The LLM update that mattered in April 2026 was not a benchmark. It was a rate-limit tier.

The month was dense. On the closed-source side: Claude Opus 4.7 went GA with the same $5 / $25 per MTok pricing as 4.6, the redesigned Claude Code desktop app and Routines research preview landed on April 14, and Anthropic disclosed the internal Mythos model under the Project Glasswing controlled rollout on April 16. GPT-5.2 and Gemini updates continued their monthly cadence. On the open-weights side: Hugging Face saw heavy traction on new Qwen, DeepSeek, and Gemma releases, and GLM-5.1 variants kept arriving. But the operational story, the one that changes behavior inside consumer apps, is smaller and not on any roundup: Anthropic enforced a per-model seven_day_opus weekly rate limit that made Opus unworkable as a default in a desktop agent that runs multiple sessions in parallel.

Anthropic's API has four rate-limit type labels that Fazm's bridge recognizes (ChatProvider.swift lines 537 to 543): five_hour (session limit), seven_day (general weekly limit), seven_day_opus (Opus-specific weekly limit), and seven_day_sonnet (Sonnet-specific weekly limit). The seven_day_opus bucket is separate from the general seven_day bucket. In practice, a Pro-plan user who picks Opus as their default in a desktop agent like Fazm, which runs a main session, a floating-bar session, and a background observer session in parallel, can burn the Opus weekly quota in under a day. That turns the default 'Smart' tier into a brick for the rest of the week.

Two things. First, on April 4 in version 2.0.6, Fazm added a Smart / Fast model toggle in the chat header bar so users could switch between Opus and Sonnet without leaving the conversation (from CHANGELOG.json). Second, a one-time migration was written into the app launch path that silently moved every existing 'Smart (Opus)' user over to 'Fast (Sonnet)' the next time they opened the app. The migration lives at lines 287 to 296 of /Users/matthewdi/fazm/Desktop/Sources/FloatingControlBar/ShortcutSettings.swift, gated by a UserDefaults flag named shortcut_didMigrateFromOpus. The inline comment reads: 'One-time migration: switch existing Opus users to Sonnet (Opus burns through rate limits too fast).' That sentence is the whole April 2026 LLM update story compressed into a code comment.

No. Opus still scores higher on hard reasoning and long-horizon planning. The decision is a product decision, not a capability decision. For a consumer desktop agent that runs continuously and multiplexes many sessions, a model that hits a weekly wall is a worse default than a model that does not. Sonnet 4.6 is strong enough that most users will not notice the downgrade during normal use, and the 'Smart' tier is still one click away when they need it. The migration preserves user choice for anyone who explicitly re-picks Opus.

The ShortcutSettings.swift updateModels(_:) function at lines 168 to 189 pulls a dynamic model list from the ACP (Agent Client Protocol) SDK response on startup. When a new Claude model ships, the list refreshes at the next app launch. A substring matcher at lines 158 to 162 maps known family substrings ('haiku', 'sonnet', 'opus') to friendly short labels ('Scary', 'Fast', 'Smart'). Unknown families fall back to the raw API display name (with 'Claude ' stripped). This is why Fazm did not need a release to pick up Opus 4.7 or any subsequent point release; the picker simply rebuilt itself.

Version 2.2.0 on April 11, 2026 added 'Custom API endpoint setting for proxies and corporate gateways' (CHANGELOG.json line 29 of the v2.2.0 entry). This exposes the ANTHROPIC_BASE_URL environment variable to the user so they can route Claude traffic through a corporate proxy, a self-hosted gateway like LiteLLM, or a region-specific endpoint. The addition was demand-driven: enterprise rollouts in early April kept hitting corporate egress policies that blocked direct calls to api.anthropic.com. The feature is orthogonal to model choice but became table stakes for any app expecting to be adopted inside a company after the April update cycle.

The in-app picker defaults to three Claude tiers (Haiku 4.5, Sonnet 4.6, Opus 4.6) but the updateModels function accepts whatever the ACP SDK reports at runtime. Beyond Claude, the app delegates to skills in ~/.claude/skills and to MCP servers that can themselves target other models, so the effective set the app can talk to is broader. For direct chat, the picker is Claude-only. For programmatic automation via skills, any model that exposes a compatible API can be wired up.

ChatProvider.swift lines 537 to 543. The five_hour label is a short-window session limit that resets every five hours, typical of Claude Pro. The seven_day label is the general weekly usage limit. The seven_day_opus label is Opus-specific and the one most users hit in April 2026 because Opus is more expensive per call. The seven_day_sonnet label is Sonnet-specific and rarer in practice. The label determines the user-facing message. A seven_day_opus hit surfaces as 'Opus weekly limit, resets <date>' with a suggestion to switch to Sonnet. The app does not silently switch models on a rate limit; it tells the user and lets them choose.

The specific Swift source lines apply to Fazm's Claude integration. The general pattern, that a consumer app's choice of default model is dictated by rate-limit economics and not by benchmark rank, applies across providers. OpenAI has its own per-model rate limits that shift with each release. Gemini has region-specific quotas. Open-weights models on Hugging Face shift the same problem onto the user's GPU. The anchor fact in this guide is Claude-specific, but the lesson is provider-agnostic.

Yes. Open /Users/matthewdi/fazm/Desktop/Sources/FloatingControlBar/ShortcutSettings.swift. Lines 150 to 156 declare the three default models. Lines 158 to 162 hold the shortLabelMap for friendly 'Scary / Fast / Smart' labels. Lines 168 to 189 are updateModels(_:), the dynamic refresh. Lines 287 to 296 are the Opus to Sonnet one-time migration, gated by shortcut_didMigrateFromOpus. Line 290 is the match condition (savedModel == 'claude-opus-4-6' && !didMigrateOpus). Lines 291 to 293 are the swap. Every claim in this guide maps to one of those line ranges.

Worse, in fact. Screenshot-based agents send every frame of a decision as a pixel array to a vision model. Each vision call consumes tokens at a much higher rate than text, so a screenshot-based agent on Opus burns the Opus weekly quota faster than any text-based app. Fazm uses macOS accessibility APIs instead of screenshots: the app reads the actual UI element tree of any running Mac application as structured text, which costs a fraction of the tokens and means a cheaper model like Sonnet 4.6 can make accurate decisions. This is the second reason the Opus to Sonnet migration was safe in April 2026: the pipeline was already designed to work with a mid-tier model.

Open the Fazm desktop source tree. The shortcut_didMigrateFromOpus flag is defined and read in ShortcutSettings.swift lines 287 to 296. The rate-limit type labels are in ChatProvider.swift lines 537 to 543. The v2.0.6 changelog entry dated 2026-04-04 mentions the 'Smart/Fast model toggle in the chat header bar for quick switching between Opus and Sonnet.' The v2.2.0 changelog entry dated 2026-04-11 mentions the 'Custom API endpoint setting for proxies and corporate gateways.' Those two changelog lines and the Swift migration line are the three atoms that compose the April 2026 LLM-update story as seen from inside a shipping consumer app.