What Anthropic shipped in April 2026 and what actually changes for a real Mac agent
Opus 4.7 went GA. Mythos got a codename and a gated program. Claude Design launched. The top results tell you what shipped. This page translates that list into what a consumer desktop agent running on macOS actually picks up, and what it safely ignores. Written from inside one such agent's source tree.
THE SHORTLIST
What actually dropped
Three releases you will see cited across the April 2026 coverage, with one line each on what they are and are not.
Claude Opus 4.7, generally available
GA upgrade at the same $5 / $25 per MTok pricing as Opus 4.6. Anthropic frames the gains as stronger software engineering, better vision at higher resolution, sharper instruction following, and more reliable long-running agent work. No protocol or SDK shape change.
Claude Mythos, via Project Glasswing
Internal codename Capybara. Described by Anthropic as a step change above Opus 4.6, with unprecedented accuracy at identifying and exploiting software vulnerabilities. That capability is exactly why it is not shipping broadly. Glasswing limits access to roughly 50 partner organizations with no public release date.
Claude Design, new Labs tool
Launched inside Claude Apps for creating designs, prototypes, slides, and one-pagers. This is a product surface, not a model release. It does not change the API, the model list, or the rate-limit machinery your client sees.
THE REFRAME
Now translate that for a desktop agent
Every release note reads the same: better vision, better tool use, better instruction following. The bundle sounds monolithic, but those three gains hit different classes of agent very differently.
A Computer Use style agent, the one that screenshots the screen each turn and asks the model to OCR it, spends most of its tokens on the picture and most of its reliability budget on the model picking the right pixel. That agent gets a real uplift from Opus 4.7's higher-resolution vision.
A macOS-native agent that reads the accessibility tree first and only screenshots when it needs to sees the world as structured text already. The window, its children, their labels, their roles, their bounds. That agent's bottleneck is not vision, it is tool selection and long-horizon coherence. Which is where Opus 4.7's other two gains land, and why an agent like this should upgrade even though the splashy bullet in the release notes does not describe its hot path.
How a Mac agent actually moves a bit through the April 2026 stack
THE ANCHOR FACT
What this looks like inside the source
Everything above is abstract until you look at two files. Here is the accessibility read that powers the agent's view of the frontmost app, lifted verbatim from the desktop source at AppState.swift lines 439 through 441:
And here is the single line that tells you which Claude model Fazm's ACP bridge routes traffic to by default, from the Node.js side of the subprocess at acp-bridge/src/index.ts:1245:
Two things to notice. One, the state read does not start with a screenshot. It starts with a native accessibility call that returns structured data the model can reason over without vision. Two, the bridge picks its default in a single line. Swapping to Opus 4.7 for a user, or for a class of long-horizon tasks, is a constant change and a restart of the subprocess, not an architectural migration. That is why Opus 4.7 GA slots in cleanly.
“To change the model the entire agent talks to.”
acp-bridge/src/index.ts:1245
WHAT THE STATE CALL RETURNS
AX tree vs screenshot, same moment in time
To make the vision-vs-structure tradeoff concrete: this is roughly what a Mac agent gets back when it probes the focused window of the frontmost app with the accessibility call above, compared to what a screenshot-first agent ships to the model as a base64 image.
A screenshot-first agent has to do OCR and layout inference against what is essentially a picture. A structured-first agent already has the window role, the children, the labels, and the text content before any token is spent on vision. Opus 4.7's higher-resolution vision makes the first path cheaper and more reliable, which is real progress. But it does not close the gap with a well-built accessibility-first agent on tasks the AX tree already answers.
What each kind of agent picks up from Opus 4.7
Same release, two different impact profiles.
| Feature | Screenshot-first Computer Use | AX-first Mac agent |
|---|---|---|
| Higher-resolution vision | Big win. Fewer misreads on small UI text, truer pixel targeting. | Nice to have. Only matters on the rare tasks a screenshot is used for. |
| Stronger instruction following | Helps, but noise from OCR is still the dominant error mode. | Direct win. Tool selection and long-run coherence are the hot path. |
| More reliable long-running agent work | Helpful, bounded by how much state the screenshot can convey. | Biggest win. Structured tool output compounds well across turns. |
| Same $5 / $25 per MTok pricing | True of both paths. | True of both paths. |
| Image tokens per turn | Up, because the screenshot is the ground truth. | Near zero on structured tasks. Image tokens only when visual verification is genuinely needed. |
| Works with non-browser Mac apps | Yes via screenshot, but layout inference is lossy. | Yes via AX tree and native MCP binaries (macos-use, whatsapp-mcp). |
The three things a Mac agent should do this week
Expose Opus 4.7 in your model picker
Do not make it the default if Sonnet is faster for your workload, but make it selectable for long-horizon tasks where tool-use reliability is the bottleneck. One constant change in one file.
Keep reading structured state first
AX tree on Mac, DOM for web, native tool JSON everywhere else. Vision uplifts are real but expensive, and Anthropic's roadmap rewards agents that use structure when it exists.
Do not plan for Mythos
It is Glasswing-only and restricted by design. Build on the GA ceiling (Opus 4.7), not the preview. If Anthropic ever opens Mythos, you will get a week's notice like everyone else.
Separate vision tokens from reasoning tokens
Log how often each run actually needs the screenshot path vs the AX path. That's the ratio that tells you how much of Opus 4.7's vision uplift you will realize.
Leave room for a mid-release pricing surprise
Opus 4.7 shipped at Opus 4.6 pricing. That is a strategic signal. Plan your extra-usage caps assuming future Opus refreshes stay at the same MTok price until Anthropic says otherwise.
Why this matters if you use a Mac agent every day
The honest version of the story is: Anthropic shipped a lot in April 2026, and most of the splash was aimed at Computer Use, at the browser, at the general developer. If you run a Mac-native consumer agent that speaks to Finder, Notes, Mail, WhatsApp, Settings, and the half dozen other apps you live in, the headline improvements are not the ones that change your day. The quieter ones, better instruction following and more reliable long-running tool use, are. And the reason those quieter ones land so cleanly is that the agent already does the unglamorous work of reading the accessibility tree instead of screenshotting everything.
Want to see Opus 4.7 driving real Mac apps, not a browser demo?
Fifteen minutes. We open Fazm on a Mac, flip the model, and drive a real task through the accessibility tree while you watch.
Book a call →Frequently asked questions
What did Anthropic release in April 2026?
Three things matter for people shipping apps on top of Claude. First, Claude Opus 4.7 went generally available at the same $5 / $25 per MTok pricing as Opus 4.6, with stronger software engineering, better vision, sharper instruction following, and more reliable long-running agent work. Second, Claude Mythos (internal codename Capybara) was unveiled, but only through Project Glasswing, a gated early-access program limited to around 50 partner organizations, no public release date. Third, Claude Design launched inside Claude Apps as a Labs tool for creating designs, prototypes, slides, and one-pagers.
Is Claude Mythos available to the public?
No. Mythos is gated behind Project Glasswing and is distributed only to a small set of partner orgs Anthropic has selected. Per Anthropic's own description, Mythos can identify and exploit software vulnerabilities with unprecedented accuracy, which is why the rollout is so restricted. If your product strategy depends on shipping on top of Mythos today, that is not available. Plan for Opus 4.7 as the current ceiling for GA access.
Does Opus 4.7's improved vision matter for a macOS agent?
Less than people assume. Opus 4.7 sees images in greater resolution, which clearly helps a Computer Use style agent that screenshots the screen and asks the model to OCR it. But a Mac-native agent that reads the focused window via AXUIElementCreateApplication and AXUIElementCopyAttributeValue(appElement, kAXFocusedWindowAttribute, ...) gets app state as structured text already. For that class of agent, the bigger uplift from Opus 4.7 is sharper instruction following and more reliable long-running tool use, not the vision improvement.
How does Fazm use Claude under the hood?
Fazm runs a long-lived Node.js subprocess called the ACP bridge (acp-bridge/src/index.ts) that speaks Anthropic's Agent Client Protocol and manages the Claude API connection. The Swift desktop app (Desktop/Sources/Chat/ACPBridge.swift) talks to the bridge over stdin and stdout as JSON lines. The bridge's DEFAULT_MODEL constant is set on a single line at acp-bridge/src/index.ts:1245, and the Swift layer handles permissions, native UI, and the accessibility tree reads that give Claude a structured view of whatever app is frontmost.
What does Fazm read from macOS when you ask it to do something in an app?
For state, it creates an AX element for the frontmost app with AXUIElementCreateApplication(frontApp.processIdentifier), then calls AXUIElementCopyAttributeValue(appElement, kAXFocusedWindowAttribute as CFString, &focusedWindow) to get the focused window. This is at AppState.swift lines 439 through 441. It only falls back to CGWindowListCreateImage based screenshots when the AI actually needs a pixel rendering, for example to verify a visual state an AX tree cannot describe. That hybrid is why Opus 4.7's vision gain is a nice-to-have rather than a step change for a well-built Mac agent.
Is Fazm running on Claude Opus 4.7?
The bridge picks a default model in one place, acp-bridge/src/index.ts:1245. At the time of this post Fazm ships with Sonnet 4.6 as the default because it is the best speed-to-quality tradeoff for short-turn desktop work. Opus 4.7 is available as a model the user can pick in the model list, and it is the right choice for the long-horizon agent runs that benefit most from Opus's stronger tool-use reliability and instruction following.
Why does Fazm work with any Mac app and not just the browser?
Because it does not depend on a browser extension or a DOM. Fazm exposes a set of native macOS MCP tools. macos-use (binary mcp-server-macos-use inside Contents/MacOS) drives Finder, Settings, Mail, and any other AX-compliant app via the accessibility tree. whatsapp-mcp drives the native WhatsApp Catalyst app. Playwright is used only when the task is actually in the browser. The tool-routing section of the desktop system prompt is explicit: capture_screenshot for the full desktop, macos-use tools for native apps, playwright only for web pages.
What is Project Glasswing?
Project Glasswing is Anthropic's gated early-access program for Mythos. According to the reporting around it, participants are a small number of partner organizations, not the general public. Anthropic has said publicly that Mythos was too capable at certain offensive security tasks to ship broadly, and Glasswing is the mechanism for letting select partners evaluate it without a full GA rollout.
What is Claude Design?
Claude Design is a new Labs tool inside Claude Apps for creating designs, prototypes, slides, and one-pagers with Claude. It is separate from the model line. Think of it as a product surface rather than a model release. It does not change the API, the SDK, or the rate-limit shape that third-party clients see.
If I am building on Claude today, what should I actually do this week?
Three things. First, set Opus 4.7 as an option in your model picker and let users flip to it for long-running agent tasks where tool-use reliability matters more than per-token cost. Second, do not rebuild around vision just because Opus 4.7 sees better, structured state (accessibility tree, DOM, JSON tool output) is still cheaper and more deterministic. Third, do not make product plans around Mythos. It is Glasswing-only and you will not be on that list.
Can I use Opus 4.7 in a Claude Pro or Max subscription?
Yes. Opus 4.7 is available on paid Claude plans and via the API at the standard Opus pricing. If you exhaust your plan's included usage, you move to extra usage (pay-as-you-go) at API rates, the same rate_limit_event machinery described in Fazm's other guide applies unchanged. No shape change to the protocol for this release.