9. Models and cost

Aliases move. On Bedrock and Google Cloud's Agent Platform, sonnet still means Sonnet 4.5; on Foundry, opus means Opus 4.6. To pin, use the full model name (claude-opus-5-5) or ANTHROPIC_DEFAULT_OPUS_MODEL.

In priority order: /model in the session, claude --model at launch, ANTHROPIC_MODEL, the model field in settings, then ANTHROPIC_DEFAULT_MODEL.

• /model <name> saves. Typing a name, or pressing Enter in the picker, switches now and saves your default. Press s in the picker for this session only.
• Project and managed settings reapply every launch, and the startup header names the file that set the model.
• A skill's model: lasts for the rest of the turn that invoked it. A subagent resolves its model from the per-call choice, then its frontmatter, then CLAUDE_CODE_SUBAGENT_MODEL, then the session's model. A resumed session keeps the model it was saved with.
• Admins restrict choices with availableModels in managed settings. The allowlist covers /model, --model, settings, subagents, skills and fast mode.
• Each model has its own prompt cache. A mid-session switch rereads the whole conversation uncached; Claude Code asks you to confirm while the cache is warm.

• Resolution order: explicit (CLAUDE_CODE_EFFORT_LEVEL, --effort, /effort), then settings (a level saved per model under modelSettings, or effortLevel), then the model's default.
• A top-level effortLevel in user settings doesn't count for Opus 5.5. It is the older form; Opus 5.5 uses a per-model saved level or its own medium default.
• In Anthropic's testing, Opus 5.5 at medium matches or exceeds Opus 5 at high. Moving from Opus 5, start at medium.
• ultrathink in a prompt asks for deeper reasoning on that one turn. ultracode is a setting, not a level. /effort auto clears your saved level for the active model.
• Admins cap effort with maxEffortLevel (any plan or provider) or per-role limits on Enterprise; the lower cap wins.
• On Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 with an API key or subscription, changing effort keeps the prompt cache. On most other setups it doesn't.

• Native 1M: Fable 5.1, Fable 5, Sonnet 5 and later, Haiku 5.5, Opus 4.7 and later. No [1m] needed.
• Opus 4.6 and Sonnet 4.6 reach 1M only through their [1m] variant, which depends on your plan.
• Native-1M sessions auto-compact at about 967K tokens. /autocompact 500k compacts earlier for the current model (range 100K to 1M); CLAUDE_CODE_DISABLE_1M_CONTEXT=1 holds every model to 200K.
• On the Anthropic API, tokens past 200K carry no premium. The exception is Haiku 5.5, which costs more on prompts over 100K tokens.

• The same Opus, up to 2.5x faster, at a higher price per token: $8 / $40 per million input / output tokens on Opus 5.5, $10 / $50 on Opus 5 and Opus 4.8. Those three are the only supported models.
• Research preview. Anthropic API and subscriptions only: not on Bedrock, Agent Platform, Foundry, or Claude Platform on AWS.
• On subscriptions it draws on usage credits only, never the plan's included usage. Team and Enterprise start with it off until an Owner enables it.
• Turn it on at the start. The first switch-on bills the whole conversation so far as uncached input at fast rates, and fast and standard speed keep separate caches.
• Toggle with /fast or Option+O on macOS. fastModePerSessionOptIn: true makes every session start with it off.

Reading /usage (aliases /cost and /stats)

• The Session block is a local estimate at list price, unless an admin sets modelPricing to show contracted rates. It resets on /clear. For subscribers it is not billing.
• The Prompt cache line shows the share of input served from cache, misses and their likely cause, and whether the cache is warm.
• On Pro, Max, Team and Enterprise, the plan breakdown attributes recent usage to skills, subagents, plugins and MCP servers, and recommends changes that touch at least 10% of usage. Press f to apply the ones Claude Code can, d or w for 24 hours or 7 days. It covers this machine only.
• Background jobs (summaries for resume, status checks) use under $0.04 per session.

Across enterprise deployments, the average is about $13 per developer per active day and $150 to $250 per developer per month, with 90% of users under $30 per active day. Pilot a small group to set your own baseline.

OpenTelemetry works on every setup and is the only option that streams per-user tokens and cost into your own monitoring in near real time.

Console rate-limit recommendations per user, by team size:

Habits that move the number: /clear between unrelated tasks, Sonnet for most coding with Opus for hard reasoning, plan mode before big changes, workflow text moved from CLAUDE.md into skills, and subagents for noisy work like test logs. Agent teams use about 7x the tokens of a standard session when teammates run in plan mode. When spend is unexpectedly high, the docs trace it to long sessions that were never cleared and Opus left as the default.