AI model releases in 2026

The verified shortlist

Five model families carry most of the weight when people ask about 2026 releases. Each row below links to a primary or first-tier source. Where I could not re-verify a precise day, the row says "Spring 2026" rather than invent one.

Counting open-weight and regional models, the real H1-2026 list runs to dozens. The five above are the ones a working developer is most likely to reach for.

The list is the wrong thing to optimize

Look at what happened with Fable 5. Anthropic shipped its most powerful public model on June 9. Thirteen days later, on June 22, it was pulled from general availability under a US export directive, and while it was live some queries silently routed to Opus 4.8 instead. If your workflow was welded to "the newest Claude," you got a moving target, a fallback you did not choose, and then a removal, inside two weeks.

That is not an anomaly. It is the texture of 2026. GPT-5.5 held back API access at launch for a day. Grok 5 slipped past its target more than once. Gemini moved in a stream of point releases with no single headline you could plan around. The cadence is fast, uneven, and occasionally reversible. Optimizing your setup around which model is best this week means rebuilding it every few weeks.

The thing that does not churn is the layer around the model: your chats, your forks, your accumulated context, the tools the agent can reach. Most coding tools couple that layer tightly to one provider, so a model change is a migration. Fazm decouples them on purpose, and it does it in a way you can read in the source.

What a new release looks like in the code

Fazm wraps the real Claude Code agent loop over ACP, and bundles Codex as a second backend. The Codex side lives in CodexBackendManager.swift, which keeps a list of the models the adapter exposes and defaults the picker to gpt-5.5/medium, hiding older generations like gpt-5.4 and gpt-5.3-codex once a newer one works. When GPT-5.5 shipped in April, "adopting" it meant the picker showed it. That is the whole upgrade.

Pointing the Claude loop at a different provider entirely is one field. Here is the exact block from Desktop/Sources/Chat/ACPBridge.swift that reads your custom endpoint setting and rewires the agent:

if let raw = defaults.string(forKey: "customApiEndpoint")?
     .trimmingCharacters(in: .whitespacesAndNewlines), !raw.isEmpty {
  if let endpoint = Self.validCustomAPIEndpoint(raw) {
    env["ANTHROPIC_BASE_URL"] = endpoint
    env["FAZM_CUSTOM_API_ENDPOINT"] = "true"
    // never leak the bundled key to a user proxy
    env["ANTHROPIC_API_KEY"] = "sk-fazm-custom-endpoint"
  }
}

That is the anchor. The setting is validated first (a malformed value like localhost:8766 is rejected and chat falls back to the default rather than bricking), then the same Claude Code loop is pointed at GitHub Copilot, a corporate gateway, or a local Anthropic-compatible server like LM Studio or Ollama. A brand-new open-weight model behind a compatible shim is reachable the day it ships, with no Fazm update required. The model became an environment variable. Everything you built on top of it, the persistent window, the forks, the screen and browser reach, stays put.

So what do you actually do with the 2026 firehose

Stop treating each release as a project. Treat it as a setting you flip when a benchmark or your own testing says it is worth flipping. Concretely:

• Keep two windows open, one on Claude and one on Codex, and route each task to whichever is winning that week. When Opus 4.8 led on agentic coding by Anthropic's reported numbers, that was a window choice, not a migration.
• Pin the model your project depends on so a surprise deprecation (or a Fable-5-style pull) is a one-click fallback, not an outage.
• Bring your own Claude Pro or Max account so a price change on a new model hits your existing plan, not a per-seat tool markup.
• For anything you cannot or will not send to a cloud, point the same loop at a local endpoint and keep the exact same chat UX.

Questions people actually ask