AI: A single prompt alone consume almost all of the 128k tokens of gpt5/sonnet4

[UdittLamba]

Summary

Previously, a single 128k context lasted for multiple prompts, but now it is a single response for a single prompt that consumes most of the tokens. Continuing to a new thread also doesn't;t help, as the "new from summary" is bugged out as well

Description

Steps to trigger the problem:

1. Select sonnet 4 or gpt 5
2. Enter a prompt
3. Often token consumption would hit 100k plus within the return of first response

Expected Behavior: single context should last for multiple complex prompts
Actual Behavior: it barely lasts a single prompt after the last update

Model Provider Details

• Provider: Anthropic via API key, Copilot Chat
• Model Name: Sonnet 4, GPT 5
• Mode: Agent Panel
  Other Details (MCPs, other settings, etc): none

Zed Version and System Specs

Zed: v0.201.8 (Zed)
OS: Linux Wayland manjaro unknown
Memory: 31.1 GiB
Architecture: x86_64
GPU: AMD Radeon RX 7900 XTX (RADV NAVI31) || radv || Mesa 25.2.1-arch1.4