The Opus 4.7 release, from inside the desktop app that had to render it

Claude Opus 4.7 went generally available in April 2026 at the same $5 / $25 per million tokens pricing as Opus 4.6. Anthropic frames the upgrade as stronger software engineering, higher-resolution vision, sharper instruction following, and more reliable long-running agent work. The same month, Anthropic unveiled a second model, Claude Mythos (internal codename Capybara), through Project Glasswing, a gated early-access program limited to roughly 50 partner organizations with no public release date. A third surface, Claude Design, launched as a Labs product for creating designs, prototypes, slides, and one-pagers, but it is a product on top of the existing models rather than a new model.

Because a long-running agent emits more events. A short turn might be one text block plus one tool call. A long turn on Opus 4.7 can spawn nested subtasks, stream tool progress while a task is running, post a final tool-use summary when several calls finish, and split its output across multiple text blocks separated by tool use. If a client only listens for text_delta and tool_call, those richer turns arrive as silence punctuated by jumps in the text. Fazm had to forward five additional Agent Client Protocol update types to the Swift app so the UI can represent what Opus 4.7 is actually doing: task_started, task_notification, tool_progress, tool_use_summary, and compact_boundary.

Two hops. First, the ACP bridge, a long-lived Node.js subprocess, handles the raw ACP update at acp-bridge/src/index.ts:2402. It pulls taskId and description out of the update, adds taskId to a per-turn Set<string> at line 2405, and sends a typed task_started event to the Swift app. Second, the Swift side, ChatProvider.swift:2600-2607, receives the event and calls addToolActivity with toolName 'Subtask', status .running, and the taskId as toolUseId. The chat view already knows how to render a tool activity row, so the subtask inherits the existing animation, status pill, and completion behavior.

On a long turn it is possible for a task_notification to arrive after the turn has ended, because the underlying Claude Code SDK can deliver late updates for a task that was logically finished or cancelled. Without guarding for that, you see yesterday's subtasks blink into today's chat. The bridge keeps a per-turn Set<string> called currentTurnTaskIds, declared at acp-bridge/src/index.ts:1531 and cleared at line 1728 when the turn ends. When a task_notification arrives, the handler at line 2416 checks whether the task's taskId is still in the set. If it is not, the handler calls onStaleNotification (wired at line 1577 to increment staleTaskNotificationCount) and tags the log line with 'STALE'. The client still forwards the event so the Swift side can decide whether to suppress it, but it is now labeled.

No. Short turns look the same as they did on Opus 4.6: agent_message_chunk for text, tool_call for tool use, a closing stop reason. task_started appears when the model decomposes the work into named subtasks, typically on multi-step requests that involve planning across files or apps. Fazm's rate-limit behaviour, streaming throttle, and session pre-warming did not need to change for Opus 4.7, which is exactly the kind of release note that does not ship: 'we changed nothing and users still saw every new feature.'

The set is cleared at acp-bridge/src/index.ts:1728 inside the per-turn cleanup block, right after the prompt resolves and before the handler runs its post-turn bookkeeping. Clearing on turn end, instead of on session end, is what makes the detector correct: if the user sends a new prompt in the same session a few seconds later, a late task_notification from the previous prompt gets flagged as stale instead of being attributed to the new prompt's taskIds.

task_started marks the beginning of a named subtask (a logical unit of work the model has planned). tool_progress is a periodic heartbeat emitted while a single tool call is running, carrying toolUseId, toolName, and elapsed time in seconds. On Opus 4.6 with shorter turns, tool_progress was nice-to-have. On Opus 4.7, where a single tool call can run long enough that the UI would look frozen without feedback, tool_progress is the thing that lets the activity row show '[WebSearch] running 12s' instead of a spinner that just keeps spinning. It is handled by Fazm's bridge at acp-bridge/src/index.ts:2427-2433 and logged on the Swift side at ChatProvider.swift:2616-2617.

No. Fazm does not hardcode a model list. On each session start, the ACP SDK reports the models the user's auth has access to via an availableModels array, the bridge forwards it as a models_available event, and the Swift picker rebuilds from scratch. When Anthropic flipped Opus 4.7 on, it appeared in the picker the next time the floating bar opened, relabeled 'Smart (Opus, latest)' by the substring family map in ShortcutSettings.swift. Zero signed-and-notarized desktop release, zero Sparkle auto-update, zero user action. The task-tracking event handlers described above were already there from a previous release, waiting for a model that emits them frequently enough to matter.

A Computer Use style agent screenshots the screen each turn and asks the model to OCR it. Opus 4.7's higher-resolution vision is a clear win for that class of agent because the bottleneck is reading the picture. But on long-running work, that agent still has to figure out how to render progress to the user, which means wiring up task_started, task_notification, and tool_progress the same way. The client-side plumbing problem is orthogonal to the vision-vs-structure tradeoff. Every Opus 4.7 host has to solve it.

Chronologically yes, practically no. Mythos is distributed only to the small set of partner orgs inside Project Glasswing and is not available through any public API, plan, or console. Anthropic has described Mythos as capable of identifying and exploiting software vulnerabilities with high accuracy, which is why the rollout is gated. If you are building on Claude today, the GA ceiling is Opus 4.7 and that is what you should plan your next six months around. If Anthropic ever opens Mythos GA, Fazm's substring family map and availableModels forwarding will pick it up the same way they picked up Opus 4.7, with no new build.

The handleSessionUpdate function in acp-bridge/src/index.ts (roughly lines 2071 through 2489) is the canonical switch. The cases relevant to the April 2026 new-model release are: agent_message_chunk (text), agent_thought_chunk (thinking delta on a separate channel), tool_call and tool_call_update (tool lifecycle), plan, compact_boundary, status_change, compaction_start, compaction_delta, task_started, task_notification, tool_progress, tool_use_summary, rate_limit, api_retry, and usage_update. Anything not in that switch logs an 'Unknown session update type' line and is ignored, which is the right behavior when the ACP surface adds something new in the future.