LLM updates and announcements in 2026, read through three lines of code a desktop wrapper had to touch

The three places in one wrapper’s source where every 2026 LLM update lands

A wrapper on top of Claude Code and Codex sits in front of every model announcement of the year. The interesting question is not which model wins on a leaderboard. It is which ones forced the wrapper to ship a new build. After tracking the 2026 announcements against the repo, the answer is that most did not, because the variability is parked in three places that already cover the common cases.

1. Place 1

   A one-line regex that decides Claude vs Codex routing

   The very first thing the bridge does on a new query is check whether the requested model ID matches a regex. If it matches, the call is routed to the codex-acp adapter. If not, the call falls through to the @agentclientprotocol/claude-agent-acp adapter. The whole branch is one line of source, and it covers any future gpt-, codex-, or o-prefixed model without a code change.

2. Place 2

   An AppStorage key that points the SDK at any Anthropic-compatible gateway

   A single TextField in the settings page binds to customApiEndpoint at Desktop/Sources/MainWindow/Pages/SettingsPage.swift line 885. When the user enters a URL there, the bridge sets the ANTHROPIC_BASE_URL env var on the spawned ACP subprocess at Desktop/Sources/Chat/ACPBridge.swift lines 379 to 382. The SDK respects the env var, so every /v1/messages call from the agent then targets the user’s gateway instead of api.anthropic.com. That gateway can be Bedrock, Vertex, GitHub Copilot, or a local bridge that translates Anthropic-shaped calls to DeepSeek V4, Qwen 3.5-Omni, or Gemma 4.

3. Place 3

   A dynamic models emitter that picks up new IDs from the ACP SDK

   The picker in the chat header is not driven by a hard-coded array. The function emitModelsIfChanged in acp-bridge/src/index.ts around line 2362 reads availableModels from the SDK’s session/new response, filters out the “default” pseudo-model, and emits a models_available message to the Swift host. When Anthropic adds an ID server-side (Opus 4.7 on Apr 16 is a clean example), the next session opens with the new model already in the dropdown, with no app-side release.

The regex, verbatim

This is the entire branch point. If a 2026 announcement of a new gpt-, codex-, or o-prefixed model lands, the wrapper handles it without touching any other file.

What each 2026 announcement looked like on the wrapper side

The list below covers the events that either moved the agent loop a wrapper ships against, or moved the lane a wrapper sits in. Capital announcements are included because they shape the supply side and the gateway story (which is now first-class in Claude Code). Executive interviews and fundraising leaks are not on the list.

1. Feb 5, 2026Anthropic, Model GA

   Claude Opus 4.6 generally available

   First major release of the year. $5 / $25 per Mtok cost line, the same one Opus 4.7 inherited in April. For a wrapper reading availableModels from the SDK, the new ID appeared in the picker on the next session/new. For a wrapper with hard-coded defaults, it was a one-line bump.

   Cost on the wrapper side: no code release if availableModels is dynamic; one-line bump otherwise

2. Feb 17, 2026Anthropic, Model GA

   Claude Sonnet 4.6 generally available

   Became the day-to-day default for tool-heavy agent loops on a Pro or Max subscription. Most desktop wrappers still pin this as their everyday default in May. In the Fazm source it shows up as the string literal claude-sonnet-4-6 in acp-bridge/src/index.ts line 2336 and on the Swift side in the session config at Desktop/Sources/Providers/ChatProvider.swift.

   Cost on the wrapper side: one default-model constant change; the picker auto-detects the new ID

3. Apr 7, 2026Anthropic, Preview + SDK

   Mythos preview via Project Glasswing, ACP SDK 0.25.0

   Anthropic announced a gated cybersecurity preview of Claude Mythos for 11 partner orgs (Amazon, Apple, Microsoft, Nvidia named), with preview pricing of $25 / $125 per Mtok. Same day, @agentclientprotocol/claude-agent-acp shipped 0.25.0, the first of three SDK bumps in two weeks. For desktop wrappers Mythos is a watch-list item; the SDK bump is what shipped that day.

   Cost on the wrapper side: SDK bump release; Mythos itself stays off the picker

4. Apr 15, 2026Anthropic, Product

   Redesigned Claude Code desktop app

   Anthropic shipped a redesigned Claude Code desktop app with a parallel-sessions sidebar, drag-and-drop workspace layout, integrated terminal and file editor, faster diffs, expanded previews, SSH support on Mac, and full plugin parity with the CLI. Anthropic now competes with every host that wrapped Claude Code for the same workflow. The wrappers stop being a coverage play and start being a reach play.

   Cost on the wrapper side: no SDK release; a positioning rewrite on every wrapper landing page

5. Apr 16, 2026Anthropic, Model GA

   Claude Opus 4.7 generally available

   Opus 4.7 went GA at the same $5 / $25 cost line as Opus 4.6, with better long-run agentic behavior and high-resolution image understanding. For a wrapper that reads available models from session/new (which is the May default once ANTHROPIC_BASE_URL is set), the new ID showed up in the dropdown automatically. For a wrapper with hard-coded defaults, this was a one-line bump.

   Cost on the wrapper side: no code release for the dynamic picker; one-line bump for hard-coded defaults

6. Apr 20, 2026Anthropic, Capital + SDK

   Amazon $5B + up to $20B, $100B AWS commitment, ACP SDK 0.29.2

   Amazon committed another $5B of equity with up to $20B more tied to commercial milestones, paired with a $100B Anthropic commitment to AWS over ten years and up to 5GW of new capacity to train and run Claude. Same day, @agentclientprotocol/claude-agent-acp shipped 0.29.2. The capital number is the headline; the SDK bump is what wrappers actually shipped against.

   Cost on the wrapper side: SDK bump release; gateway-routed defaults start landing in the changelog

7. Apr 24, 2026Anthropic, Capital

   Google up to $40B, 5GW dedicated TPU

   Google committed up to $40B in cash and compute on the same week, lifting the post-money valuation to roughly $350B and reserving 5GW of dedicated TPU. Combined with the Amazon deal, that places roughly 10GW of training and inference power on the supply side under one tenant. The visible end of this for a Mac developer is the slow normalization of gateway routing as the default path.

   Cost on the wrapper side: no immediate release; informs the gateway-first defaults shipping through May

8. Spring 2026OpenAI, Model

   GPT-6 GA and GPT-5.5 agentic preview

   OpenAI shipped GPT-6 as the next mainline model and previewed GPT-5.5 as the agentic step. Both IDs start with gpt-, so they match the CODEX_MODEL_PATTERN regex at acp-bridge/src/codex-query.ts line 67. A user selecting either model from the picker routes through the @zed-industries/codex-acp adapter (pinned at ^0.14.0) rather than the Claude-ACP path. The wrapper does not need a code change.

   Cost on the wrapper side: no wrapper release; the regex already covers the prefix

9. April 2026Open weights, Models

   DeepSeek V4, Qwen 3.5-Omni, Gemma 4, Muse Spark

   Inside one month, the open-weights and challenger side stacked four heavyweight releases. DeepSeek V4 Flash and V4 Pro at a reported ~$5.2M training spend; Alibaba's Qwen 3.5-Omni accepting 10+ hours of audio and 400 seconds of 720p video per turn; Google's Gemma 4 family under Apache 2.0, 31B Dense leading on intelligence-per-parameter; Meta's Muse Spark out of Alexandr Wang's Superintelligence Labs as the first proprietary frontier model from that group.

   Cost on the wrapper side: no wrapper release; reachable through the customApiEndpoint TextField + a user-run bridge

10. Late Apr - May, 2026Anthropic, Product surface

    Claude Code env vars, slash commands, picker

    A steady drip of small developer-surface changes. ANTHROPIC_BEDROCK_SERVICE_TIER (default, flex, priority) for Bedrock tiering. ANTHROPIC_WORKSPACE_ID for workload identity federation. CLAUDE_CODE_PLUGIN_PREFER_HTTPS to clone plugin sources over HTTPS. /tui to switch to flicker-free rendering, claude project purge to delete all Claude Code state for a project, /resume now finds a session by pasting a PR URL into its search box. None of these are model news; all of them touch the daily workflow.

    Cost on the wrapper side: no wrapper release; the env vars are passed through to the spawned subprocess

11. Mid May, 2026Anthropic, SDK

    @agentclientprotocol/claude-agent-acp at the 0.33 line

    By mid May, current Mac wrappers on top of Claude Code are pinned to the 0.33 line of the ACP package. In the Fazm repo this is the version declared in acp-bridge/package.json: "@agentclientprotocol/claude-agent-acp": "^0.33.1". The session/fork RPC, the rate_limit_event forwarding, and the compactBoundary status updates are all surfaces a host has direct access to at this version. The official desktop app does not expose any of these as end-user buttons yet, even though they are in the same SDK.

    Cost on the wrapper side: SDK bump release; fork button and rate-limit forwarding wired end-to-end

The 2026 model surface, grouped by vendor

Three families of models accumulated through the first five months of 2026. Anthropic stayed on the Opus / Sonnet / Haiku / Mythos cadence. OpenAI moved the gpt- and o- family forward. Open weights pulled within reach of the closed leaders on specific axes (intelligence-per-parameter for Gemma 4, multimodal capacity for Qwen 3.5-Omni, training cost efficiency for DeepSeek V4). Every entry below is a model the wrapper can route to today through the three places above.

How a new model ID actually reaches the chat header

The flow below is what happens when Anthropic or any compatible gateway publishes a new model ID server-side. The wrapper never restarts, and no app release is needed. The picker simply has a new entry the next time a session is opened.

The one announcement that actually moved the lane

Reading the year by impact on a wrapper builder, the redesigned Claude Code desktop app on Apr 15 is the one that mattered most. Before it, hosts on top of Claude Code (Zed, Fazm, hand-rolled wrappers) were filling the UI gap Anthropic had not addressed itself: session management, multi-window state, a chat surface for a CLI tool, a way to run several agents at once. After it, Anthropic was in the same lane, and the wrappers had to be clearer about what they did that the official app did not.

The honest answer is that the official desktop app and an ACP-based wrapper aim at different things. The official app is the cleanest first choice for terminal-shaped Claude Code work: parallel runs, plugins, an integrated editor, SSH on Mac, file diffs, streaming responses. A wrapper takes the same agent loop and runs it outside the developer workflow, through macOS accessibility APIs rather than screenshots, so the same agent that just touched your code keeps going in the browser, in Mail, in Calendar, in Numbers. That is reach the official app does not aim for. The 2026 announcement stack made the boundary between those two answers sharper, not the other way around.

What I am watching through the rest of 2026

• Mythos general availability and how the price drops from preview ($25 / $125 per Mtok) once the gated cohort expands. Even at half that, it stays a high-value run only.
• How the ACP SDK shape changes once gateway-routed Anthropic-compatible endpoints become the default expectation rather than a side door. The 0.33 line is already further down that path than 0.25.0 was.
• Whether OpenAI ships its own agentic protocol parity with ACP, or whether the codex-acp adapter stays the cleanest in-process bridge for gpt-, codex-, and o-models on Mac.
• How Qwen 3.5-Omni’s multimodal capacity gets exposed through Anthropic-shaped bridges. Most of the open-weights bridges in May 2026 still drop the audio and video channels when translating to /v1/messages.
• Whether the official Claude Code desktop app ever exposes the SDK’s session/fork RPC or compactBoundary status events as end-user buttons. As of May 2026 it does not, even though they have been in the SDK since spring.