Claude Sonnet 4 now supports 1M tokens of context
TL;DR Highlight
Claude Sonnet 4's context window expanded 5x from 200K to 1M tokens, with a new tiered pricing model that doubles input costs above 200K.
Who Should Read
Developers using Claude API for large codebase analysis or multi-step agents, or developers using Claude Code who struggle with context management.
Core Mechanics
- Claude Sonnet 4 context window expanded from 200K to 1M tokens — roughly 75,000+ lines of code in a single context.
- Tiered pricing introduced: $3/MTok input and $15/MTok output up to 200K tokens; above 200K, input jumps to $6/MTok and output to $22.50/MTok. First non-linear pricing in the LLM industry.
- Concerns that longer context makes LLMs 'distracted' and reduces output quality — without evals proving accuracy at long context lengths, cost-to-value is hard to judge.
- Prompt caching is essential for managing costs with the larger context window.
Evidence
- Community skepticism about long-context claims — previous 200K context already had quality degradation reports, and 1M claims need evaluation data to be credible.
- The 2x price jump above 200K creates a natural incentive to optimize context usage and stay under the threshold when possible.
- Double-escape checkpointing in Claude Code helps manage context efficiently during long sessions.
How to Apply
- For agents handling large codebases, feed full source into context but always pair with prompt caching to reduce repeated call costs. If you can stay under 200K tokens, prices are half — so evaluate context pruning strategies first.
- In Claude Code, use double-escape checkpointing after filling context to manage long sessions efficiently.
- Consider the 200K threshold as a budget boundary — for many tasks, aggressive summarization and selective file inclusion can keep you in the cheaper tier.
Terminology
Related Papers
Training an LLM in Swift, Part 1: Taking matrix mult from Gflop/s to Tflop/s
Apple Silicon에서 Swift로 직접 행렬 곱셈 커널을 구현하며 CPU, SIMD, AMX, GPU(Metal)를 단계별로 최적화해 Gflop/s에서 Tflop/s 수준까지 성능을 높이는 과정을 상세히 설명한 글이다. 프레임워크 없이 LLM 학습의 핵심 연산을 밑바닥부터 구현하고 싶은 개발자에게 Apple Silicon의 성능 한계를 체감할 수 있는 드문 자료다.
Removing fsync from our local storage engine
FractalBits가 fsync 없이 SSD 전용 KV 스토리지 엔진을 구현해 동일 조건 대비 약 65% 높은 쓰기 성능을 달성한 설계 방법을 공유했다. fsync의 메타데이터 오버헤드를 피하기 위해 사전 할당, O_DIRECT, SSD 원자 쓰기 단위 정렬 저널을 조합한 구조가 핵심이다.
Google Chrome silently installs a 4 GB AI model on your device without consent
Google Chrome이 사용자 동의 없이 Gemini Nano 4GB 모델 파일을 자동 다운로드하고, 삭제해도 재다운로드되는 문제가 발견됐다. GDPR 위반 가능성과 수십억 대 기기에 적용될 때의 환경 비용 문제가 제기되고 있다.
How OpenAI delivers low-latency voice AI at scale
OpenAI redesigned its WebRTC stack to serve real-time voice AI to over 900 million users, detailing the design decisions and trade-offs of a relay + transceiver split architecture.
Efficient Test-Time Inference via Deterministic Exploration of Truncated Decoding Trees
Deterministic Leaf Enumeration (DLE) cuts self-consistency’s redundant sampling by deterministically exploring a tree of possible sequences, simultaneously improving math/code reasoning performance and speed.