OpenAI's new model release in May 2026, and the 17-line filter that lets your agent pick it up

The May 2026 ledger

Two announcements landed in May. The full GPT-5.5 family arrived a week earlier, in late April; it is included for completeness because it is what an agent loop actually consumes.

Ledger compiled from OpenAI model release notes and Fazm source on May 26, 2026.

What the headline numbers say

The 52.5% number is an OpenAI-published internal evaluation against its own previous default. It is the claim, not an independent benchmark.

Where the new model gets accepted: a 17-line Swift filter

Fazm wraps the Codex CLI through the Agent Client Protocol. When the ACP bridge probes the Codex adapter, it gets back a list of model ids that look like gpt-5.5/high, gpt-5.5/xhigh, gpt-5.4-mini/low, gpt-5.3-codex/high. That raw list is everything the upstream adapter knows about. Fazm does not hardcode a copy of it. The picker in the floating bar runs every id through a single static function that decides whether the model belongs to the current frontier generation.

Here is the function verbatim, copied from Desktop/Sources/CodexBackendManager.swift (lines 186 to 205 as of the May 26, 2026 build):

// Desktop/Sources/CodexBackendManager.swift
//
// Returns true when the modelId belongs to the current frontier
// generation the picker should expose (gpt-5.5 or newer).
// Inputs look like "gpt-5.5/high", "gpt-5.4-mini/low",
// "gpt-5.3-codex/high".
static func isPickerEligible(modelId: String) -> Bool {
    let family = modelId
        .split(separator: "/").first.map(String.init) ?? modelId
    // Strip variant suffixes ("-mini", "-codex") so we only
    // compare the base version.
    let base = family.split(separator: "-")
        .prefix(2).joined(separator: "-")
    guard base.hasPrefix("gpt-") else { return false }
    let version = String(base.dropFirst("gpt-".count))
    let parts = version.split(separator: ".")
    guard parts.count == 2,
          let major = Int(parts[0]),
          let minor = Int(parts[1]) else { return false }
    // Keep gpt-5.5 and newer (e.g. 5.5, 5.6, 6.0)
    if major > 5 { return true }
    if major == 5 && minor >= 5 { return true }
    return false
}

Why a filter beats an allowlist when OpenAI ships weekly

The shape of the code matters more than the line count. Most desktop AI apps enumerate the models they support as an enum or a switch statement. That couples every release to an app version. Fazm inverts the relationship: the adapter is the source of truth for what exists, the app only judges whether a new id looks like it belongs to the frontier.

enum Model: String, CaseIterable {
    case gpt5_3_low      = "gpt-5.3/low"
    case gpt5_3_medium   = "gpt-5.3/medium"
    case gpt5_3_high     = "gpt-5.3/high"
    case gpt5_4_low      = "gpt-5.4/low"
    case gpt5_4_medium   = "gpt-5.4/medium"
    case gpt5_4_high     = "gpt-5.4/high"
    case gpt5_5_low      = "gpt-5.5/low"
    case gpt5_5_medium   = "gpt-5.5/medium"
    case gpt5_5_high     = "gpt-5.5/high"
    case gpt5_5_xhigh    = "gpt-5.5/xhigh"
}

// To support gpt-5.6: edit the file, add four cases, ship.

What happens the day codex-acp adds GPT-5.6

Four things, in order. None of them require a Fazm release.

About GPT-5.5 Instant specifically

GPT-5.5 Instant is the ChatGPT product default. It is tuned for the chat UI: smarter answers, more concise responses, personalization from past chats, files, and connected Gmail for Plus and Pro web users. It is not the same SKU as the developer-side gpt-5.5 tiers that an agent loop consumes through the Codex CLI.

Fazm reaches GPT-5.5 a different way. When you connect your ChatGPT subscription through the CodexAuthSheet in Settings, the bridge negotiates OAuth, writes the resulting credential to ~/.codex/auth.json, and the Codex adapter starts surfacing the reasoning-effort variants: gpt-5.5/low, gpt-5.5/medium, gpt-5.5/high, and gpt-5.5/xhigh. Those are the strings that flow through the filter and into the picker. The Instant variant lives inside the ChatGPT app where its product tuning makes sense; the effort tiers live inside the agent loop where reasoning depth and latency are parameters you actually want to control per turn.

About the May 7 realtime voice models

OpenAI's May 7 release was on a different surface: the realtime API, not the Codex CLI. The new voice models reason, translate, and transcribe in a streaming session aimed at building voice agents. That capability is real, but it is not what an agent loop on your Mac normally talks to. It would only land in Fazm if codex-acp added it to the model list, which it has not as of May 26, 2026.

The practical answer for a Fazm user who wants voice today: Fazm already has its own voice path. Hold the configured hotkey, talk, the audio is captured on-device and transcribed by WhisperKit before being handed to whichever agent backend the chat window is set to. The audio never leaves your machine until the transcribed prompt is sent. That works against any model in the picker, Claude or Codex, including GPT-5.5.

What this should change about how you read release notes

Release calendars do half the work. They tell you a new model exists. They do not tell you the path from the announcement to the tool you actually run. For a developer who lives inside an agent loop on a Mac, the second half is the one that matters. The whole point of a frontier-version filter (instead of a hand-maintained allowlist) is that the path from announcement to picker is zero steps on the app side.

When GPT-5.6 ships next, the Fazm binary you already have on disk will see it the next time codex-acp does. When GPT-6 ships later this year, the same is true. The model layer is where the velocity is; the harness layer (persistent sessions surviving a restart, one-click forking, no auto-compacting of context, real accessibility-API reach into your browser and Mac apps) is the part that should stay stable. That is the whole point of the split.