The April 2026 LLM release that actually reached a Mac user was one substring match, not a benchmark chart.

The month was unusually dense. Anthropic shipped Claude Opus 4.7 GA on April 2 along with a Claude Mythos Preview on April 7 (limited to ~50 partner organizations via Project Glasswing). OpenAI released GPT-5 Turbo on April 7 with native image and audio generation inside a single model. Google released the Gemma 4 family under Apache 2.0 on April 2. Zhipu released GLM-5.1 (754B MoE, MIT license) and GLM-5V-Turbo. Alibaba released Qwen 3.6-Plus with 1M-token context. DeepSeek released V3.2. xAI released Grok 4.1. At least nine significant models shipped in the first two weeks alone, and that count undercounts because it excludes point-release updates like Opus 4.7.1 and Haiku minor bumps.

Inside Fazm, close to nothing. The ACP bridge reads the model list from the Claude Agent SDK's session response (result.models.availableModels), hashes it against the last broadcast, and if it changed calls send({ type: 'models_available', models: filtered }) at acp-bridge/src/index.ts line 1278. The Swift UI subscribes to that message, runs it through a three-row substring table at ShortcutSettings.swift lines 160 to 163, and re-renders the button row. When the Anthropic SDK starts returning 'claude-opus-4-7' or 'claude-mythos-preview' in its availableModels, a Fazm user with zero app updates sees new buttons appear the next time they open the app. The one place an app update is required is when the SDK itself rev-locks to a new major protocol shape.

The entire pipeline is anchored on three const strings and one substring-match table. First const: DEFAULT_MODEL = 'claude-sonnet-4-6' at acp-bridge/src/index.ts line 1245. This is the only place in the whole bridge where a model name is hardcoded, and it exists solely as the fallback when no model is specified on an incoming query. Second const: SONNET_MODEL = 'claude-sonnet-4-6' one line below, used by pre-warmed sessions. Third const: the events literal 'models_available' used at index.ts line 1278 inside send({ type: 'models_available', models: filtered }). The substring-match table is the four-line modelFamilyMap at ShortcutSettings.swift lines 160 to 163 that bins any new model ID into one of three consumer-facing families: Scary (Haiku), Fast (Sonnet), Smart (Opus). When OpenAI's GPT-5 Turbo or Zhipu's GLM-5.1 show up via a custom endpoint with names that do not contain any of those three substrings, ShortcutSettings.swift lines 188 to 190 fall through to 'use the API name directly' and the model renders as its raw ID.

Five hops. First, the provider adds the new model to its server-side catalog (Anthropic's console, the OpenAI API, a self-hosted Gemma 4 server exposing an Anthropic-compatible interface). Second, the Claude Agent SDK's session/new response returns it inside the models.availableModels array. Third, the ACP bridge receives that response inside handleQuery's pre-warm path and calls emitModelsIfChanged (index.ts lines 1271 to 1280). Fourth, the bridge sends a single stdout JSON line: { type: 'models_available', models: [{ modelId: 'claude-opus-4-7', name: 'Opus 4.7, latest' }, ...] }. Fifth, ACPBridge.swift's read loop decodes it and calls ShortcutSettings.updateModels, which re-bins every new ID through the substring family table at lines 180 to 216 and republishes availableModels on the @Published var. The SwiftUI ForEach at SettingsPage.swift lines 2013 to 2031 re-renders the button row on the next tick.

Yes. That is what the session key is for. Every incoming query carries an optional model field, and acp-bridge/src/index.ts line 1390 computes incomingSessionKey = msg.sessionKey ?? (msg.model || DEFAULT_MODEL). If two surfaces (say main and floating) each specify a different model, the bridge routes them to two separate ACP sessions with independent sessionIds, and acpRequest('session/set_model', { sessionId, modelId }) at lines 1341 and 1366 sets the per-session model server-side. The practical consequence: you can run Opus 4.7 in the main chat and Claude Mythos in the floating bar simultaneously in one Node.js process. The per-session isolation also means one session's compaction, tool calls, and rate-limit hits do not leak into the other session's state.

ShortcutSettings.swift lines 184 to 190 handle that case explicitly. The family map compactMap returns nil for unknown model IDs, then the fallback branch constructs a ModelOption with label = model.name (the SDK-provided display name, e.g., 'Gemma 4 12B Instruct' or 'GLM-5.1-754B') or falls back to the raw modelId, and gives it order=99 so it sorts after the three named families. The button still renders, just with its literal SDK name instead of the Scary/Fast/Smart alias. This is why a user running Fazm against a Gemma 4 or GLM-5.1 custom endpoint sees those models in the picker with their real names rather than being blocked.

DEFAULT_MODEL = 'claude-sonnet-4-6' at acp-bridge/src/index.ts line 1245 is the only piece of the pipeline that has to change when a new month's minor upgrade ships. When Anthropic rotates the default alias from sonnet-4-6 to sonnet-4-7, Fazm updates that one constant in the next release. Everything else, including the user's picker, rate limit routing, and session pre-warming, reads from that constant. The rest of the 2,700-line bridge never hardcodes another model ID. This is why a keyword like 'llm release april 2026' ends up being answered by 'change one string' rather than 'ship a feature'.

Through the customApiEndpoint setting, yes, with one constraint: the endpoint must speak an Anthropic-compatible wire protocol. LiteLLM, Cloudflare AI Gateway, and any self-hosted proxy that translates Anthropic's /messages shape into OpenAI, Gemini, or vLLM work. When you point Fazm at a proxy fronting GPT-5 Turbo or Gemma 4, the ACP SDK still negotiates available models over its own protocol, and whatever IDs come back are fed through emitModelsIfChanged. The models_available stdout line then surfaces them in the picker with their real names. The exact user-visible consequence: switching Fazm to a GPT-5 Turbo proxy and opening the picker shows a new 'gpt-5-turbo' button next to Scary/Fast/Smart, with no recompile.

Because the labels are the product, not the model IDs. A non-technical Mac user reading 'claude-sonnet-4-6' vs 'claude-opus-4-7' has no way to tell what they should pick. ShortcutSettings.swift line 160 hardcodes the answer: Haiku is Scary (means 'fast and cheap, so it runs silently in the background'), Sonnet is Fast (general-purpose default), Opus is Smart (for hard problems). The labels are deliberately evaluative, not descriptive. When a new model family ships in April 2026, the substring table preserves that evaluative shelf: opus-4-7 is still Smart, sonnet-4-7 is still Fast. The shelf is stable even as the underlying model IDs rotate monthly.

Nothing. The per-model session isolation at acp-bridge/src/index.ts line 1390 means each model has its own ACP sessionId, its own CWD, its own ongoing query state, and its own abort controller. When you click from Smart to Fast in the picker, your previous session keeps running (or stays paused) under its own key, and the bridge routes the next query to whatever session key the new model maps to. The explicit code at lines 1449 to 1504 handles three cases: session key already exists (reuse it, possibly calling session/set_model if the model changed), session key new (create a new ACP session), and model swap inside a pre-existing key (call session/set_model at line 1475). The result is that model switching is essentially free: no connection drop, no context loss for either session.

The three-const anchor is at /Users/matthewdi/fazm/acp-bridge/src/index.ts lines 1245 to 1290 (DEFAULT_MODEL, emitModelsIfChanged, and the broadcast). The substring-match table is at /Users/matthewdi/fazm/Desktop/Sources/FloatingControlBar/ShortcutSettings.swift lines 151 to 216 (defaultModels, modelFamilyMap, normalizeModelId, updateModels). The per-model session key is at index.ts line 1390. The session/set_model calls are at index.ts lines 1341, 1366, 1475, 1495, 1503, and 1810. The UI button row that re-renders on every models_available broadcast is at /Users/matthewdi/fazm/Desktop/Sources/MainWindow/Pages/SettingsPage.swift lines 2013 to 2031. Everything is visible to anyone who opens the repo; there is no hidden configuration server or LaunchDarkly flag gating any of it.