How much context coding agents really use
A week of agent traffic through MagmaRouter: the average request carried 117,905 tokens of context, one in ten went past 262,144 and the largest filled the full million.
A week of requests
In the seven days to 12 November 2026, MagmaRouter answered 60,126 requests. The average one sent 117,905 tokens of context. 36% sent more than 100,000, 10% sent more than 262,144 and 3% sent more than 500,000. The largest sent 1,048,122 tokens, a few hundred short of the most our models take.
Requests that size come from coding agents such as Claude Code, Cline and OpenCode, not from chat.
Why an agent needs so much
An agent sends the whole session again on every turn. That is the system prompt, the definitions of every tool it can call, each file it has opened and each tool result it got back, plus the new message at the end. Reading a dozen source files and running a test suite a few times passes 100,000 tokens well before the work is done, and none of it can be dropped without the model losing track of what it has already seen.
What happens at the limit
When a session outgrows a model's context, the request is refused outright. The agent then has to compact the history into a summary, losing the detail it was relying on, or the session ends. On last week's traffic a model with a 262,144-token window would have turned away about one request in ten, and those are the longest and most valuable sessions, the ones deepest into a task.
The average is the wrong number to plan around. Pick a model whose window covers your longest sessions, not your typical one.
Context on the line
Every model on the line reads close to a million tokens:
- GLM 5.3 Xploded: 1,048,576 tokens
- GLM 5.3 Flash Xploded: 1,048,576 tokens
- Kimi K3.1 Pro Xploded: 1,048,576 tokens
- Hy4 Xploded: 1,048,576 tokens
- DeepSeek V4.1 Flash 1017 Xploded: 1,048,576 tokens
- MiMo V2.6 Flash Xploded: 1,048,576 tokens
- Qwen3.8 27B Xploded: 1,010,000 tokens
Long context is cheaper than it looks, because most of each request repeats the one before it and is billed at the cached input price. On GLM 5.3 Flash Xploded that is $0.20 per million cached tokens against $0.30 for new ones.
More from the blog
- Most of an agent's input is a cache hit
- Kimi K3.1 Pro, uncensored, is on MagmaRouter
- How a model gets into Xploded
- Why most uncensored models cannot run an agent
- Which Xploded model to use
- Introducing Xploded
- Run Claude Code on an uncensored model
- A spend cap that cannot be overrun
- How MagmaRouter compares, honestly