The April 14-15, 2026 model releases, and the ten days of routing code a Mac app had to ship first

The window covered NVIDIA's Ising family of open AI models for quantum error correction (up to 2.5x faster, 3x more accurate decoding), OpenAI's GPT-5.4-Cyber variant released alongside the Trusted Access for Cyber program on April 14, NVIDIA's open-source models for next-gen autonomous driving (Alpamayo line), continued uptake of Llama 4 Maverick and OLMo 2 32B in the open-weights community, Codestral 2 (Apache 2.0) for local coding, and a batch of agent frameworks landing in Google ADK and Goose. Per llm-stats's running ledger, April 2026 is tracking as one of the densest open-source release months on record.

Because the news roundup is the SERP-crowded half of the story. The half nobody writes is: once GPT-5.4-Cyber or Llama 4 Maverick exists, what does it take to actually route to it from a Mac, fall back when it errors, and stream its response without UI flicker? Fazm shipped all three pieces of that plumbing in the ten days before April 14-15: Smart/Fast Opus/Sonnet toggle on April 4 (v2.0.6), ACP protocol v0.25.0 with credit-exhaustion and rate-limit error handling on April 7 (v2.1.2), and a Custom API Endpoint setting plus 600ms typing-indicator debounce on April 11 (v2.2.0). All three land inside /Users/matthewdi/fazm/CHANGELOG.json with an exact date and version.

It is the chat-header control added in v2.0.6 on April 4, 2026, that lets the user flip between Claude Opus (Smart) and Claude Sonnet (Fast) in one click, mid-conversation, without opening Settings. The CHANGELOG line reads: 'Added Smart/Fast model toggle in the chat header bar for quick switching between Opus and Sonnet.' The reason it matters for an April 14-15 news page is that the two tiers cost and respond differently per query, so routing becomes a first-class UI decision when you are juggling a frontier reasoning model and a fast conversational one for the same agent session.

v0.25.0 adds explicit error types for credit exhaustion and rate limits inside the Agent Client Protocol frames Fazm uses to talk to the Claude agent backend. Before that, the same conditions looked like generic stream errors, and onboarding would silently fail when credits ran out. The April 7, 2026 CHANGELOG entry for v2.1.2 makes this explicit: 'Upgraded ACP protocol to v0.25.0 with improved error handling for credit exhaustion and rate limits' and 'Fixed onboarding silently failing when credits run out, now shows an error with options to connect Claude or skip.' This is the error-handling half of a router: not just which model to pick, but what to do when the pick fails.

The line in the April 11, 2026 v2.2.0 CHANGELOG entry reads: 'Added custom API endpoint setting for proxies and corporate gateways.' In practice it sets the ANTHROPIC_BASE_URL that the chat agent speaks against, so a user or admin can route to a proxy, a corporate gateway, or an open-model backend that speaks the Anthropic API shape (vLLM, llama.cpp server, LiteLLM proxy, etc.). That is the hook that connects a page about April 14-15 open-source releases to the app on your Mac: the endpoint is the seam where a newly-released open-weight model plugs in.

CHANGELOG v2.2.0, April 11, 2026, third line: 'Fixed typing indicator flickering during network retries by adding a 600ms debounce.' It looks cosmetic, but it is a symptom of a streaming layer that has to tolerate mid-stream retries (which is what you get when the backend ACP frames hit a credit or rate-limit error and the client recovers). A 600ms debounce means 'wait for six 100ms-scale retry events to settle before flipping the indicator off and on again.' That is exactly the sort of thing you harden right before a week of model switching, not as decoration.

Auto-fallback to the built-in Fazm-hosted account. CHANGELOG v2.0.7, April 5, 2026: 'Fixed chat failing silently when personal Claude account lacks access to the selected model, now auto-falls back to built-in account.' So if you point Fazm at your own Anthropic API key and try to call a model your billing does not include, the chat quietly reroutes through the bundled account rather than leaving you staring at a silent failure. That is the third leg of the routing story: you do not only switch up and down the model ladder, you also fall over sideways when your credentials cannot reach the model.

No. Fazm's desktop control goes through native macOS accessibility APIs (AXUIElementCreateApplication, AXUIElementCopyAttributeValue) via the bundled mcp-server-macos-use binary, not screenshots piped into a vision model. The April 14-15 news cycle has a lot of 'computer-use' framing because GPT-5.4-Cyber and the new agent frameworks are partly aimed at that use case. Fazm's position in that space is opposite by design: any model you route to inherits a text-shaped accessibility-tree tool surface, not a pixel-coordinate one. The model changes, the input format does not.

Yes. Open /Users/matthewdi/fazm/CHANGELOG.json. The relevant entries are 2.0.6 (2026-04-04, Smart/Fast toggle), 2.0.7 (2026-04-05, auto-fallback to built-in account), 2.1.2 (2026-04-07, ACP v0.25.0, credit-exhaustion error handling), 2.1.3 (2026-04-09, Manage Subscription button, pop-out error surfacing), 2.2.0 (2026-04-11, Custom API Endpoint, 600ms debounce, pop-out shortcut), 2.2.1 (2026-04-12), and 2.3.2 (2026-04-16, tightened privacy language from 'nothing leaves your device' to 'local-first'). Seven numbered releases in thirteen calendar days. Every claim in this guide maps to a version and a date in that file.

CHANGELOG v2.3.2, April 16, 2026, first line: 'Tightened privacy language in onboarding and system prompts to accurately say "local-first" instead of "nothing leaves your device."' It is a quiet but meaningful edit: a router that can hit a hosted endpoint (GPT-5.4-Cyber, Claude Opus) cannot honestly claim nothing leaves the device. 'Local-first' is the honest frame: the default path is local, the model-routing layer can reach out when you opt into a hosted model, and the copy reflects that. For a news-cycle page about hosted-vs-open releases, this is the disclosure primitive that makes the rest of the story coherent.

Every other page lists releases (what shipped). This page lists routing infrastructure (what shipped on the Mac side to use it). The difference is concrete: this page names versions (v2.0.6, v2.1.2, v2.2.0, v2.3.2), dates (April 4, 7, 11, 16), exact CHANGELOG lines, environment variables (ANTHROPIC_BASE_URL), timing constants (600ms), and file paths (/Users/matthewdi/fazm/CHANGELOG.json). A reader who installs Fazm can verify every single one by opening Settings, opening the chat header, or opening the JSON file. That is the uncopyable half of the story.