Every major 2026 AI model update, absorbed by a 3-line substring map

The constant that absorbed the year

Here is the actual code, unmodified, from the shipping app. Three rows. No release knobs. No per-model code path. The order field is the display order in the picker. The family field is the human-readable persona name. The substring is the only thing that matters at runtime.

When the Agent Client Protocol bridge advertises a new alias like claude-opus-4-7 in April, or claude-sonnet-4-6-20260203 in February, the matching logic at line 256 picks the first family whose substring appears in the ID. That alias is now routed into the right persona slot for the rest of the app's lifetime, until the bridge advertises something different. Fazm did not ship a Swift update for any of the seven Claude alias changes year to date.

How a new alias travels from Anthropic to the picker

The pipeline has three stops. Anthropic publishes the alias and the SDK announces it. The Node.js ACP subprocess pipes the announcement through the bridge as a models_available message. The Swift side substring-matches and rewrites the picker. No build, no release, no user action.

The 3-line ACP message handler

The models_available case in the bridge is literally three lines. The bridge converts the array of dicts into a typed tuple list and hands it to the Swift settings store. Everything downstream is the substring match and a sorted, atomic swap into a published property.

On the Swift side, ShortcutSettings.updateModels at line 251 walks the tuples, runs modelFamilyMap.first(where:) with the contains check, surfaces any bracket annotations like sonnet[1m] so context-variant aliases are distinguishable, sorts by family order, and replaces lastClaudeModels. Then recomputeAvailableModels() merges with the Codex and Gemini halves and the picker rerenders. No additional logic is needed for the SwiftUI picker to redraw because availableModels is a @Published property on an ObservableObject.

The year, month by month

The 2026 release wave is not random. There is a rhythm. Anthropic shipped two Claude families in the first half of the year. Google held Vertex GA for Gemini 3.1 Pro through the spring. OpenAI shipped GPT-5 Turbo and then the GPT-5.5 step in close succession. DeepSeek went from R2 to V4 Preview inside a single quarter. Open weights kept landing. Below is the picture from a shipping app's vantage point, where every one of these arrived as a runtime announcement, not a build trigger.

February: Anthropic moves first

Sonnet 4.6 landed early in the month and Opus 4.6 followed in the same window. Both were straight family rolls: same substring, new version. The picker did not need a code change. The label string is built dynamically from the family name and any bracket annotation in the ID, so claude-sonnet-4-6-20260203 rendered as Fast (Sonnet, latest) the first time a user opened the bar after the bridge picked it up.

On the Gemini side, the picker held two whitelist entries (gemini-flash-latest, gemini-pro-latest) which auto-rolled to Gemini 3.1 Pro on Vertex GA without code changes. The whitelist exists because un-versioned aliases are not in the bridge's curated probe response but resolve correctly when passed to session/set_model. Tested against gemini-cli 0.42.0 on 2026-05-21.

March: cadence accelerates

OpenAI's GPT-5.4 step landed in the first half of March, picked up via the Codex side of the bridge through codex_probe_result. The Codex name is used verbatim from the adapter response, so the picker showed exactly what the provider advertised. Anthropic continued Sonnet 4.6 iteration with point-version updates that the substring match absorbed without any user-visible churn.

By late March, the median time between frontier-class announcements across providers had compressed to under three weeks. The decision shifted from "which model do we hard-pin against" to "how do we keep the surface area of the model picker constant while the underlying list shuffles every two weeks". The substring family map is one answer; the per-provider probe path is the other half.

April: the densest month

April was the heaviest month of the year so far. On April 16, Anthropic released Claude Opus 4.7 across the API, Bedrock, Vertex, and Foundry. The benchmark deltas published in the announcement: SWE-bench Verified from 80.8% to 87.6%, Terminal-Bench 2.0 from 65.4% to 69.4%, GPQA Diamond from 91.3% to 94.2%, Finance Agent from 60.7% to 64.4%. Pricing held at $5 input and $25 output per million tokens; the tokenizer changed (1.0 to 1.35x more tokens for the same input) and high-resolution image input went from 1568px / 1.15MP up to 2576px / 3.75MP.

GPT-5 Turbo shipped with native multimodal in the same window, then the GPT-5.5 step. DeepSeek V4 Preview dropped roughly 24 hours after one of the OpenAI announcements: trillion-parameter, open weights, built on Huawei Ascend chips, Flash variant priced at $0.14 per million input tokens. Meta pushed Llama 4 Maverick and Scout into mainstream open weights, and Meta Superintelligence Labs (under Alexandr Wang) shipped Muse Spark. Mistral announced Large 3 with EU data-residency positioning. Microsoft shipped three foundational models covering text, voice, and image.

From the Fazm side, every one of these arrived through the same path. The bridge advertised the model. The substring match (for Claude) or the verbatim display name (for Codex and Gemini) ran. The picker rerendered. Zero Swift releases shipped that month for model wave reasons. The releases that did ship in April were for unrelated bug fixes and feature work.

May: the billing reshuffle

May was quieter on model releases but louder on billing. On May 6, 2026 Anthropic permanently doubled the 5-hour limits on Pro, Max, Team, and seat-based Enterprise plans. On May 14, 2026 Axios reported that Anthropic moved third-party agent tool usage to a separate monthly credit meter, so a Claude Code wrapper invoked through certain OAuth paths spends out of a different pool than direct Claude.com usage. The two changes pull in opposite directions: more headroom on the main plan, less crossover credit from third-party tooling.

Neither change required a Fazm release. The picker still asks the bridge for the current model list. The wrapper still uses the user's own Claude Pro or Max account. The new credit meter is upstream of the picker, not inside it. If a user hits the new third-party limit, the bridge surfaces a model_entitlement_missing message (parsed at ACPBridge.swift:1747) with the downgraded-to model, and the app handles it the same way it would any other entitlement issue.

Watch a fresh alias travel through the system

Imagine Anthropic ships claude-sonnet-4-7-20260901 at 9 a.m. one morning. Here is what happens inside Fazm the next time the app launches, end to end.

What is transferable

The 2026 release rhythm has changed the shape of what a shipping product needs to absorb. The lesson from a half year of bridge logs:

• Do not put the model list in the binary. Read it at runtime from whatever upstream actually knows.
• Match on family substrings, not exact aliases. The aliases churn every two to three weeks; the families churn once a year.
• Surface bracket annotations like [1m] in the display label. Context-variant aliases are real and need to be distinguishable.
• Use @Published (or your framework's equivalent) so the UI updates without an event bus.
• Keep the per-provider probe paths separate. Claude, Codex, and Gemini have different conventions and the merged list looks cleaner if they stay split until the last step.
• When pricing or entitlement changes (May 6 and May 14), let the upstream signal it. The picker does not need to know about credit meters.