Only a 200K context window shows in actual use - Claude Code

Quick fix

Change the default model in ~/.claude/settings.json to the [1m] variant, or run /model claude-opus-4-8[1m] in the session, to activate the 1M context.

Symptom

Error outputCopy

在实际使用中只显示200K上下文窗口了

English translation

Only a 200K context window shows in actual use

Cause

At the heart of this: the model capabilities CC Switch declares do not match how Claude Code picks a model. CC Switch marks `supports1m: true` in the routing config but never synthesises a separate model ID with the `[1m]` suffix on the `/v1/models` endpoint. At the same time, the `model` field in a third-party provider response (DeepSeek, for example) is not mapped back to a Claude model name, so the Claude Desktop lookup fails. Claude Code ends up selecting the bare model without the suffix (`claude-opus-4-8`) and manages the context window against the 200K limit.

The cluster also holds two related variants under Codex: history reasoning content being wrongly counted into the context during Chat Completions protocol conversion, filling it up fast (#3867); and Codex judging a model non-official GPT once `model_catalog_json` model mapping is enabled, removing built-in editing tools such as `apply_patch` (#5947).

Fix

1. 1

   Session-level stopgap: in the current Claude Code session, run the command to switch to the logical model variant carrying the 1M marker.

2. 2

   Global default fix: edit the Claude Code user config file to point the default Opus model at the 1M variant.

   // ~/.claude/settings.json
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-4-8"
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-4-8[1m]"

3. 3

   Protocol fix for Codex users: if the context fills up fast or costs look wrong, give up on the generic CC Switch Responses conversion and configure the Anthropic protocol endpoint per the official DeepSeek Codex integration documentation.

Affected Versions

ToolClaude Code

VersionUnknown

PlatformsWindowsmacOS

Source Issues

This page is distilled from 5 real issues

• #3867Codex发送给 DeepSeek 的消息长度超过了模型允许的最大上下文长度。
• #4353Claude Desktop 模型未映射为1M上下文窗口
• #4525Claude Desktop 第三方网关模式下 DeepSeek 响应 model 字段未被反向映射，导致 Code 模式 /context 始终显示 200K
• #5947[Codex] 启用模型映射生成 model_catalog_json 后，apply_patch 等内置工具全部消失，Codex 审阅视图不触发
• #6192Codex app使用cc switch的deepseek v4 flash通用Responses配置异常导致实际费用翻倍

FAQ

I ticked 1M in Claude Desktop settings — why is Claude Code still 200K?

Ticking it in the Claude Desktop UI is not the same as what Claude Code selects for a session. Claude Code relies on a logical model entry with the `[1m]` suffix to activate the 1M budget, so the current session must have `claude-opus-4-8[1m]` selected rather than the bare model.

Codex says the maximum length is exceeded before the context reaches 1M. What now?

CC Switch counts the reasoning content a reasoning model produces into the history messages during protocol conversion. Switch to the Anthropic protocol endpoint, or configure per the official DeepSeek Codex documentation.

Why did apply_patch disappear after enabling model mapping in Codex?

That is upstream Codex behaviour: when it detects `model_catalog_json`, Codex decides you are on a non-official GPT model and stops delivering built-in editing tools such as `apply_patch`. For now it is a choice between showing the model name and keeping the built-in tools.

Related problems

Claude Official check failed: could not extract base_url - CodexCC Switch proxying and load balancing do not support an official Claude subscription — use a third-party API key. Forcing it errors and risks a ban.4 issues merged"is not a model this version of Claude Code recognizes" - CC SwitchSet CLAUDE_CODE_MAX_CONTEXT_TOKENS or append [1m] to the model name to define the correct context window size.3 issues mergedconfig.toml and auth.json still hold the previous third-party APIUpgrade CC Switch to v3.16.1+, restart, and switch Codex provider again. If official OAuth broke, run codex logout then codex login --device-auth.11 issues merged