Claude API Extended Output, Prompt Generator, and Usage Dashboard: A Walkthrough of the New Features

This Claude API update mainly addresses three things: longer answers, faster debugging, and more transparent costs. Below, following the most common developer workflow, we break down Claude API’s new features clearly and provide key invocation tips you can copy directly.

Claude API Extended Output: Sonnet 3.5 Maximum Output Doubled

In the Claude API, Claude Sonnet 3.5’s maximum output token limit has been raised from 4096 to 8192, making it suitable for longer summaries, code generation, and multi-step explanations. To enable extended output, you need to add the specified beta request header to your request.

The approach is straightforward: when calling the Claude API, add the request header anthropic-beta: max-tokens-3-5-sonnet-2024-07-15, and set max_tokens to the value you need. It’s also recommended to define a clear output structure (e.g., bullet points, paragraphs, JSON fields); otherwise, longer output won’t necessarily be easier to read.

Console Workbench Upgrade: The Prompt Generator Is Better for “Quick Drafting”

The Claude Console workbench now includes a prompt generator. You only need to describe the task in one sentence (e.g., “classify and handle inbound customer support requests”), and it will produce a more complete prompt template. For those who frequently write system instructions and need standardized output formats, this can save a lot of back-and-forth trial and error.

A more practical workflow is: first, let the prompt generator produce a “runnable” version, then add your real constraints—such as field validation, failure fallbacks, output length, and language style. Finally, paste the finalized prompt back into the Claude API’s system message or the first user instruction.

Evaluation Mode: Compare Prompts Before Using the Claude API

The workbench’s evaluation mode supports side-by-side output comparisons for two or more prompts, and lets you score results on a 5-point scale. It’s especially useful for A/B testing “same task, different wording”: for example, for the same field extraction task, one prompt may be more robust while another is more concise.

Before going live, it’s recommended to lock in a small fixed sample set: 3–10 typical inputs plus 1 extreme input, then use evaluation mode to select the most stable approach. This makes it less likely that after deploying with the Claude API you’ll run into format drift, missing fields, or overly long explanations.

Usage and Cost Dashboard: Make Claude API Costs Clear

The developer console now adds “Usage” and “Cost” tabs, allowing you to track consumption by USD amount, token count, and API key. This change is crucial for setups with multiple environments (test/staging/production) or multiple teams sharing the same account.

In practice, it’s recommended to split API keys by business line, and after each prompt change or max_tokens increase, return to the dashboard to compare token usage changes. Claude API costs are often not driven by the model itself, but are quietly amplified as prompts get longer and outputs get longer.

Release Notes and Learning Resources: Stop “Guessing Updates”

The official documentation now includes more complete release notes covering updates to the Claude API, the console, and applications, making it easier to check the reasons for changes and their impact scope. New courses and Cookbook guides have also been added, focusing on high-frequency capabilities such as citations, retrieval-augmented generation, and classification.

If you’re already using the Claude API for RAG, classification, or structured extraction, it’s recommended to treat the “release notes” as a routine checklist item: if interface behavior, limits, or recommended practices change, you can catch issues earlier than discovering them through production errors.