FAQ

gpt-5.5 stops around 265k context. What should I do?

Use a Claude 1M-context model for compaction, then switch back to gpt-5.5[1m].

A practical sequence inside Claude Code:

/model opus[1m]      # or /model sonnet[1m]
/compact
/model gpt-5.5[1m]

Why this helps:

• gpt-5.5 routing is handled by llmux, but Claude Code still owns local context accounting and compaction behavior.
• In long sessions, Claude Code can block a gpt-5.5 session around the mid-200k range, even when gpt-5.5[1m] is selected.
• Switching temporarily to a Claude model with a 1M context window gives Claude Code a known large-window model for /compact.
• After compaction reduces the active transcript, switching back to gpt-5.5[1m] continues routing through llmux to the Codex group.

The ~265k cutoff is empirical client behavior, not an llmux routing limit. The [1m] suffix improves Claude Code's context-window display, but it does not guarantee every un-compacted long transcript will be accepted unchanged.

This does not change your llmux account configuration. It is a Claude Code session-management workaround.

Why does [1m] matter in /model gpt-5.5[1m]?

Claude Code derives its displayed context window from the model-name string. Bare gpt-5.5 can be treated as an unknown or smaller-window model by the client. The [1m] suffix tells Claude Code to use a 1M context display while llmux still routes the request by the gpt- prefix.

llmux strips one trailing [1m] before resolving the codex model, so the suffix never reaches upstream and works on any codex id or alias (gpt-5.6-sol[1m], sol[1m]). gpt-5.6-sol[1m] and gpt-5.6-terra[1m] are also catalog rows advertising a 1000000 window — probed 2026-08-21 at 910,229 input tokens accepted / ~936k rejected against the ChatGPT-account backend (OpenAI publishes 1,050,000 total for the gpt-5.6 family). gpt-5.5 is a 272k model, so gpt-5.5[1m] remains a display-only workaround.

See operational-reference.md for the routing details, and models.md for the catalog rows.

Does gpt-5.5[1m] still route to Codex?

Yes. The gpt- prefix still matches the Codex group. llmux strips the display suffix for routing and usage attribution.

Does llmux replace Claude Code?

No. llmux intentionally keeps Claude Code as the harness. It sits behind Claude Code as a local Anthropic-compatible proxy so the account/model layer can move while your harness stays fixed.

Can I use llmux with only Claude accounts?

Yes. The durable core is multi-account Claude scheduling and Claude Code integration. Codex routing is optional.

Is llmux for sharing accounts across a team?

No. llmux is for one human using their own accounts. It is not for credential pooling, resale, or shared subscription brokerage.