Model Aliases & 1M Context

Claude Code resolves the short aliases (opus, sonnet, haiku, fable, opusplan) client-side, before any request reaches shunt. The table it resolves against is compiled into the CLI binary, and two of its branches key off how the session is connected. Point ANTHROPIC_BASE_URL at shunt and both branches change — so the same /model opus can select a different model than it does against api.anthropic.com.

shunt cannot influence this: the choice is already made by the time it sees the request. What shunt can do is remap the result — see Recover the intended tiers.

The two gates

Every difference on this page traces back to these:

The second gate is a strict host match. Any base URL that isn’t api.anthropic.com fails it, including a shunt instance that proxies straight through to Anthropic with a passthrough credential.

Alias resolution

opus and sonnet carry per-deployment overrides. fable and haiku do not, so they resolve the same way everywhere.

The gateway column is the surprising one: a gateway session pins opus two releases behind and sonnet one behind. Those pins have not moved across the releases checked (2.1.217–2.1.220) — what changed is the other column. Claude Code 2.1.219 moved the non-gateway opus default from Opus 4.8 to Opus 5, which widened a 4.8-vs-4.7 gap into a 5-vs-4.7 gap and made the pin visible.

Anthropic documents the equivalent table for third-party providers (Bedrock, Vertex, Foundry) but does not list a gateway row, so this divergence is easy to miss.

[1m] never reaches shunt

The suffix is a client-side hint, not part of the model id. Claude Code strips it before sending and expresses the request as a beta header instead:

POST /v1/messages?beta=true
{"model": "claude-opus-5", ...}
anthropic-beta: claude-code-20250219,context-1m-2025-08-07,...

So shunt never sees claude-opus-5[1m] on the wire — it sees claude-opus-5 plus context-1m-2025-08-07. (shunt also strips a trailing [1m]/[1M] before route matching and availableModels enforcement, so a hint that does arrive — from a hand-written client — still routes correctly.)

1M context is not applied automatically

Claude Code sizes the context window in this order: an explicit [1m], then the context-1m beta header, then whether the model is natively 1M on this deployment — and that last check requires the api.anthropic.com host gate. Behind any other base URL it fails, so natively-1M models fall back to the 200K default.

Anthropic states the same constraint: “when ANTHROPIC_BASE_URL points at a gateway, Claude Code can’t verify 1M support” (extended context).

The practical rule behind shunt is to always write the suffix — /model opus[1m], /model fable[1m] — or bake it into the pin:

export ANTHROPIC_DEFAULT_OPUS_MODEL='claude-opus-5[1m]'

Two details worth knowing:

• A bare alias has never auto-applied [1m] on any release checked (2.1.217–2.1.220). If 1M stopped appearing after an upgrade, the cause is the window check above, not a change in alias handling.
• claude-fable-5 has no [1m] variant in its catalog entry — it is natively 1M and nothing else. Against api.anthropic.com a typed fable[1m] is silently reduced to fable; behind shunt the suffix is kept, because it is the only way to reach the full window.

CLAUDE_CODE_MAX_CONTEXT_TOKENS is not an escape hatch here — it applies only to ids that do not start with claude-, as covered in Effort & Context.

Fable disappears from the picker

Fable is gated on the same host check: when the deployment is firstParty and the base URL isn’t api.anthropic.com, Claude Code filters every Fable entry out of /model and reports a configured claude-fable-5 as absent. A "model": "claude-fable-5[1m]" in settings.json then silently falls back to Opus.

Gateway-login sessions are exempt — Fable stays selectable there. Two other ways to reach it on the base-URL path:

# Bypasses the availability check entirely
export ANTHROPIC_DEFAULT_FABLE_MODEL=claude-fable-5

An explicit --model fable / /model fable also resolves, because the availability gate governs the picker and the saved default rather than direct id resolution. This is unchanged across 2.1.217–2.1.220.

Recover the intended tiers

shunt can’t change what the client picks, but it can rewrite the model on the way through. If your users are on gateway login and you want the opus/sonnet aliases to land on the current tiers, remap the pinned ids:

[[routes]]
model = "claude-opus-4-7"        # what a gateway session sends for `opus`
provider = "anthropic"
upstream_model = "claude-opus-5" # what you want it to actually run

[[routes]]
model = "claude-sonnet-4-6"
provider = "anthropic"
upstream_model = "claude-sonnet-5"

Curated [[models]] entries are unaffected: a discovered id is selected verbatim from the picker and never passes through the alias table.

Version history

Verified by extracting the alias table from each installed CLI binary and confirming the result against a request capture.

Opus 5 requires Claude Code 2.1.219 or later; on 2.1.217–2.1.218 the opus alias cannot reach it at all, on any deployment.