The 2026 model updates also broke our microphone. We fixed it with 19 words.

The thing nobody writes about

You can ship a perfect model picker. You can wire up dynamic catalogs from the agent SDK so new Anthropic releases show up without a build. You can teach the floating bar to route Smart to Opus and Fast to Sonnet and Scary to Haiku. None of that helps if the user says “switch to Sonnet” into a microphone and the speech-to-text model returns “switch to sonnet” in lowercase, because the downstream parser was looking for the proper noun.

The 2026 LLM update cycle has been brisk enough that this is now a daily problem in production. Anthropic shipped Sonnet 4.6 as the everyday default. Opus 4.6 went GA, then Opus 4.7 followed weeks later. Haiku 4.5 is the Scary-button backing on most consumer Macs. Each of those names is, from the perspective of a 2024-era ASR model, a poem, a body part, a Japanese verse form, and a condiment. None of them is a model name yet. None of them is even a noun yet, on the priors that mattered when the speech model was trained.

The agent SDK fixed the model picker. Nothing fixes voice unless the agent fixes voice itself.

The fix is a 19-string array

At Desktop/Sources/DeletedTypeStubs.swift line 631 there is a Swift literal called systemVocabulary. It is a static let, so it is compiled into the binary at build time. There are 19 entries. Five of them are Anthropic-family LLM names. Two are protocols around those names. The rest are developer-adjacent product names that the team learned, by reading actual production voice queries, were getting heard as common English on the way in.

Five LLM names. Two protocols. Twelve developer-adjacent product names. The list is curated, not generated. Nova-3 will accept up to 500 keyterms per session, but the doc comment above the array, lines 629 and 630, is honest about the practical drop-off: effectiveness falls past about 30 terms. So the list stays at 19. The cost of adding a name is a build, a notarization, a Sparkle update. The benefit is that the next time someone says it, the microphone hears it correctly.

What goes on the wire

The plumbing is at Desktop/Sources/TranscriptionService.swift. Line 94 sets private let model = "nova-3". Lines 294 through 296 are a three-line for-loop that emits one URL query item per term:

// Add keyterm parameters for custom vocabulary (Nova-3 uses "keyterm" not "keywords")
for term in vocabulary {
  queryItems.append(URLQueryItem(name: "keyterm", value: term))
}

The comment is doing real work. The older Deepgram keywords parameter accepted boost weights and could hallucinate the term if you pushed too hard. Nova-3 keyterm is designed to be safe at high recall: the model treats the listed strings as in-vocabulary for the session and false-positive risk stays low because the strings are not common English. The 19 query items get appended to the WebSocket URL on every voice session, every push-to-talk press, every dictation toggle. There is no per-utterance prompt. The bias is at the session level.

The five mishears the list is built to stop

The internal motivation for shipping this list, recorded in the 2026-04-28 changelog entry for version 2.6.2, is that “product names and tech terms (Claude, Sonnet, Opus, MCP, Tauri, Supabase, etc.)” were coming back wrong. Five of those failure modes are below. None of them is hypothetical. They were observed in actual voice queries during the run-up to the release.

The fix is not magic. It is the same ASR model, the same audio, the same network round-trip. The only delta is 19 query items.

Two layers, two release cadences

Fazm absorbs the 2026 LLM update cycle on two surfaces, and they have different update rules. The picker is dynamic. The voice vocabulary is static. Both deserve to exist.

The picker reads from a @Published array fed at runtime by the Claude agent SDK. New Anthropic models show up without a Fazm build, get substring-matched against a four-row family map, and route to one of three labels. That is the dynamic layer. New names absorb the second they appear in the SDK.

The vocabulary is a Swift literal in source, compiled into the binary, shipped as part of a notarized release. New names absorb on the next point release. That is the static layer. The two layers do not contradict each other; they cover different costs. Adding a model to the picker is free for the Fazm side, and we want it to be free, because Anthropic ships fast and the user should not wait. Adding a name to the vocabulary is the cost of a release, and we accept that cost, because the wrong place to make a speech model decide what is a word is at request time, and the right place is at session-open time.

The user can override

The Settings panel at SettingsPage.swift renders each built-in term with a small X button next to it. Click the X and the term is appended to a Set called disabledSystemVocabulary, which is persisted across launches. The effective list shown to Nova-3 is user-added terms first, then any active system terms not already present, deduplicated case-insensitively. A Restore built-in defaults button clears the disabled set in one click. The 2026-04-28 release notes for version 2.6.2 list this UI as a deliberate ship: “Added a Vocabulary section to Settings to view, disable, and restore the built-in transcription terms.”

A user can also add their own. The Dictionary section accepts a free-text term and appends to a published array called transcriptionVocabulary. That is the path for company-specific words: customer names, internal codenames, product lines that nobody outside the building has ever heard. The session-level keyterm cap is 500, so there is room.

Why this matters more in 2026 than it did in 2024

In 2024, the LLM you were saying out loud was almost always called GPT, with a number after it. There were maybe three product names that mattered for voice. The Whisper-class ASRs of the time hit them well enough on common pronunciations that the keyterm trick was a niche concern.

In 2026, the active vocabulary that a serious user touches in a normal week includes Claude, Sonnet, Opus, Haiku, GPT, the OpenAI numbered series, DeepSeek, Gemini, Llama, Mistral, plus MCP, ACP, and a half-dozen agent-protocol acronyms. Anthropic alone has shipped four named families with version numbers attached. The combinatorial space is larger, the confidence interval on any single ASR pass is lower, and the failure mode is no longer rare. It is the median voice query for an LLM-using developer.

The right time to teach a speech model about Sonnet 4.6 is the day Anthropic announces Sonnet 4.6. The wrong time is six months later, when the next version is already named Sonnet 4.7 and the original mistake has been quietly mistranscribed into a thousand chat logs.