Anthropic's new model in May 2026, and the one thing that decides whether your tool can use it

People typing this question are usually doing one of two things: checking whether they are running the latest Claude, or deciding whether to wait for something newer before they commit to a tool. Both are reasonable. But the more useful answer is not the model name. It is whether the app you run will even let you pick the next model when it lands, or whether you will be stuck on whatever string the developer hardcoded until they cut a release. That second question is the one this page actually answers, with code.

The recent Anthropic timeline, in plain dates

Here is the short, verifiable version of what changed leading into May 2026. Dates are from Anthropic's own posts and the GitHub changelog.

So if you arrived hoping for a fresh May drop to chase, there is not one for everyday use. The interesting question moves elsewhere.

The real question: can your tool use the next model on day one?

Most apps that wrap a Claude model ship with the model identifier baked into the binary. When Anthropic renames or releases something, that app cannot offer it until the maintainer notices, updates the string, and ships a build. You wait on their release cadence, not Anthropic's. This is exactly the friction the May 2026 searches are circling: the model exists, but my tool still shows me the old one.

Fazm is built the other way around. It wraps the real Claude Code agent loop over ACP (the Agent Client Protocol), and it asks that backend which models exist instead of deciding for you. When a new model becomes resolvable on your Claude account, it shows up in the picker without a Fazm update. The diagram below is the whole path.

The code that makes that true

This is not a marketing claim. The model list in Fazm's control bar is a published property that gets recomputed from the ACP response. The relevant method lives in Desktop/Sources/FloatingControlBar/ShortcutSettings.swift. It takes whatever models the backend advertises and merges them into the picker, falling back to a small static set only until the live list arrives:

/// Available models for Ask Fazm. Updated dynamically from ACP SDK;
/// falls back to defaults.
@Published var availableModels: [ModelOption] = ShortcutSettings.defaultModels

/// Update the Claude half of the model list from the ACP SDK response.
func updateModels(_ acpModels: [(modelId: String, name: String, description: String?)]) {
    guard !acpModels.isEmpty else { return }
    // ...map each advertised modelId into a picker entry...
    lastClaudeModels = newModels
    recomputeAvailableModels()
}

There are two details in that file worth calling out, because they are the reason a model transition does not break you:

• The default is a fallback, not a pin. The static defaultModels list exists only so the picker is not empty before ACP replies. The moment the backend reports its real list, recomputeAvailableModels() replaces it. Your visible options track the backend, not a frozen constant.
• Older pinned ids are migrated forward. A helper, normalizeModelId, maps the legacy default alias and prior-pinned Opus versions (the file literally cites claude-opus-4-5) onto the current canonical Opus id, so a stored preference from an older build still resolves to a real, current model instead of silently falling back to something else.

The same pattern is used for the other backends. The Gemini entries are deliberately the un-versioned gemini-flash-latest and gemini-pro-latest aliases, with an in-file note that this is so the picker auto-rolls when Google ships a new generation. The design philosophy is consistent: prefer the latest-resolving alias over a hardcoded version, so the tool ages gracefully instead of needing a release every time a vendor moves.

What Mythos does and does not mean for you

Because Mythos is the only genuinely new model in the May 2026 conversation, it is worth being precise. Mythos is a security-focused Claude that, per Anthropic's own preview, found thousands of real vulnerabilities during testing, including a 27-year-old OpenBSD bug and a 16-year-old FFmpeg flaw that prior tooling had missed. Anthropic explicitly said it does not plan to make the preview generally available, and on May 25, 2026 framed a future public release as conditional on safeguards that, by its own admission, no company has built yet.

For day-to-day coding, browser work, or desktop automation, none of that changes your options today. The model you should expect to be running is Opus 4.7. If and when a Mythos-class model reaches general release, a tool that reads its model list from the backend will surface it the same way it surfaced every Opus before it: as a new entry, when your account can resolve it, without you reinstalling anything.