Claude Code v2.1.243 adds loop token visibility, curated model picker, contracted pricing, and granular prompt-cache controls

Claude Code v2.1.243, released on 24 August 2026, is a focused release aimed squarely at the people who run Claude Code at scale: developers managing agentic loops, IT admins governing model access on Bedrock or Vertex, and finance teams trying to reconcile AI spend against negotiated contracts. Four new settings and one new /usage view address the gaps that tend to surface once an organisation moves past experimentation and into production operation.

Loops now show up properly in /usage

If you have ever watched a Claude Code session burn through tokens without a clear sense of which task was responsible, the new Loops breakdown in /usage is the fix you have been waiting for.

The breakdown surfaces per-loop run count, total tokens consumed, tokens per run, and the timestamp of the last run. In practice, this means you can open /usage and immediately see which background loop is responsible for an unexpected spike, rather than hunting through raw session logs or relying on external telemetry.

The stakes here are real. Claude Code’s query loop resends the full message history, system prompt, and tool schemas on every retry. A poorly scoped prompt on a long session with several retries can quietly consume hundreds of thousands of tokens before producing anything useful. Visibility is the first step to controlling that.

Curate the model picker with modelPicker

The new modelPicker managed setting gives administrators an ordered, labelled list of models to display in the /model picker. Critically, it accepts any ID spelling, including Vertex and Bedrock model IDs, and can append to or entirely replace the built-in model lineup.

For organisations running Claude Code through Amazon Bedrock or Google Cloud Vertex AI, this matters for two reasons. First, model IDs in cloud-provider deployments differ from Anthropic’s native API IDs, and the picker previously had no clean way to surface them with accurate labels. Second, governance: if your organisation has approved a specific set of models for cost or compliance reasons, you can now ensure developers only see those models in the picker rather than the full catalogue.

One thing worth knowing: /model changes the model for the current session only. Pressing d in the picker sets a new default. The existing model setting and ANTHROPIC_MODEL environment variable continue to control the persistent default, so modelPicker is purely about what appears in the menu, not about overriding those underlying controls.

Contracted rates in /cost with modelPricing

Public list prices and actual negotiated rates rarely match, which creates a persistent headache when developers or finance teams use /cost output to estimate or report spend. The new modelPricing managed setting resolves this by letting you specify contracted per-model rates and a discount multiplier.

Once configured, those rates flow through to /cost, the status line, and telemetry cost figures. That means the numbers a developer sees in their session, and the numbers your observability tooling captures, both reflect what your organisation actually pays rather than what Anthropic publishes on its pricing page.

For teams that feed Claude Code telemetry into a cost allocation dashboard or chargeback system, this is a meaningful improvement. It removes the manual step of multiplying reported costs by a discount factor after the fact.

Independent prompt-cache TTLs for main sessions and subagents

Prompt caching is one of the more effective ways to reduce token spend on repetitive agentic tasks, but the optimal cache lifetime varies considerably depending on what is being cached. A long-running main conversation benefits from a longer TTL; short-lived subagents spinning up and down rapidly often do not need the cache held for as long.

v2.1.243 introduces two independent settings: promptCacheTtl for the main conversation and subagentPromptCacheTtl for subagents. The example from the changelog is instructive: a 1-hour TTL on the main conversation while subagents use 5 minutes. This lets you optimise cache cost at each layer of the agent hierarchy rather than accepting a single setting that is always a compromise.

These settings are available to API-key users and cloud-provider deployments. If you are running multi-agent workflows at any meaningful scale, it is worth revisiting your cache configuration with these controls in mind.

What else shipped

Two other additions are worth a quick mention even though they are not the headline features.

Keyless sign-in is now available under /login via “Sign in with your Console account”. This serves organisations whose security policies prohibit API keys entirely. Previously those teams had no clean path to Claude Code without an exception process; now they have a supported route.

The /status command also gains a “Skipped sources” line, which lists any managed settings sources (such as managed-settings.json) that are present but not being applied because a higher-precedence source is active. That kind of transparency is easy to overlook, but it eliminates a confusing class of “my settings aren’t working” support tickets where the actual cause is a precedence conflict.

What this means for your team

If you are an IT or DevOps admin managing a Bedrock or Vertex deployment, modelPicker and modelPricing give you the governance primitives that production operation requires. You can now control exactly which models appear in the picker and ensure cost reporting reflects reality.

If you are a developer running agentic loops, the /usage breakdown finally makes token consumption legible at the task level. Combined with independent prompt-cache TTL controls, you have the tools to optimise both visibility and spend without guesswork.

The full changelog is available at code.claude.com/docs/en/changelog. You can check your installed version by running claude --version.