The month Anthropic announced Opus 4.7, Mythos, and a $25B Amazon deal. One Fazm pipeline stayed on Gemini on purpose.

Five threads ran through the month. Claude Opus 4.7 went generally available on April 22 at $5 per million input tokens and $25 per million output tokens, matching Opus 4.6. Claude Mythos was previewed via Project Glasswing, a gated cybersecurity program with Amazon, Apple, Microsoft, and Nvidia. Amazon expanded its Anthropic investment to the $25 billion range with Anthropic committing more than $100 billion to Amazon cloud and chips over ten years. Google and Broadcom expanded TPU supply. JPMorgan rolled Claude out internally while Dario Amodei met with Susie Wiles and Scott Bessent after the Pentagon flagged supply-chain concerns about Mythos. On the SDK side the Agent Client Protocol moved from 0.25.0 on April 7 to 0.29.2 on April 20, which is the version Fazm 2.4.0 upgraded to.

Because Fazm has two very different jobs and they do not belong to the same model. The interactive chat, every tool call, every prompt you type, every accessibility read of another Mac app, and every floating bar ask run through Claude via the ACP bridge. That is the low-latency, tool-heavy, action-taking side. A second, separate job exists: watch ~60 minutes of short active-window screen recordings and decide whether an AI agent could take one task off the user's plate. That job wants a long multimodal video context and does not need to hold a conversation. At GeminiAnalysisService.swift line 67 Fazm hard-codes `private let model = "gemini-pro-latest"` and routes those chunks there. None of April's Anthropic announcements changed the answer for that one job.

All four live at the top of /Users/matthewdi/fazm/Desktop/Sources/GeminiAnalysisService.swift. Line 67 sets `private let model = "gemini-pro-latest"`. Line 68 sets `private let maxChunks = 120` as the hard cap on buffered chunks. Line 69 sets `private let targetDurationSeconds: TimeInterval = 3600`, so Fazm waits for roughly 60 minutes of recordings before firing one analysis. Line 71 sets `private let inlineSizeLimit = 1_500_000`, the 1.5 MB threshold above which a chunk uses the Gemini File API resumable upload instead of an inline base64 body. The chunk format itself is set by the feeder: SessionRecordingManager.swift lines 76 through 83 build a `SessionRecorder.Configuration` with `framesPerSecond: 2.0`, `chunkDurationSeconds: 60.0`, and `captureMode: .activeWindow`. Two frames per second, one minute per chunk, only the frontmost app window.

No, and that is the point. Opus 4.7 GA only affected the ACP bridge path, which is Fazm's Claude side. The relevant work there was the ACP v0.29 model ID rename, the partial models_available label fix in v2.4.1 on April 22, and the dynamic model list picked up at ShortcutSettings.swift line 180. The Gemini path shipped no change during the April release train. The same model string, the same 60-minute window, the same 1.5 MB cutoff. You can verify this by scanning `git log --since=2026-04-01 --until=2026-04-30 -- Desktop/Sources/GeminiAnalysisService.swift` and comparing the result to the Claude-side files.

Three reasons, all spelled out in SessionRecordingManager.swift around line 76. First, 2 FPS is called out in the source as `Lower FPS than research recorder to save CPU` — this is a background observer that must not heat the user's laptop. Second, `captureMode: .activeWindow` restricts each frame to the frontmost app's window only, both for privacy and to make the resulting video actually useful, since a full desktop frame with every other window in shot is mostly noise. Third, chunking at 60 seconds lets the analyzer reason about a window of activity rather than a single moment. The Gemini analysis prompt literally tells the model: `You are watching ~60 minutes of a user's screen recording. Each video clip captures the active window of whatever app the user was using at that moment.`

Technically yes, but the constraints fight the use case. Claude's vision path accepts images, not multi-hour video. Feeding 120 sixty-second clips would require chunking to stills, which throws away exactly the temporal signal the task needs. Gemini's File API accepts the MP4 directly and lets a single request span the whole hour. The separation is not a judgment about which model is better at reasoning, it is a judgment about which transport fits the input shape. Claude stays on conversational agent work where it is strongest. The video summarizer stays on the pipe that was built for video.

They land in the `observer_activity` table as rows with `type = 'gemini_analysis'`. The main window surfaces them through /Users/matthewdi/fazm/Desktop/Sources/MainWindow/Pages/DiscoveredTasksSection.swift. Each row carries the verdict (`NO_TASK`, `TASK_FOUND`, or `UNCLEAR`), a task summary, a description, and a fuller markdown document. The prompt in GeminiAnalysisService.swift explicitly forces `Return UNCLEAR if: you can't make out what the user is doing, the content is ambiguous, or you'd be guessing` and requires checking the last ten rows of `observer_activity` and the last ten chat messages before suggesting anything new. That guard is what keeps the same suggestion from appearing twice.

No. `lastFailedAnalysis: Date?` at GeminiAnalysisService.swift line 77 plus `retryCooldown: TimeInterval = 300` on line 78 set a five-minute cooldown after any failed analysis. During April 2026 this was a useful property, because Claude saw a lot of capacity turbulence around the Opus 4.7 launch window, but the Gemini path was insulated from that by definition. When Gemini itself had a blip, the cooldown prevented spamming the API. Claude outages on the chat path never affected task discovery, and Gemini outages on the task path never affected chat. That is the whole value of keeping them split.

The analysis path uses a Fazm-managed Gemini key on a capped budget, the same managed pattern Fazm uses for the built-in Claude account. Users do not bring a key to watch their own screen recordings. This is consistent with the privacy position documented in the v2.3.2 release on April 16: `Tightened privacy language in onboarding and system prompts to accurately say 'local-first' instead of 'nothing leaves your device'`. Chunks leave the device only when the 60-minute analysis fires, and the model decides one task or returns NO_TASK or UNCLEAR.

Screenshot-based agents take a still image on demand, send it with a prompt, and ask the model to reason about what is on screen right now. That is a synchronous, foreground interaction. Fazm's Gemini analysis path is none of those things. It is an asynchronous background summarizer watching an hour of 2 FPS active-window video after the fact, and it is reading the user's own macOS accessibility tree for everything Claude needs in real time. The two mechanisms exist because they answer two different questions. Screenshots are good for `what is happening now`. An hour of video is good for `what have you been doing and could I help`. Fazm ships both, on different models, on purpose.