What Is Ultracode in Claude Code? Who Gets It and What It Costs

Ultracode is not max effort, ultrathink or the Max plan

It is often confused with three other things: the max effort level, the ultrathink keyword and Claude's Max and Ultra plans. Read the table by its first column; only the first two rows are ultracode.

Anthropic introduced ultracode with dynamic workflows on May 28, 2026. Until Claude Code v2.1.284 (September 28, 2026), turning it on also set effort to xhigh; since then it is a separate toggle that keeps the session's level. Starting Claude Code with claude --effort ultracode still does both: it turns ultracode on and sets effort to xhigh.

That makes the session's effort level matter. Claude Code runs Opus 5.5, Sonnet 5.5 and Haiku 5.5 at medium effort by default and Fable 5.1 at high, so an ultracode session on Opus 5.5 plans its workflows at medium unless you raise it. Ultracode also needs a model that supports xhigh effort: Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 do; Opus 4.6, Sonnet 4.6 and models without effort support, such as Sonnet 4.5, do not, and Claude Code does not offer ultracode on them.

The keyword counts only in a prompt you type yourself, at the interactive prompt, in an IDE extension panel or in a Remote Control client. A prompt passed with -p, a scheduled task, a webhook or a pull request comment that contains the word does not start a workflow.

Every paid plan has it; Pro turns workflows on first

Anthropic makes dynamic workflows available on all paid plans, with Anthropic API access, and on Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry, and ultracode comes with them on any model that supports xhigh effort. Workflows are on by default for Max, Ultra, Team, Enterprise and API use, and they run in the CLI, the desktop app, the IDE extensions, claude -p and the Agent SDK.

• Pro: turn on the Dynamic workflows row in /config before ultracode appears. Pro also starts on the small size guideline, which tells Claude to aim for fewer than five agents per workflow; other accounts start on medium, fewer than ten. The guideline is advice to Claude, not a cap.
• Organizations: an admin can turn workflows off for everyone, in managed settings or on the Claude Code admin settings page, and that removes ultracode as well. No setting removes ultracode on its own, and an effort cap lowers the effort a session with ultracode runs at without turning it off.

A substantive request can chain several workflows, and every agent spends tokens

In Anthropic's example, a single request turns into several workflows in a row: one to understand the code, one to make the change and one to verify it. Each workflow runs up to 16 agents at once by default, and each agent runs on your session's model unless something else assigns one, such as a smaller model the script names for a stage. Anthropic's launch post says workflows can use substantially more tokens than a typical Claude Code session and suggests starting with a scoped task; on a subscription those tokens count toward your plan's usage and rate limits.

Turning ultracode on also switches off three safeguards, because Anthropic treats turning it on as consent to large runs:

• the Large workflow warning, shown when a workflow schedules more than 25 agents or its projected total passes 1.5 million tokens;
• the session's limit of 20 concurrent subagents for subagents Claude starts with the Agent tool;
• in auto permission mode, the approval prompt before the first workflow launch.

The workflow runtime's own caps, such as 1,000 agents per run, still apply. On a subscription, Fable 5.1 usage can bill usage credits instead of plan limits, depending on the plan and seat. On API billing the cost is direct, and the monthly API credit that comes with Max, Ultra and Team does not pay for interactive Claude Code sessions. As an illustration, not an Anthropic figure, and ignoring prompt caching: if each agent in a nine-agent workflow on Opus 5.5 reads 40,000 tokens and writes 2,000, the run costs about $1.80 at Opus 5.5's $4 input and $20 output per million tokens, where one such turn in conversation costs $0.20.

A run waits out a five-hour reset, not a distant weekly one

When an agent in a workflow hits your Claude usage limit, the run pauses instead of failing, and the waiting agents start again shortly after the limit resets. That only happens when every condition holds: an interactive session signed in with a Claude subscription, Claude Code v2.1.271 or later, the autoContinueAtUsageLimit setting on, a reset within 24 hours, and fewer than two earlier waits in the run. A run started with -p, the Agent SDK or in a background session does not wait, and neither does one whose limit resets more than 24 hours later, as a weekly limit can: those agents fail. Anthropic's error guide adds that one heavy burst, such as a large workflow fan-out, can use up the weekly allowance before the five-hour window resets. The /workflows view shows each agent's token use while the run goes, and you can stop the run there, usually keeping the work already finished.

When ultracode is worth the extra usage

Anthropic points workflows at work that splits into many parallel pieces: codebase-wide bug sweeps, migrations across hundreds of files, research that needs sources checked against each other, and plans worth drafting from several angles. For one task of that size, the keyword gets you a workflow without leaving the whole session on ultracode. For routine edits, run /effort ultracode off: while it is on, every substantive request is planned as one or more workflows.

On a subscription, the plan decides how far a long run gets. Pro starts on the small size guideline and shares one five-hour limit with your chats, while a verified Ultra account has no five-hour limit at all, which suits long runs that would otherwise pause at the five-hour reset. If your workflows regularly run past your plan, API billing has no five-hour or weekly limit and charges per token instead.