1M context is now generally available for Opus 4.6 and Sonnet 4.6

Original Article ↗ Hacker News Discussion ↗

Anthropic’s move to make a 1 million‑token context window standard for its Claude Opus 4.6 and Sonnet 4.6 models is being greeted as a major boost for long‑running coding and analysis workflows, especially multi‑agent setups and large codebases. Many developers report that bigger windows reduce painful manual compaction and enable overnight or multi‑hour autonomous work, but they also note persistent “context rot” beyond ~150–700k tokens, higher costs, and the need for careful orchestration, sub‑agents, and written plans. Overall, the change is seen as expanding what’s practically possible with AI coding assistants, while reinforcing that good prompting, session management, and model choice still matter more than raw context size.

Overall reaction to 1M context

  • Many are excited: fewer forced compactions, easier long-running coding/debugging and multi-hour agent workflows.
  • Others report that quality still degrades well before 1M tokens (often ~150–200k, sometimes ~600–700k) and that 1M is mainly useful for avoiding compaction, not for “using” the whole window intelligently.
  • Several say the “dumb zone” problem has improved in recent months but not disappeared.

Pricing, plans, and token usage

  • Key change: standard Opus/Sonnet pricing now applies across the full 1M context; previous “long-context premium” goes away.
  • Confusion and complaints around subscription tiers (Pro/Max/extra usage, 5x vs 20x plans, fast mode, 1M access on Pro).
  • Long sessions are very expensive: large contexts multiply per-call cost; people report burning through hundreds of dollars or hitting Max limits in minutes if they don’t manage context aggressively.
  • Prompt caching alleviates some costs but doesn’t remove them.

Context rot, compaction, and workflows

  • Broad agreement that “context rot” is real: models start forgetting design decisions, constraints, and earlier reasoning as sessions grow.
  • Compaction is widely seen as dangerous for Claude Code: can drop key steps, reintroduce solved bugs, or lose CLAUDE.md and other instructions.
  • Mitigations discussed:
    • Frequent intentional compaction via explicit summaries and fresh sessions (RESEARCH → PLAN → IMPLEMENT).
    • Using project-level memory files (CLAUDE.md, spec.md, goal.md, log.md, task.md) and reloading them into new chats.
    • Subagents/agent teams to keep the main orchestrator’s context small while workers run in fresh windows.
    • Editing/rewinding session history, or external tools to slice/summarize JSONL logs.

Comparisons with other models and harnesses

  • Some find Opus 4.6 clearly ahead; others report better long-context stability, compaction, and review quality from OpenAI’s Codex, or niche strengths from Gemini.
  • Several emphasize that harness quality (Claude Code vs Codex CLI vs third-party tools) matters as much as the base model.

Use cases and limits

  • Positive reports: large refactors, CI-driven agents, reverse engineering, emulator and game implementations, SEO/content generation, data analysis.
  • Negative reports: infrastructure work, tricky debugging, or highly constrained/low-level domains still often require strong human oversight.
  • Debate over claims that Opus 4.6 is “AGI”; many point to brittleness, looping, and architectural blind spots as counterevidence.