Claude Code 2.1.283 keeps new models out until admins say so

Claude Code 2.1.283 adds a setting called availableModelsMatch. Set it to "exact" and each entry in a model allowlist covers only the version it names, so the next model release stays blocked until an admin adds it. The release notes on Github also add a deny list for models, an audit for outdated prompts and a header that lets LLM gateways group requests.

At a glance

• Admins get two new managed settings for models: availableModelsMatch, which pins allowlist entries to exact versions, and deniedModels, which blocks named models even when the allowlist would let them through.
• The new /doctor prompt-audit command, also reachable as /checkup prompt-audit, scans CLAUDE.md files, skills, agents and commands for prompting patterns that were written for older models.
• One default changes: on third-party providers or with telemetry off, interactive sessions with no configured permission mode start in auto mode, although permissions.defaultMode still overrides that.

If you haven't been following, Claude Code has been tightening its model allowlists one step at a time. According to GitHub release notes, v2.1.175 added enforceAvailableModels, which makes the availableModels list apply to the Default model too and stops user or project settings from widening a managed list. According to TheRouter.ai, an earlier release added requiredMinimumVersion and requiredMaximumVersion, which make the agent refuse to start outside an approved version range.

availableModelsMatch and deniedModels give admins two ways to keep a model out

The first setting, availableModelsMatch, changes how Claude Code reads the existing availableModels list. With "exact", an entry allows only the model version it names. When a newer release of that model ships, it stays blocked until someone adds it to the list, so a team can try a model before developers can pick it.

The second setting, deniedModels, works from the other side. It blocks specific models even when availableModels allows them, so an admin can keep a broad allowlist and still block individual models. Both are managed settings. The Claude Code docs describe these as applying above every other level, so no user, project, local or --settings value overrides them, apart from a few security-sensitive exceptions.

The model picker also gets some cleanup. /model stops accepting Sonnet 4.6 or Sonnet 5 with [1m] when the id carries a date or -v1:0 suffix, in the cases where the plain id was refused. The picker also stops showing a hardcoded Haiku version and price when ANTHROPIC_DEFAULT_HAIKU_MODEL pins a different model.

Why would a team want new models blocked by default?

According to Codingnomads, when you deploy through Bedrock, Vertex or Foundry, Claude Code's model aliases resolve to the latest version available in your account. Anthropic ships a model, the cloud provider enables it, and the alias can end up pointing at a model your account hasn't enabled yet. Codingnomads says the result is authentication errors for every developer on the team.

Think of the allowlist as a standing order at a shop. An order for "milk" gets whatever milk the shop has that week. An order for one brand and one carton size gets exactly that or nothing. Setting availableModelsMatch to "exact" makes every entry work like the second kind of order.

Where the list lives matters too. According to GitHub documentation, availableModels arrays set at several levels are merged and deduplicated, so a strict list belongs in managed or policy settings. The Claude Code docs list four ways to deliver them: the claude.ai admin console, polled hourly; MDM or OS policy, checked every 30 minutes; a managed-settings.json file, reloaded on change; and the HKCU registry on Windows and WSL.

x-claude-code-prompt-id lets gateways group the requests behind one prompt

For teams that route traffic through an LLM gateway, the release adds x-claude-code-prompt-id to the gateway hint headers. With CLAUDE_CODE_GATEWAY_HINT_HEADERS=1 set, a gateway can group every request that serves one user prompt. The Claude apps gateway gains an opt-in load_test_mode block, which builds and signs requests without sending them upstream and returns a canned reply to clients.

The same gateway gets a mantle upstream provider for Amazon Bedrock's Mantle endpoint. For observability, OTEL_LOG_TOOL_CONTENT=1 now also adds MCP tool, WebFetch and WebSearch outputs to the tool.output OpenTelemetry span event.

The new /doctor prompt-audit command audits CLAUDE.md files, skills, agents and commands for prompting patterns written for older models. Its report opens with stale paths, stale commands and instruction files that contradict each other. Thinking keywords that Claude Code documents stay in place.

Sessions on third-party providers start in auto mode when nothing is configured

One change affects a default rather than adding a feature. Interactive sessions on third-party providers or with telemetry off start in auto mode when no permission mode is configured, and permissions.defaultMode still overrides that. Managed sandbox settings used to be ignored entirely when one nested value was invalid. Now the invalid value fails closed and the rest of the block still applies.

There are also fixes. On Windows, the PowerShell tool no longer lets cmd /c rd, rmdir, del or erase delete drive roots, the home folder and other folders that Remove-Item refuses. A brief HTTP 404 from a stateless remote MCP server, such as a proxy mid-redeploy, no longer leaves that server unusable for the session. Code Review stops charging for reviews that hit their time limit with nothing verified.

The release notes don't say exactly how an availableModels entry matches when availableModelsMatch is left unset. They only suggest that newer releases can get through. They also don't say how "exact" treats the date and -v1:0 suffixes mentioned in the /model fix. In our view, admins should read the auto-mode default first, because it changes behavior for teams that configured no permission mode at all.

When the next model ships

The first real test of "exact" matching will come with Anthropic's next model release. Teams that turn it on will have to add the new version by hand before anyone can select it, and the notes don't say when that release is due. According to TheRouter.ai, Claude Code has averaged roughly one release per business day in 2026, so these settings may change again soon.