# Context Window

> **TL;DR.** A [context window](https://www.

- **Category:** AI / Developer Tools / Infrastructure
- **Stage:** established
- **Age:** 338 days
- **Origin date:** 2025-06-30
- **First detected:** 2026-04-16
- **Canonical URL:** https://earlyterms.com/term/context-window
- **Sources:** 8 primary URLs

## Definition

A [context window](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents) is the span of tokens an LLM reads and reasons over in a single forward pass. In 2025-2026 the phrase shifted from a spec buyers compare (8K vs 1M) into an engineering resource agent teams actively budget, observe, and pay for per session.

The operational pivot was crystallized by three landmarks: Phil Schmid's [June 30, 2025 essay](https://www.philschmid.de/context-engineering) declaring context engineering the new skill, Chroma's July 14, 2025 [Context Rot](https://www.trychroma.com/research/context-rot) research showing non-uniform long-context decay, and Anthropic's March 13, 2026 move making 1M-token context GA on Opus 4.6 and Sonnet 4.6 at standard pricing.

## Example

The Red Beard's Feb 19, 2026 instrumentation traced 3,177 API calls across Claude Opus, Sonnet, GPT-5.3 Codex, and Gemini 2.5 Pro on one debugging task: context loads ranged from ~27K tokens (Opus) to ~258K (Gemini). Manus's agent team publicly designs around KV-cache hit rate as 'the single most important metric for a production-stage AI agent.'

## Analogy

Context used to be RAM on a spec sheet — a number you compared. Now it's cloud compute on an invoice.

## Why it's emerging now

Four forces pushed context window from a model spec into a managed resource in 12 months: 1M-token tiers went GA at standard pricing (Anthropic, Mar 13 2026), Chroma's 'Context Rot' quantified hidden decay, sub-agent fan-out made token spend unpredictable, and Karpathy's 'context engineering' vocabulary took hold.

## Related terms

- *alias:* 1M context
- *alias:* long-context
- *alias:* context length
- *parent:* context engineering
- *child:* context compaction
- *child:* context rot
- *child:* context budget
- *child:* context observability
- *related:* KV cache
- *related:* prompt caching
- *related:* Managed Agents
- *related:* Agent Harness
- *related:* Claude Agent SDK
- *related:* Coding Agents
- *competitor:* RAG

## Sources

1. [Anthropic — Effective context engineering for AI agents (Sep 29, 2025)](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents)
2. [Anthropic — 1M context GA for Opus 4.6 and Sonnet 4.6 (Mar 13, 2026)](https://claude.com/blog/1m-context-ga)
3. [Chroma Research — Context Rot: How increasing input tokens impacts LLM performance (Jul 14, 2025)](https://www.trychroma.com/research/context-rot)
4. [Phil Schmid — The new skill in AI is not prompting, it's context engineering (Jun 30, 2025)](https://www.philschmid.de/context-engineering)
5. [Manus — Context Engineering for AI Agents: Lessons from Building Manus (Jul 18, 2025)](https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus)
6. [The Red Beard — I intercepted 3,177 API calls across 4 AI coding tools (Feb 19, 2026)](https://theredbeard.io/blog/i-intercepted-3177-api-calls-across-4-ai-coding-tools/)
7. [Andrej Karpathy on X — +1 for context engineering over prompt engineering (Jun 2025)](https://x.com/karpathy/status/1937902205765607626)
8. [Hacker News — The RAG Obituary: Killed by agents, buried by context windows (290 pts, Oct 2025)](https://news.ycombinator.com/item?id=45439997)

---
_Generated by EarlyTerms · https://earlyterms.com/term/context-window_
