OpenAI reduces Codex Model Context Size from 372k to 272k

Original Article ↗ Hacker News Discussion ↗

OpenAI has temporarily reduced the context window for its Codex coding agent from 372k to 272k tokens, prompting debate over cost savings versus capability for large, long-running software projects. Many users report that aggressive, opaque “compaction” of conversation history and code context can derail complex workflows, especially when working across big codebases, multiple repositories, or large rule files, while others argue that good planning, subagents, and external markdown “memory” largely offset smaller windows and avoid quality loss at very long context lengths. Comparisons with Anthropic’s million-token models and DeepSeek-style caching highlight a broader tension: whether frontier tools should prioritize massive raw context or smarter, more controllable context management and harness design.

Context Size Reduction & Rationale

  • Codex’s context window was reduced from 372k to 272k tokens; commit and tweets suggest it’s a temporary cost/usage control, not a capabilities change.
  • Some point out this avoids triggering higher-priced long-context tiers that previously hit users unexpectedly.
  • One comment notes the underlying models can handle up to 1M context, but Codex’s client limits are lower to keep total (input + max output) under a billing threshold.

Understanding the Cost/Trajectory Chart

  • Several readers found the posted cost chart confusing; others explained it as cumulative cost rising roughly quadratically with turn count until compaction, with smaller context (e.g., 200k) giving a consistently cheaper trajectory than larger (300k).
  • There is disagreement about whether the curve reflects quadratic attention or just cumulative linear costs.

Impact on Real Workloads

  • A sizable group says 272k is too small for big codebases, reverse engineering, or workflows with many plans, reviews, and long-running agents; they often sit near the limit and get frequent compactions.
  • Others argue most problems can and should be decomposed into chunks under ~200–300k, with smaller contexts yielding better quality and lower cost.

Compaction: Quality, Control, and Workarounds

  • Many complain Codex’s auto-compaction:
    • Triggers around 10–20% remaining context with no way to disable or roll back.
    • Sometimes causes the model to “forget” recent tasks or previously read code, forcing re-scans and burning tokens.
  • Others report Codex compaction has improved significantly since earlier releases and works fine for them.
  • Common strategies to mitigate:
    • Use markdown “plan”/“rules”/“report” files as durable memory instead of relying on chat history.
    • Aggressively use subagents, hierarchical planning, and external tools that store and re-inject prior context.
    • Manually restart sessions or clear context around 200–300k.

Comparisons to Other Models & Harnesses

  • Several users say they’ll stick with or move to models with ~1M context (e.g., Anthropic, DeepSeek, others), despite similar degradation past a few hundred thousand tokens.
  • Others claim long-context marketing overstates real usable context; they see a “dumb zone” starting around 120–300k on many models.
  • Some prefer alternative open-source or third-party harnesses with tunable compaction, history trees, and stronger user control.

Safety & Harness Design

  • The same commit adds stricter system-prompt guidance before destructive actions (e.g., don’t recursively delete home or root directories), responding to real cases of accidental mass deletion.
  • There is support for further isolation (containers, guarded rm wrappers) and skepticism about letting agents run destructive commands at all.