The Anthropic API changelog for April 2026, sorted by what actually broke a build

The headline entries were Claude Opus 4.7 reaching GA on April 16 at the same $5 / $25 per million token pricing as 4.6, Managed Agents going public beta on April 8 under the managed-agents-2026-04-01 header, the Advisor tool public beta on April 9 under advisor-tool-2026-03-01, a formal deprecation of Claude Sonnet 4 and Opus 4 announced April 14 with retirement scheduled for June 15, 2026, the scheduled retirement of the 1M context beta for Sonnet 4 and 4.5 on April 30, Claude Haiku 3 retiring on April 20, and an April 23 admission from Anthropic that a caching bug had been dropping thinking history and a verbosity prompt regression had quietly degraded Claude Code output quality. The caching bug and the verbosity regression were reverted and affected users had their usage limits reset.

No. Most of them absorbed automatically. Fazm consumes the Claude API through the Agent Client Protocol (ACP) bridge that wraps the Claude Code agent SDK. When Anthropic flipped Opus 4.7 on at the SDK level, the bridge's models_available event fired with the new model ID, ShortcutSettings.updateModels rebuilt the picker with a new Smart (Opus, latest) row, and the user saw it the next time they opened the floating bar. The one entry that did require code was the v0.29 SDK rename of the Opus slot, which broke the saved user preference until Fazm 2.4.0 (2026-04-20) shipped the normalizeModelId function to migrate it.

normalizeModelId lives at Desktop/Sources/FloatingControlBar/ShortcutSettings.swift line 170. It substring-matches a model ID. If the string contains 'haiku' it returns 'haiku'. If it contains 'sonnet' it returns 'sonnet'. If it contains 'opus' it returns 'default'. The third case exists because ACP SDK v0.29+ renamed the Opus slot from 'opus' to 'default' in its availableModels payload. Users who had previously selected Smart (Opus) had the string 'opus' saved in UserDefaults under the key shortcut_selectedModel. On v2.4.0 launch, the init code at lines 325-333 reads that saved string, runs it through normalizeModelId, writes the result back to UserDefaults, and the picker keeps matching the newly renamed slot. Without this function, Opus users silently fell back to Sonnet after updating Fazm.

The ACP SDK used by Claude Code reports availableModels as a small list of human-ish slots, not raw API model IDs. Before v0.29 the three slots were haiku, sonnet, and opus, mapped server-side to whichever versions the user's auth had access to. In v0.29 the top slot was renamed 'default' so the SDK could ship the user's best-available model (Opus 4.7 today, whatever comes next tomorrow) without the slot name implying a specific family. The rename is forward-compatible for any new user. For existing users with stored selections it is a breaking change, which is why a client-side migration is required.

Opus 4.7 GA on April 16 landed on day zero because Fazm does not hardcode model IDs; the picker rebuilds from the ACP availableModels response. The April 23 cache-and-verbosity revert required nothing on the client side because the bug was entirely server-side. The April 14 Sonnet 4 / Opus 4 deprecation notice is a future concern rather than a breaking change today. The April 30 1M context beta retirement is silently absorbed because Fazm's main chat does not request the beta header. Haiku 3 retirement on April 20 had no effect because Fazm was already pointing the Scary slot at the newer Haiku family. The Messages API landing on Amazon Bedrock on April 7 was irrelevant because the app talks to Anthropic-hosted Claude through Claude Code, not Bedrock.

Not directly, and the reason is the most interesting one for consumers. Managed Agents run inside Anthropic's sandbox. They cannot see an individual user's Messages.app, Mail, Notes, Figma document, or Logic Pro session because they have no access to the Mac's accessibility tree and no local file-system permissions. Fazm is the opposite shape: it runs locally on the user's Mac with Accessibility permission granted and drives real apps through AXUIElement calls in Desktop/Sources/AppState.swift. The Advisor tool is interesting as an architectural pattern (a fast executor paired with a stronger advisor), and Fazm's three-session warmup in ChatProvider.swift lines 1047-1051 already echoes it in spirit, but we did not rewire anything for April's beta.

Yes, visible in /tmp/fazm-dev.log as longer observer-batch latencies and a few turns where the assistant re-asked a question it had already answered. Because Fazm talks to Claude through Claude Code's SDK, the thinking-history drop showed up as the main chat losing short-range context on long sessions. No Fazm code change was needed. Anthropic reverted the bug on April 23, affected accounts had their usage limits reset per Anthropic's own statement, and the symptom cleared without an app update.

The acp-bridge subprocess calls session/new against Claude Code on every warmup. The response includes models.availableModels, scoped to the user's authenticated plan. The bridge filters out the sentinel default pseudo-model if the real list has other entries, dedupes against the last emitted list, and sends a models_available event up to the Swift app at acp-bridge/src/index.ts lines 1271-1281. On the Swift side, setModelsAvailableHandler routes to ShortcutSettings.updateModels, which runs the family-map substring match and publishes @Published availableModels. SwiftUI re-renders the picker automatically. No user action required.

Arguably the v0.29 ACP SDK rename itself. If Fazm called the Messages API directly with a hardcoded claude-opus-4-7 model string, the rename would not have applied and normalizeModelId would not exist. The trade is that direct Messages API calls lose automatic model availability from the user's auth, lose the ACP session resume protocol, lose the dynamic model list, and lose the memory system that ships with Claude Code. The migration function is a one-time cost for durable benefits everywhere else.

Opus 4.7 shipped an updated tokenizer on April 16, 2026. Token counts for the same input string may differ slightly from Opus 4.6 output. For Fazm this is invisible because the app does not enforce a local token budget per-turn. It relies on the Anthropic API to report actual usage, which Claude Code surfaces through the ACP protocol. If the app had been computing local token estimates for a progress bar or paywall gate, those estimates would be off by a few percent until the tokenizer was swapped client-side. This is a silent absorbing case.