Claude Code's New /claude-api hillclimb Command
Claude CodeAnthropic's September 8, 2026 blog post on reducing Claude Platform costs introduced /claude-api hillclimb, a new Claude Code command that automates the search for a cheaper model, effort level, and prompt configuration without sacrificing accuracy. The tool splits a user-supplied evaluation into train and test sets, proposes configuration changes, reads failing examples to refine the prompt, and reports a final configuration scored on held-out data. In Anthropic's own customer-support benchmark, hillclimb took a setup running Opus 4.8 at high effort and found a Sonnet 5 configuration at low effort that scored 90.5% on held-out tickets, up from 78.6% for the original setup, at roughly one-fifth the cost. The command joins two existing claude-api skill commands, prompt-audit and cost-optimize.
Key Takeaways
- hillclimb is documented in Anthropic's blog post rather than the published CHANGELOG.md, making the blog the primary record that Claude Code added it.
- The tool automates a three-way tradeoff between model choice, effort level, and prompt content, searching for the cheapest configuration that holds accuracy steady.
- A customer-support benchmark example improved held-out accuracy from 78.6% to 90.5% while cutting cost to roughly one-fifth of the original setup.
- hillclimb builds on two existing commands,
prompt-auditandcost-optimize, chaining their outputs into a single automated search loop. - The search guards against overfitting by scoring its final configuration on a held-out test set the optimizer never saw during tuning.
- No additional install is required; the command ships inside Claude Code's bundled
claude-apiskill, available to any developer building on the Claude API.
Sources & Mentions
1 external resource covering this update
A New Command for Automated Cost-Performance Tuning
On September 8, 2026, Anthropic published a blog post, "Reducing cost and improving performance with Claude Platform," describing how development teams can cut Claude API spend without sacrificing quality. Documented in the post is a new Claude Code command that does not appear in the published CHANGELOG.md: /claude-api hillclimb, added to Claude Code's built-in claude-api skill.
How Hillclimb Works
Where cost-optimize profiles an existing codebase's Claude API spend and prompt-audit flags outdated prompting patterns, such as verification rituals and mandatory procedures written for older models, hillclimb goes a step further by running an actual search. Given an evaluation, it splits the test cases into a training set and a held-out test set, then works through model, effort, and prompt configurations, reading the training examples that fail at each step to decide what to change next.
A Worked Example
Anthropic demonstrated the tool on an internal customer-support benchmark. Starting from an Opus 4.8 baseline running at its default high-effort setting, hillclimb first tried Opus 5 at low effort combined with prompt-audit's cleanup of mandatory tool-call rituals and scratchpad steps, matching the original 98.9% training accuracy at a fraction of the cost (2.6 cents per ticket). It then tried stepping down further, to Sonnet 5 at low effort, which cut cost again to 1 cent per ticket but dropped accuracy to 88.9%. Reading the newly failing tickets, the tool added routing rules and a refund-cap cross-reference to the prompt, recovering the original 98.9% accuracy at the cheaper price point. On 14 held-out tickets the search never saw during tuning, the final configuration scored 90.5%, against 78.6% for the original setup, at about one-fifth the cost.
Part of a Broader Cost Toolkit
Hillclimb joins prompt-audit and cost-optimize as the third command Anthropic has added to the claude-api skill specifically for managing API spend, and the blog post frames all three as complementary: cost-optimize finds where the money is going, prompt-audit removes waste from prompts written for older models, and hillclimb automates the harder judgment call of trading off model choice, effort level, and prompt content against measured accuracy. Because the feature lives inside Claude Code's bundled claude-api skill rather than as a separate product, any developer using Claude Code to build on the Claude API can invoke it directly with no separate installation.