allowed, it clears showUpgradeClaudeButton on its own:

.sink { [weak state, weak provider] status, resetsAt in
    guard status == "allowed" || status == nil else { return }
    guard state.showUpgradeClaudeButton else { return }
    let now = Date().timeIntervalSince1970
    let resetElapsed = (resetsAt ?? 0) <= now
    // ... clears the banner once the window resets
    state.showUpgradeClaudeButton = false
}

The practical result of the two paths together: regardless of how the May 2026 limit changes shake out for your bucket, when you do hit a cap in a long Fazm window you get a calm upgrade banner, not a crash, and when your limit window resets the banner disappears without you touching it. The limit moved. The handling did not have to.

When the next Claude model ships, it appears without an app update

May 2026 was light on model launches, but the next Claude model is a matter of time. The thing worth knowing is that a wrapper does not need to ship a release to expose it. Fazm builds its model list from whatever the agent reports, on every session. The flow looks like this:

The function in the middle is emitModelsIfChanged, at line 2368 of acp-bridge/src/index.ts. It receives the agent's availableModels on every new session, deduplicates against the last list it sent, and forwards a models_available message to the Swift app. The picker in ShortcutSettings.swift then maps each ID against a substring table: a model whose ID contains haiku fills the Scary slot, sonnet fills Fast, and opus or default fills Smart. There is even a migration note in that file: ACP SDK v0.29 and later use the literal default for the latest Opus, and the code rewrites a stored opus choice to match. So whenever Anthropic publishes the next Sonnet or Opus, it slots into the picker on the next warmup. No rebuild, no download.

Claude for Small Business vs a local Mac agent

The May launch of Claude for Small Business is the update most likely to make a small-business owner ask which tool they actually need. The honest answer is that they solve the same goal through two different doors, and which door is right depends on where your work lives.

Claude for Small Business reaches your work through cloud integrations. It connects to QuickBooks, HubSpot, Canva, Docusign, Microsoft 365 and similar products through their APIs and runs workflows inside that connected set. When your month-end close, your invoicing, and your CRM all already live in supported SaaS tools, that is a clean fit and the work happens in the cloud.

A local Mac agent reaches your work through the operating system. Fazm drives your real browser through an extension and native Mac apps through macOS accessibility APIs rather than screenshots, so the same agent can act inside an app that has no API at all: a legacy desktop tool, an in-house program, a vendor portal that was never built to be integrated. It runs locally, and you bring your own Claude account so usage hits your existing plan.

Where Claude for Small Business wins: your stack is modern SaaS, fully supported, and you want Anthropic to host and maintain the connectors. Where a local agent wins: your work spans apps with no integration, you want it to run on your machine, or you want the same agent that writes code to also click around the rest of your desktop. Many people will end up using both, one for the connected cloud tools and one for everything the cloud cannot reach.

Questions about the May 2026 Claude updates