New AI model releases and announcements, May 24-25 2026

The gap every recap leaves open

Read the week's coverage and you will know that Gemini 3.5 Flash is fast and that Chinese open-weights models now move more tokens through OpenRouter than the US frontier labs. What none of it tells you is the thing you would actually do next: try one of these in the agent you already use, today, without throwing away your current session.

Here is the friction. The bare Claude Code CLI talks to Anthropic. Two of the three headline models this week (Gemini 3.5 Flash and the Chinese open-weights families) are not Anthropic models, so the CLI cannot reach them. The usual answer is to go install a second tool, wire up a second config, and lose whatever context you had built in the first one. That is the part the leaderboards never mention, because it is not a property of the model. It is a property of the harness you run the model in.

Fazm wraps the same Claude Code agent loop in a native macOS app, and it exposes one setting that closes this gap: a Custom API Endpoint. The mechanism is small enough to read in full, so the rest of this page is the actual code, the exact constraint, and the steps to use it. For the wider argument about why this layer matters more than which model wins a given month, see why agent tooling beats model upgrades.

The anchor fact: 12 lines that decide what your agent talks to

When you fill in the Custom API Endpoint, Fazm does not blindly trust it. Before the value is handed to the agent subprocess as ANTHROPIC_BASE_URL, it has to survive a guard. This is the whole of it, from Desktop/Sources/Chat/ACPBridge.swift lines 633 to 650.

// Custom API endpoint (allows proxying through Copilot, corporate gateways, etc.)
// Only forward it when it's an absolute http(s) URL with a host. A malformed value
// (missing scheme, "localhost:8766", stray text) otherwise lands in ANTHROPIC_BASE_URL
// and makes the Anthropic SDK throw "API Error: Invalid URL" on every query, silently
// bricking built-in chat. Falling back to the default keeps chat working.
if let customEndpoint = defaults.string(forKey: "customApiEndpoint")?
  .trimmingCharacters(in: .whitespacesAndNewlines), !customEndpoint.isEmpty {
  if let url = URL(string: customEndpoint),
    let scheme = url.scheme?.lowercased(), scheme == "http" || scheme == "https",
    let host = url.host, !host.isEmpty {
    env["ANTHROPIC_BASE_URL"] = customEndpoint
  } else {
    log("ACPBridge: ignoring malformed customApiEndpoint '\(customEndpoint)' " +
        "(expected an http(s) URL with a host); falling back to default endpoint")
  }
}

The guard is the interesting part. It requires the string to parse as a URL with an http or https scheme and a non-empty host. That is why "localhost:8766" is rejected (no scheme) while "http://localhost:8766" is accepted. The reason is spelled out in the comment: a malformed value would otherwise reach the Anthropic SDK, throw "API Error: Invalid URL" on every query, and silently brick chat. So the failure mode is a clean fallback to the default endpoint, not a dead window. The setting can only ever break loudly into the log, never quietly into your workflow.

Pointing your loop at one of this week's models

Five steps, none of which require a second tool or a fresh session. The model behind the endpoint can be Gemini 3.5 Flash, a Chinese open-weights model, or anything else, as long as the endpoint itself speaks the Anthropic Messages format.

What the launch log prints

The guard is not theoretical. Here is what the bridge logs on two saves: one valid endpoint pointed at a local Anthropic-compatible bridge, and one malformed value that gets rejected and falls back.

The second case is the one most tools get wrong. A missing scheme is an easy paste mistake, and the lazy implementation lets it through and dies on every query. Here it is caught, logged, and survived.

Which of this week's models you can actually run, and how

Honest accounting. "Native" means the bare Claude Code CLI reaches it without help. The last column is what Fazm adds.

The pattern is the same for the bottom three rows: put something that speaks Anthropic in front of the model, point the endpoint at it. The Settings help text says it plainly, a raw Gemini or OpenAI key will not work in that field. For the deeper story on how Fazm absorbs an entire month of releases without an app update, see the April 2026 LLM release breakdown.

Why the week's real story is the harness, not the model

Look at what the labs themselves emphasized. Gemini 3.5 Flash was sold as an agent model, benchmarked on Terminal-Bench and MCP Atlas rather than on chat quality. The OpenRouter data showed programming now drives more than half of all token usage on the platform. Both signals say the same thing: the value is shifting from the raw model to the loop you run it in, the tools it can call, and whether your context survives long enough to finish a real task.

That is exactly the part a Custom API Endpoint protects. Swap the model behind it as often as the news cycle changes, and the loop you built, the persistent window, the forks, the full uncompacted history, stays put. The model is the part that churns weekly. The harness is the part you keep.