Claude Opus 4.7 with 1M context: what changes for the exam

Anthropic shipped Opus 4.7 this week with a 1M-token context window (alongside the standard 200K). The exam blueprint hasn't moved, but a handful of scenario answers should.

What's new in 4.7

• 1M context as an opt-in window on Opus 4.7. Pricing tiers shift above 200K input.
• Sharper agent steering. Tool-call accuracy on the public agent benchmarks jumped again.
• Better long-context retrieval. Needle-in-haystack scores improve roughly linearly out to 800K.

What this changes on the exam

• Compaction questions still favor compaction — 1M doesn't mean "skip compaction." It means *the threshold at which compaction is required* moves later.
• "Choose the model" questions now have three production tiers: Opus 4.7 for hardest reasoning, Sonnet 4.6 for default, Haiku 4.5 for high-volume fast paths.
• Agent loops that previously failed at ~150K can now run longer without summarization. The cost calculus shifts at the breakpoint.

What this doesn't change

• Static-first prompt ordering still drives cache hits. Bigger context windows don't fix cache misses.
• You still want ~70% utilization headroom. 700K of 1M, not 999K of 1M.
• Tool design is unchanged.

The exam ID for this model is claude-opus-4-7. Memorize it — there's a multiple-choice question that hinges on knowing the exact identifier.