Opus 4.7, Mythos, the $800B round, and the 12-line Swift function that decides whether any of it reaches your Mac

The April 2026 news cycle around Anthropic was dense. The most-cited items: (1) the general-availability release of Claude Opus 4.7 with stronger software-engineering and long-running agent reliability, at the same $5 / $25 per MTok pricing as 4.6; (2) the redesigned Claude Code desktop app and the Routines research preview, announced April 14; (3) the Mythos safety story (Bloomberg, April 16) where Anthropic disclosed a model called Mythos that was held back from general release after internal red-teaming flagged cyber-offense capabilities; (4) the Project Glasswing controlled rollout to roughly 40 enterprises including Amazon, Apple, Microsoft, and JPMorgan Chase; (5) the $800B-valuation funding offers reported on April 14; (6) the April 6 expanded Google Cloud / TPU partnership and the related Broadcom multi-gigawatt deal; (7) the Long-Term Benefit Trust appointing Vas Narasimhan to the board on April 14. None of these items, by themselves, tell a Mac user whether their chat session will keep working after the rollout.

Because Opus 4.7 was released at the same price point as 4.6 but not every Claude consumer plan got access on day one. Free-tier accounts and certain regional plans saw a window where requests for Opus models returned the error string 'model not found' or 'may not exist or you do not have access to it.' Apps that hard-coded a personal-OAuth-only architecture surfaced that error to the end user. Apps with a fallback path swallowed it. Fazm's ChatProvider.swift implements that fallback in a 12-line static function (isModelAccessError, line 1362) that matches three specific substrings, then a control-flow branch at lines 2944 to 2956 that sets pendingBridgeModeSwitch = 'builtin' and auto-retries the same query under a bundled Anthropic API key.

Mythos is an internal Anthropic model that, per a Bloomberg report on April 16, 2026, performed strongly enough on cyber-offense red-team evaluations that Anthropic's own safety team concluded it could not be released as a general API. Project Glasswing is the restricted access program where Mythos is being made available to roughly 40 organizations (Amazon, Apple, Microsoft, JPMorgan Chase, others) under a controlled rollout. For consumer Mac apps the effect is straightforward: most users will never see a Mythos endpoint, and apps must continue to assume Sonnet, Opus, and Haiku as their available model surface.

Routines is a Claude Code feature, previewed alongside the redesigned desktop app on April 14, that lets a user define repeatable agent workflows that can be triggered later (similar in spirit to scheduled or saved tasks). It is a research preview, not a stable API. Fazm's existing ScheduledTasks subsystem (Sources/MainWindow/Pages/ScheduledTasksPage.swift) is unrelated and predates Routines. They cover overlapping ground but do not interoperate as of April 2026.

Three locations. First, the bridge mode is declared as @AppStorage("bridgeMode") var bridgeMode: String = "builtin" at line 439 of /Users/matthewdi/fazm/Desktop/Sources/Providers/ChatProvider.swift. The two valid values are 'builtin' and 'personal'. Second, the BridgeMode enum lives at /Users/matthewdi/fazm/Desktop/Sources/Chat/ACPBridge.swift lines 192 to 200, with cases .personalOAuth and .bundledKey(apiKey:). Third, the fallback decision happens in the post-query error handler at lines 2944 to 2956 of ChatProvider.swift: if the bridge is in personal mode and isModelAccessError(msg) returns true, the handler sets pendingBridgeModeSwitch = 'builtin' and retryAfterModelFallback = true, then re-runs the original message under bundled credentials.

Lines 1362 to 1369 of ChatProvider.swift: (1) lower.contains("may not exist") && lower.contains("not have access"), (2) lower.contains("model") && lower.contains("not found"), (3) lower.contains("model") && lower.contains("not available"). All three are AND conditions on substrings of the lowercased server error. Anthropic's API has used variants of all three phrasings during the April 2026 rollout windows. The matcher is intentionally tolerant of small wording changes.

Line 476 of ChatProvider.swift: static let builtinCostCapUsd: Double = 10.0. This is a hard cap on how much Anthropic spend the bundled API key path will absorb per user, tracked locally via @AppStorage("builtinCumulativeCostUsd") and seeded from Firestore on startup. The cap exists because the bundled key is a real Anthropic key Fazm pays for. Past $10, the next outgoing query short-circuits at line 2272 and triggers a forced switch to personal mode (showCreditExhaustedAlert = true). The user is asked to sign into their own Claude account at that point. The number is conservative on purpose: it is enough to evaluate the product end-to-end without a credit card, not enough to be exploited.

Lines 2902 to 2917 of ChatProvider.swift. When the bridge throws BridgeError.creditExhausted with a raw message, the handler runs let isRateLimit = rawMessage.range(of: #"resets\s+\S"#, options: .regularExpression) != nil. The 'resets <something>' pattern is what the Anthropic API includes when the limit is time-bounded (a five_hour or seven_day window). If isRateLimit is true and the bridge is in builtin mode, Fazm shows the rate-limit message but does NOT switch modes, because the user still has bundled-key budget. If isRateLimit is false, it is real credit exhaustion: the bridge mode switches to 'personal' via switchBridgeMode(to: "personal") and the credit-exhausted alert is raised.

Lines 1370 to 1382 of ChatProvider.swift implement isTermsAcceptanceRequired, a separate matcher from the model-access matcher. It looks for four signals in the server error: 'consumer terms', 'terms of service' AND 'accept', 'terms and privacy', and 'updated our' AND 'policy'. When Anthropic ships a T&S update (as happened in early April 2026 around the rollout of new acceptable-use language), users who have not re-accepted the terms on claude.ai see API errors. Fazm classifies these explicitly because re-authing the user does NOT fix them. The user has to visit claude.ai directly. The error is surfaced verbatim instead of triggering an auth flow.

Lines 538 to 543 of ChatProvider.swift list four rate-limit type labels the bridge knows about: 'five_hour' (session limit), 'seven_day' (weekly limit), 'seven_day_opus' (Opus weekly limit), 'seven_day_sonnet' (Sonnet weekly limit). The seven_day_opus type is the one most users will encounter in April 2026 because Opus 4.7 is more expensive and Anthropic enforces the cap more aggressively. When the rate-limit type comes back as seven_day_opus, the in-app message reads 'Opus weekly limit, resets <date>'. The user is told to switch to Sonnet (or wait for the reset).

Both, in different paths. The 'personal' bridge mode launches a Node.js subprocess that talks to the Claude Agent SDK's ACP (Agent Client Protocol) layer over JSON-lines on stdin/stdout. The user's Claude.ai OAuth token is stored by the Claude Code CLI in the macOS Keychain under the service name 'Claude Code-credentials' (referenced in AppState.swift), and the bridge picks it up from there. The 'builtin' mode bypasses the OAuth path entirely and sets ANTHROPIC_API_KEY in the subprocess env (ACPBridge.swift line 351). Both modes go through the same ACP message loop, the same tool-call dispatcher, and the same skill library at ~/.claude/skills.

Nothing visible during the fallback itself. The model-access fallback path at lines 2945 to 2956 sets errorMessage = nil, swaps the bridge mode, and re-runs the query. The user sees the assistant continue typing as if nothing happened. The only visible signal is in Settings > Model selector, where the bridge mode badge updates to 'Built-in.' If the cap is hit later, the credit-exhausted alert fires and the user is asked to sign in with their personal Claude account. That second case is loud on purpose: someone has to take over paying for the conversation.

Yes. Open /Users/matthewdi/fazm/Desktop/Sources/Providers/ChatProvider.swift. Line 439: bridge mode declaration. Line 476: builtinCostCapUsd = 10.0. Lines 491 to 506: createBridge() function with the personal/builtin switch. Lines 1360 to 1369: isModelAccessError with the three substring rules. Lines 1370 to 1382: isTermsAcceptanceRequired. Lines 2272 to 2275: cap enforcement at send time. Lines 2897 to 2935: credit-exhausted handling and rate-limit-vs-cost classifier. Lines 2944 to 2956: model-access fallback control flow. The BridgeMode enum is at /Users/matthewdi/fazm/Desktop/Sources/Chat/ACPBridge.swift lines 192 to 200. Every claim in this guide maps to a specific file and line range a reader can open.