Claude Opus 5: 1M context at the old Opus price, and what it does to your RAG layer

Anthropic rolled out claude-opus-5 as the new default Opus on July 24, with providers switching on July 23 and Claude Code following on July 24. The specs are the kind that usually come with a price increase: 1M token context as both default and maximum, 128k max output, thinking on by default, a knowledge cutoff of May 2026, and a new "xhigh" reasoning level above the existing ones. The price is $5 per million input and $25 per million output, unchanged from Opus 4.8. Anthropic positions it close to Fable 5 in capability at roughly half the price, which puts it right on the line I drew in Fable 5 is back.

What actually changed

Three things, at three different layers of the stack.

The model is a straight swap for anyone on Opus 4.8. Same price, more context, more output, thinking on unless you turn it off. The xhigh level is new and is the one to be careful with: it raises both latency and tokens per task, so default versus xhigh is now an explicit cost-quality knob rather than something you inherit.

Claude Code changed underneath you at the same time. Opus 5 is the default Opus there, nested subagents and dynamic workflows were expanded, and the admin console gained two tabs, value and usage, showing active developers, session counts and top commands per day. Two new settings matter for anyone running agents on shared machines: sandbox.network.strictAllowlist and a DirectoryAdded hook. If your team's Claude Code sessions have been running since July 24, they have been running on a different model and a different orchestration model than the one you evaluated.

The seat default changed too. Sonnet 5 became the default for Pro, Team and Enterprise seats, with native 1M context and adaptive thinking on by default, at Sonnet prices. Your non-engineering users got a million tokens of context without asking for it.

What 1M at the old price does to RAG

Retrieval is a threshold problem. Chunking, embedding, a vector store and a reranker exist because stuffing the corpus into the prompt was either impossible or too expensive. At 1M tokens for $5 of input, a 300-page internal handbook fits in one call, and the question becomes whether the retrieval layer earns its maintenance cost.

A retrieval layer is a bet that the context window is too small or too expensive, and Opus 5 just changed both sides of that bet at once.

It does not kill RAG. It moves the line. Corpora under a few hundred thousand tokens that change rarely are now candidates for "load it all, cache the prefix" instead of "chunk and search". Corpora that are large, change hourly, or need per-user access control still need retrieval, and hybrid search for local RAG is still the right shape for those. Measure, per corpus, three things: total tokens, change frequency, and how often a typical answer depends on more than one passage. Small, static, multi-passage corpora are where full-context beats retrieval on both quality and engineering time.

What to do this week

• Pin model IDs where you need deterministic behavior. Default Opus now means Opus 5; anything relying on the alias silently upgraded on July 24.
• Re-baseline every agent pipeline against claude-opus-5 with thinking at default, then again at xhigh. Record cost per task and p95 latency for both and decide per route which level pays.
• Audit Claude Code settings on every developer machine and CI runner: enable strictAllowlist where agents touch the network, and wire the DirectoryAdded hook into whatever approval flow you already have.
• Revisit each retrieval pipeline with the threshold test above. Some will collapse into a cached prefix. Most will not, and that is fine.
• Check /benchmark for where Opus 5 lands next to Fable 5 and GPT-5.6 Sol on the tasks you care about before you move routing.

The honest gap

The "near Fable 5 at half the price" positioning is Anthropic's, and the MarkTechPost coverage and Claude Code changelog I am working from repeat it without independent numbers. The 1M context is a spec; whether Opus 5 uses the far end of it as well as the near end is exactly the long-context recall question vendor announcements never answer. And xhigh has no published cost multiplier, so until you measure it on your own tasks, budget for it as an unknown, not a small delta.