Claude Thinks but Returns No Answer in Cherry Studio: How to Configure Max Tokens

On September 15, a user contacted BetterToken support regarding an issue in the Cherry Studio client: Claude answered simple queries normally, but on large analytical tasks, no final text appeared. The thinking block expanded, tokens were deducted, and the application displayed no errors.

Toggling streaming (Stream) did not resolve the issue. Looking through request logs in the BetterToken console, our team noticed a shared detail: several long responses routed across different channels ended at exactly 8192 output tokens.

Because the stop status (stop_reason) was not preserved in the logs, we cannot conclusively confirm why the request finished. However, the recurring 8192 value strongly pointed toward a likely output limit: the token ceiling may have been fully consumed during the reasoning phase, leaving no remaining quota for the actual response.

How the Case Unfolded

The core of the user’s inquiry can be summarized as follows:

Initial user message (paraphrase): Short prompts receive normal responses. On complex tasks, the model thinks for an extended period, tokens are deducted, but no final text appears—the answer field remains completely blank.

We recommended enabling and increasing the Max tokens parameter in the assistant’s settings, within the limits permitted by the provider.

Follow-up user message (paraphrase): After adjusting the limit in the assistant’s settings, the issue was resolved; the user also asked whether this parameter must be configured separately for each assistant.

In Cherry Studio, this configuration is indeed managed on a per-assistant basis.

Why the Response Disappears: How Thinking Works

In Claude models with reasoning capabilities, the thinking process counts directly toward the total generation budget.

According to Anthropic’s documentation on steering thinking and controlling costs, the max_tokens parameter sets a hard ceiling on total output per request. This limit encompasses both the internal reasoning tokens (thinking) and the visible response text. The effort parameter serves as a soft guide for reasoning depth, but it does not expand the overall token allowance. If the model exhausts the entire token quota during reasoning, output generation stops immediately. When a request halts due to reaching this threshold, Anthropic’s documentation recommends either reducing the reasoning effort or increasing max_tokens, provided the model and client interface support higher values.

Step-by-Step Configuration in Cherry Studio

According to the Cherry Studio chat documentation, settings apply across all conversations for the selected assistant. Update the Max tokens setting specifically for the assistant you are using. The setting applies strictly to that profile and does not modify parameters for other assistants.

Before making changes, verify your exact model identifier and the maximum output limits supported by your API provider.

Step 1. Open Assistant Settings

In the left sidebar displaying your assistants, locate the target profile, click the three dots icon (or right-click), and select Edit Assistant.

Case study illustration: Opening the assistant settings window via the Edit Assistant menu item.

Step 2. Enable and Increase Max Tokens

Navigate to the Model tab and locate the Max tokens option.

1. Toggle the switch next to the parameter to turn it on.
2. Enter a higher value than the previous limit, staying within the specifications supported by your model and provider.

Case study illustration: Enabled Max tokens setting with a value of 128000 on the Model tab.

In this specific case, the user set the value to 128000, and long responses began completing normally. Keep these practical considerations in mind:

• The 128000 value shown in the screenshot represents a case-specific configuration for this user, not a universal recommendation.
• This number defines the maximum output token limit for a single response, not the overall context window.
• Not every model supports generating such a large volume of tokens in a single request.
• Allowing a higher output ceiling lets the model reason longer, which can increase latency and total token consumption.

Step 3. Check Custom Parameters

Scroll down the Model tab to the Custom parameters section.

In Cherry Studio, custom parameters override standard UI toggle controls. If max_tokens is already defined in this list with an older value, delete it or update it to the new number. Otherwise, the client will continue sending the older limit in requests.

How to Verify the Result

Avoid testing the setting with brief prompts: simple queries comfortably fit within default limits and will not validate the fix.

1. Start a new topic under the same assistant to clear the interrupted conversation’s context.
2. Send a complex analytical task similar to the prompt that previously cut off.
3. Check for signs of healthy completion:
   • Complete, visible text appears directly below the thinking block.
   • The response concludes logically without cutting off mid-sentence.
   • If output token statistics are visible, check them against the previous cutoff to ensure generation did not stall at 8192 again. Note that a successful response does not have to exceed 8192 tokens—it may finish cleanly well below that threshold. Be sure to evaluate output tokens specifically rather than total token usage for the request.

What to Do If the Response Still Does Not Appear

If you adjusted the setting and the output still fails to appear, the issue may stem from other causes:

1. Verify the active assistant. Confirm that you sent the query from the exact assistant where the toggle was enabled, and verify that the Max tokens switch remains active.
2. Check your provider’s allowed limits. Setting a value higher than the model or provider supports will result in parameter validation errors.
3. Inspect your output token count. In the BetterToken dashboard, examine the request metrics and focus specifically on output tokens rather than the total token count. If generation ended well below the configured ceiling, investigate stream rendering (Stream), network stability, or external tool executions (such as MCP servers and functions), though max_tokens constraints should not be completely ruled out.

When contacting technical support, provide only safe diagnostic details: the Request ID, exact timestamp, model identifier, and number of tokens consumed. Never share secret API keys or confidential prompt content.