Opus 4.7, the new weekly Opus cap, and the ten-line UserDefaults migration Fazm runs on every launch to keep your assistant answering

Anthropic landed four things that reshaped how consumer apps schedule Claude calls. Opus 4.7 went GA on April 14 at the same $5 input / $25 output per MTok as Opus 4.6. The weekly rate-limit surface got finer: alongside the existing five_hour and seven_day windows, a dedicated seven_day_opus cap now tracks Opus usage separately from Sonnet usage. The Claude Code desktop app shipped a redesign and the Routines research preview on April 14. And the Terms of Service updated early in the month, requiring users to re-accept on claude.ai before new API calls would succeed. Each of those lands as a specific string or numeric label inside a shipping Mac app.

seven_day_opus is the Anthropic rate-limit type label for the weekly Opus usage window. It is returned on agent errors when a user exhausts their Opus weekly quota, distinct from seven_day (the overall weekly limit) and five_hour (the session limit). Fazm's ChatProvider.swift lines 538 to 546 implement a rateLimitTypeLabel function that maps four type strings to human-readable labels: five_hour becomes 'session limit', seven_day becomes 'weekly limit', seven_day_opus becomes 'Opus weekly limit', seven_day_sonnet becomes 'Sonnet weekly limit', and overage becomes 'extra usage limit'. The four labels are the operational vocabulary the UI speaks when Claude says no.

Because Opus at the post-April-2026 weekly cap is not viable as the default for an always-on desktop assistant. A Fazm user types into the floating bar throughout the day: short questions, follow-ups, one-shot tasks. Opus 4.7 is strong on long multi-step engineering work, but the weekly Opus budget is sized for occasional deep work, not for being the hammer that hits every nail. Sonnet 4.6 is faster, cheaper, has a separate rate-limit bucket (seven_day_sonnet), and is the model Anthropic's own Claude Code pipeline defaults to for most operations. The one-time migration is Fazm accepting that reality on behalf of its existing Opus users.

Lines 287 to 296 of /Users/matthewdi/fazm/Desktop/Sources/FloatingControlBar/ShortcutSettings.swift, inside the ShortcutSettings singleton initializer. Ten lines total. It reads the saved model from UserDefaults under the key shortcut_selectedModel, reads a guard boolean under the key shortcut_didMigrateFromOpus, and runs the migration only if the saved model equals the string 'claude-opus-4-6' AND the guard has never fired before. On migration, it writes 'claude-sonnet-4-6' to shortcut_selectedModel, flips shortcut_didMigrateFromOpus to true, and sets the in-memory selectedModel property to match. The migration is idempotent by construction: once the guard is true, the branch is dead.

Two strings. shortcut_selectedModel stores the Claude model identifier the user's floating bar will call (default claude-sonnet-4-6, previously claude-opus-4-6 for users who manually picked Smart). shortcut_didMigrateFromOpus is a boolean that starts at false, flips to true the first time the migration runs, and stays true forever after. Both live in the app's standard UserDefaults domain. A reader with access to the app sandbox can inspect them with: defaults read ai.fazm.desktop shortcut_didMigrateFromOpus. The double-key design is how Fazm keeps the migration one-time: a user who later manually re-picks Opus does not get migrated a second time.

They are the short labels for the three Claude model tiers the floating bar surfaces. Line 151 to 155 of ShortcutSettings.swift: claude-haiku-4-5-20251001 is 'Scary (Haiku)' with short label 'Scary', claude-sonnet-4-6 is 'Fast (Sonnet)' with short label 'Fast', and claude-opus-4-6 is 'Smart (Opus)' with short label 'Smart'. 'Scary' refers to Haiku's speed, not its safety: responses come back so fast it is unsettling. The short labels are what render in the tight floating bar chrome when there is no room for full model names. After the April migration, new users default to 'Fast' and the 'Smart' option is still available but no longer auto-selected.

The ACP (Agent Client Protocol) subprocess emits a models_available notification after a session opens. ChatProvider.swift line 1018 forwards that list to ShortcutSettings.shared.updateModels(models). The updater at lines 168 to 189 of ShortcutSettings.swift maps each incoming model ID through a substring table (if the ID contains 'haiku' it gets the 'Scary' label; if it contains 'sonnet' it gets 'Fast'; if it contains 'opus' it gets 'Smart'), and unknown families fall through to the raw Claude name with 'Claude ' stripped. This is how an Opus 4.7 entry shows up as 'Smart' without any code change: the substring match catches it. The list only refreshes if it actually changed.

No. The migration is about which Claude model ID the floating bar asks for. The bridge mode (personal OAuth versus bundled Anthropic API key) is independent and lives in ChatProvider.swift under @AppStorage('bridgeMode'). A user whose shortcut_selectedModel gets migrated from claude-opus-4-6 to claude-sonnet-4-6 keeps whatever bridge mode they already had. A separate decision path handles the case where the selected model is not reachable through the current bridge. The two systems were designed to be orthogonal on purpose.

Four strings the app maps to display text. five_hour is the session limit, which resets every five hours and applies to both personal OAuth and bundled-key sessions. seven_day is the global weekly limit on the user's Claude account. seven_day_opus is the Opus-specific weekly cap that Anthropic enforces separately in April 2026 to manage Opus 4.7 capacity. seven_day_sonnet is the Sonnet-specific weekly cap. There is also overage, which tracks pay-as-you-go extra usage once a weekly cap is hit and the user has opted into overage billing. All five map to labels through the rateLimitTypeLabel static function at ChatProvider.swift lines 537 to 546.

Anthropic updated its consumer terms and acceptable-use policy early in April 2026, and users who did not re-accept on claude.ai started getting API errors containing phrases like 'consumer terms', 'terms of service', 'accept', or 'terms and privacy'. Fazm classifies these with a separate matcher called isTermsAcceptanceRequired at ChatProvider.swift lines 1370 to 1382, and when it fires the app does not try to re-auth the user (re-auth does not fix a T&S acceptance gap) but surfaces the raw error and points the user to claude.ai. The T&S matcher is deliberately distinct from the model-access matcher because the remediation is different.

In one direction. A user whose model was silently moved from Smart to Fast can re-pick Smart at any time through the model toggle in the floating bar chrome or the Ask Fazm settings panel. The migration guard (shortcut_didMigrateFromOpus) stays true after that, so the one-time migration does not fire again. If the user picks Smart and then later updates Fazm, they stay on Smart. The migration only ever moves a user once, and only in the Opus-to-Sonnet direction. There is no corresponding Sonnet-to-Opus auto-migration.

Open /Users/matthewdi/fazm/Desktop/Sources/FloatingControlBar/ShortcutSettings.swift. Lines 139 to 141: the @Published selectedModel property. Lines 151 to 155: the three default model IDs and their friendly labels. Lines 168 to 189: the ACP-driven updateModels function that relabels models dynamically. Lines 287 to 296: the one-time Opus-to-Sonnet migration. Lines 288 to 289: the two UserDefaults keys. Line 293: the in-memory write. Then open /Users/matthewdi/fazm/Desktop/Sources/Providers/ChatProvider.swift. Lines 537 to 546: the four rate-limit type labels. Line 1018: the ACP models_available handler wiring. Line 1048 to 1050: the three ACP sessions (main, floating, observer) all default to claude-sonnet-4-6.