AI model releases April 2026, read from 34 lines of routing code

Public data aggregators (LLM Stats, Releasebot, devFlokers, buildfastwithai) track seven frontier launches in the month. OpenAI shipped GPT-5.4 in Standard, Thinking, and Pro variants with a 1.05M token context window. Anthropic shipped Claude Opus 4.7 to general availability at Opus 4.6 pricing, plus a held-back internal model called Mythos. Google DeepMind shipped Gemini 3.1 Ultra and an open Gemma 4 family under Apache 2.0. Meta open-sourced Llama 4 Scout and Maverick, the latter a 400B MoE with a 10M token context window. xAI pushed Grok 4.20 with real-time web access. DeepSeek shipped V3.2 at roughly $0.28 per million tokens. This page does not repeat those specs; it covers the architecture a consumer Mac assistant needs so that all seven land in the app without a binary release.

The entire top 10 SERP for 'ai model releases april 2026' lists specs, benchmarks, and pricing. None of them describe what a consumer-facing desktop app has to do on its side when a new model drops. A Mac app that pins the string 'claude-opus-4-6' in compiled Swift exposes Opus 4.6 until the team ships a new notarized binary through Apple's process. A Mac app that pins 'gpt-5-turbo' in a dropdown stays on that model until someone cuts a release. Fazm's approach is to treat the model list as a stream, not a constant, so the Anthropic / ACP SDK can tell the running app about a new model and the user can use it before the code ever compiles.

emitModelsIfChanged at acp-bridge/src/index.ts:1271 reads the availableModels list from the ACP SDK every time a session opens, filters out the 'default' pseudo-model, and sends a 'models_available' message to the Swift app only when the serialized list actually changed. updateModels at Desktop/Sources/FloatingControlBar/ShortcutSettings.swift:178 receives that list and re-classifies each entry through the modelFamilyMap tuple: any modelId containing 'haiku' lands in slot 0 (Scary), 'sonnet' in slot 1 (Fast), 'opus' in slot 2 (Smart), anything else in slot 99 (unknown). session/set_model at acp-bridge/src/index.ts:1503 compares the user's requested model to the session's current model and, if they differ, calls the ACP endpoint to switch the model on the live session without creating a new one.

It makes unknown future models work. If Anthropic ships 'claude-haiku-5-0' tomorrow, the SDK will include it in availableModels. The updateModels function loops through the modelFamilyMap tuple and finds that 'haiku' is a substring of the new modelId. The new model automatically gets the label 'Scary (Haiku, latest)', lands at order 0 in the UI, and shows up in the floating bar's model picker the next time the user opens it. No code change. Same for a future 'claude-sonnet-5-0' or 'claude-opus-5-0'. The only time the substring match falls back to order 99 is for a genuinely unknown family, and even then the model is still selectable in the UI; it just doesn't get the three-letter label.

Because a user's conversation is context. If a user has been chatting with Fazm on Sonnet for half an hour and decides the next question needs Opus-level reasoning, the naive implementation is to kill the ACP session, start a new one with Opus, and lose the entire message history. Fazm avoids that by calling session/set_model on the existing sessionId. The ACP session stays alive, the message history stays intact, and the next reply is generated by the new model. Code path: handleQuery at acp-bridge/src/index.ts:1499-1509. The key line reads `await acpRequest('session/set_model', { sessionId, modelId: requestedModel })`.

Fazm uses the Claude Agent SDK via the ACP bridge, so at the time of writing the default path is Anthropic models (Haiku, Sonnet, Opus). The routing code does not encode any Anthropic-specific assumptions inside modelFamilyMap; the substring 'haiku' is only meaningful because Anthropic's naming convention uses family names in the modelId. A future change to route queries through a different SDK would extend the tuple, not rewrite the function. The architecture is the same: SDK streams the list, app re-sorts it, bridge calls set_model.

It doesn't. The model-routing code is about the cloud inference layer: which LLM generates the next reply. Fazm's accessibility-API architecture is about the local perception layer: how Fazm sees any app on the Mac via AXUIElement rather than by taking screenshots and OCR-ing them. Both layers are independent by design. A Claude model release touches the first layer and changes nothing about the second. A macOS update that breaks AX queries would touch the second and change nothing about the first. The April 2026 releases landed without a single AX call being modified.

Yes, and you should. The Fazm repository is public at github.com/mediar-ai/fazm. Run `git grep -n emitModelsIfChanged acp-bridge/src/index.ts` to find the function definition at line 1271. Run `git grep -n 'func updateModels' Desktop/Sources/FloatingControlBar/ShortcutSettings.swift` to land at line 178. Run `git grep -n 'session/set_model' acp-bridge/src/index.ts` to see all seven call sites. The interesting one for mid-conversation hot-swap is line 1503 inside the 'Reusing existing ACP session' branch. All line numbers are post-commit as of 2026-04-19.