Getting started with new Claude API features: long-output toggle, Workbench evaluations, and the cost dashboard

If you’ve recently been using the Claude API for summarization, coding, or generating long-form text, the most noticeable change is that it “can output more,” and the developer console is also more usable. This article breaks down several new Claude API features: how to enable long outputs, how to use the Workbench for prompt evaluation, and how to understand costs in the dashboard.

Claude API long output: Sonnet 3.5 increased from 4096 to 8192

The Claude API has increased the maximum output token limit for Claude Sonnet 3.5 to 8192, but it must be explicitly enabled. When calling the Claude API, add anthropic-beta to the request headers to enable the longer output window—useful for generating more complete reports, long code files, or multi-part summaries in one go.

The exact format is straightforward: add anthropic-beta: max-tokens-3-5-sonnet-2024-07-15 to the request headers. If you run into “output truncated” in the Claude API, first check whether you forgot this toggle and whether your max_tokens is set high enough.

A smoother Workbench: prompt generator and evaluation mode

In the Claude Console Workbench, the Claude API debugging experience has been strengthened with two key tools. The first is the “Prompt Generator”: you simply describe the task goal (for example, “classify incoming customer support requests”), and it produces a well-structured prompt draft that you can copy directly into the Claude API.

The second is “Evaluation Mode”: run two or more prompts side-by-side on the same batch of inputs, compare the outputs together, and even rate performance on a 5-point scale. For Claude API use cases that require stable output (support routing, information extraction, compliance rewrites), this step can significantly reduce guesswork in prompt tuning.

Usage and cost dashboard: accounting for Claude API costs clearly

With the new “Usage” and “Costs” tabs in the developer console, Claude API billing no longer has to be based on intuition. You can track consumption by USD amount, token count, and API key, quickly pinpointing “which key is burning money.”

It’s recommended to separate different environments into different keys (e.g., development/testing/production) and then use the dashboard to review peak periods. That way, if the Claude API ever has abnormal calls or looping requests, you can catch it before costs balloon.

More complete documentation and learning resources: release notes, courses, and the Cookbook

Claude API documentation has added more systematic release notes covering updates to the API, the Claude Console, and the Claude app—useful for teams to keep version change logs. At the same time, official courses on Claude API basics and tool usage have launched, helping newcomers quickly get up to speed on calling the API, structured JSON output, and more.

If you’re working on retrieval-augmented generation, citation/attribution, or classification tasks, you can go straight to the relevant guides in the Claude Cookbook to save time on repeated trial and error. Overall, the focus of this Claude API update isn’t on “flashy tricks,” but on solidly delivering long outputs, evaluation, and cost management.