April 2026 shipped seven frontier LLMs. Inside our binary, two of them route a token.

Two providers, four roles. On the Anthropic side, three Claude sessions are pre-warmed at startup (main, floating, observer) and all three are hardcoded to claude-sonnet-4-6 at ChatProvider.swift line 1048 through 1050. The picker can route a single user query to claude-opus-4-7 via the Smart button, but the always-on sessions stay on Sonnet. On the Google side, GeminiAnalysisService.swift line 67 pins a fourth role to the alias gemini-pro-latest, which currently resolves to Gemini 2.5 Pro after Google's April 1 GA. That is it. GPT-5 Turbo, DeepSeek V4, Gemma 4, Llama 4 Scout, Llama 4 Maverick, GLM-5.1, and Qwen 3.6-Plus do not route a single token through the shipping binary today.

The buffer is set at GeminiAnalysisService.swift line 69 to targetDurationSeconds: 3600. After every 60 minutes of session recordings, the loop concatenates 60 video clips of whatever app was focused at each moment, sends them to Gemini's File API, and asks for a verdict: NO_TASK, TASK_FOUND, or UNCLEAR. The reason this lives on Gemini and not on Claude is multimodal video. Anthropic's API does not accept native video uploads at the latency or cost Fazm needs for an hourly background task, so the screen-watching loop runs on Gemini and the conversational layer runs on Claude. They never share a session. The split is by media type, not by quality.

Because the model field at GeminiAnalysisService.swift line 67 is the literal string gemini-pro-latest, not gemini-2.0-pro or gemini-1.5-pro. Google's API resolves that alias server-side to whichever Gemini Pro is the current GA model. April 1 was the GA cutover for 2.5 Pro, and from that morning every Fazm install in the wild started pulling 2.5 Pro for screen analysis without a single download, sparkle update, or App Store review. The trace at line 264 records the resolved name as gemini-2.5-pro, which is what shows up in the user's external_llm_traces table after April 1.

Cost. The observer session is the largest token consumer in the binary — it watches every conversation batch and writes memory entries to local files. If the observer ran on Opus 4.7 (5 dollars input, 25 dollars output per million tokens) instead of Sonnet 4.6 (3 dollars, 15 dollars), the bill for an active user would roughly triple, with no measurable benefit on the kind of structured-extraction work the observer does. The same logic holds for the floating bar's default queries: most of them are short follow-ups where Sonnet's quality is already at the ceiling. Opus 4.7 is reachable, but only when the user explicitly asks for it by selecting the Smart label in the picker.

Every non-Anthropic, non-Google release. GPT-5 Turbo would need a second ACP-style bridge that talks to OpenAI's responses API, with its own emitModelsIfChanged so the dynamic picker can absorb it the same way it absorbed Opus 4.7. DeepSeek V4-Pro and V4-Flash would need either a hosted-API bridge or a local-inference bridge, depending on whether you want cloud or on-device. Gemma 4 31B and the Llama 4 family are open-weight, so a bridge to llama.cpp or vLLM running locally is the natural target. Each of these is a discrete piece of work, not a config flag — until that bridge exists, the picker simply cannot show those models.

Almost nothing visible. The dynamic picker absorbs the new ID at the next session/new call without a Fazm build. If the user had a stored selection like claude-opus-4-6, the upgrade chain at ShortcutSettings.swift line 208 through 210 finds a model whose ID contains opus and silently migrates the selection forward. The label they see is still Smart. The keyboard shortcut they bound is still bound. The Settings preference pane is still consistent. The only thing that changed is the model ID in the underlying selectedModel string and the API surface the next query lands on.

Yes. ShortcutSettings.swift line 299 through 301 holds @Published var screenObserverEnabled, defaulting to true. With it off, no clips are buffered, no 60-minute analysis fires, no chunks accumulate at Application Support/Fazm/gemini-analysis. Gemini 2.5 Pro's daily token contribution drops to zero for that install, and the cost shifts entirely onto the Anthropic side. Disabling the observer is also the cleanest way to verify how much of your token spend was background analysis versus interactive chat, because the recordLlmUsage call at GeminiAnalysisService.swift line 254 tags those tokens with account observer_gemini.

Different blast radii. Gemini's analyzer runs once an hour in the background — if the alias quietly upgrades from 2.0 Pro to 2.5 Pro, the worst case is one analysis comes back differently shaped, which the prompt at line 56 through 64 already tolerates with the NO_TASK verdict. The conversational layer is real-time and user-facing, so a silent upgrade from Sonnet 4.6 to Sonnet 5.0 mid-conversation could change the voice in the middle of someone's question. Pinning the conversational sessions to claude-sonnet-4-6 keeps the upgrade path explicit: Anthropic ships a new model, the dynamic picker shows it, the user picks Smart or Fast deliberately. Background loop gets the latest alias; foreground UI gets the pinned version.

Two changes in two weeks, both calibrated for it. April 20, 2026 shipped Fazm 2.4.0 with the dynamic picker that reads availableModels from the Claude agent SDK at runtime, which is what made absorbing Opus 4.7 on April 22 a non-event. The Gemini analyzer was already pinned to gemini-pro-latest before April, so the April 1 cutover happened on its own. Nothing about GPT-5 Turbo, DeepSeek V4, Gemma 4, or Llama 4 changed the Fazm code path because none of them run in the binary today. The architectural answer to those releases is a new bridge, and that has not been written yet.

Two places. For Gemini-handled background analysis, recordExternalLlmTrace at GeminiAnalysisService.swift line 263 writes the resolved model name (gemini-2.5-pro after April 1) into the local trace store with source fazm_observer. For Claude-handled chat queries, the log line at ChatProvider.swift line 2559 emits Chat query started (session=..., mode=..., model=..., cwd=...). Tail dev log from the Fazm menu, or read /tmp/fazm-dev.log directly, and you will see the exact model ID that handled each query. The model field is whatever the picker resolved at click time — claude-sonnet-4-6 for Fast, the current opus alias for Smart.