What happened
OpenAI’s GitHub release for Codex version 0.166.0 lists GPT-6.1 Terra as added to three model catalogs: the bundled catalog, Vertex AI, and Amazon Bedrock. Codex is OpenAI’s coding agent client, and a model catalog is the client-side list used to expose selectable model options for a given route. The release does not state that every Vertex AI or Bedrock account receives access to the model, nor does it give pricing, quota, regional availability, or provider-configuration instructions. It establishes that the CLI now includes Terra in those catalog definitions.
The release also makes several interaction changes. Users can select transcript text from the keyboard with word, line, and page motions; the selection experience includes search and configurable shortcuts. Markdown is rendered consistently in asynchronous questions, choices, answers, and transcript replies. These are client-interface changes rather than changes to model behavior.
For queued work, version 0.166.0 adds codex queue --resume, described as a way to resume unloaded threads after a message has been queued. The terminal UI also receives clearer safety-block explanations and selectable safer follow-up prompts. A new codex doctor --network-only option narrows diagnostics to configuration, connectivity, and WebSocket issues. That gives operators a more focused first check when a client cannot connect, without the release claiming to diagnose every possible failure.
The bug-fix list focuses on continuity across interrupted or resumed work. Codex now preserves acknowledged approval and permission settings when a thread is resumed or forked. It preserves local image attachments when restoring killed composer text, restores spawned descendant threads when a parent is unarchived, and improves terminal UI reconnection recovery. The release also preserves reasoning-token accounting across resume and fork, which it says avoids premature compaction. Separately, it includes packaging and build work, including shipping the standalone executor in canonical CLI and app-server release packages.
What the archive adds
This is a new fact on a known Terra story: the model was already covered as available across OpenAI surfaces in OpenAI releases GPT-6.1 Terra, a cheaper fast model for computer use, in ChatGPT Work, Codex and the API and as a live API model in GPT-6.1 Terra Model | OpenAI API. This release narrows the additional fact to Codex CLI’s bundled, Vertex AI, and Bedrock catalogs.
Trend signal
This is modest evidence for Harness architecture outlives model choice. The release separates the evolving model menu from the rest of the harness: the same client also changes transcript handling, thread recovery, approvals, diagnostics, and packaging. It does not prove that the harness is more durable than any individual model, but it shows that the operational surface is broader than a model name.
Why it matters
Teams rarely operate a coding agent as only a prompt box. They need a repeatable route from client to model provider, durable session state, explicit approvals, and useful diagnostics when the route fails. Adding Terra to the relevant catalogs removes one client-side gap for teams that use Codex through Vertex AI or Bedrock, but it is only one part of that route. Access controls, provider enablement, credentials, regions, and commercial terms can still determine whether the model is usable.
The workflow changes matter for longer-running agent work. Resuming unloaded threads, retaining permission acknowledgements, and maintaining reasoning-token accounting can reduce accidental loss of context or repeated approval handling. The new network-only diagnostic is also a concrete operational affordance: it can help distinguish connectivity and WebSocket problems from task or model-selection problems before a team changes prompts or code.
Limits and uncertainty
The release notes are a changelog, not an access policy. They do not specify which Vertex AI or Bedrock regions expose GPT-6.1 Terra, which credentials or account tiers qualify, the model identifier a user should select, or whether provider-side limits differ from OpenAI’s own route. They also provide no benchmark results, latency measurements, reliability data, or comparison with other Codex-supported models. “Added to catalogs” should therefore be read as a client capability update, not as proof of universal availability or a recommendation to migrate production workloads.
Skill to build now
Treat model selection as a deployment check rather than a static configuration. This week, update a non-production Codex environment to 0.166.0, inventory whether it uses the bundled route, Vertex AI, or Bedrock, and verify what models the configured route actually exposes. Run codex doctor --network-only if connectivity is in doubt. Then resume or fork a representative thread and confirm that the approval settings your team expects remain intact. Record the result alongside the provider, region, and account context so model availability does not become tribal knowledge.
Reader takeaway
Codex CLI 0.166.0 makes GPT-6.1 Terra a recognized option in its bundled, Vertex AI, and Amazon Bedrock catalogs and improves the mechanics of operating long-lived sessions. For builders, the next move is verification: confirm the provider route exposes the model and that resumed threads retain the controls your workflow relies on.
Open question
Which Vertex AI and Amazon Bedrock account, region, and permission combinations actually make GPT-6.1 Terra selectable after the CLI update?