Opus 5 ships with a 1M-token context window as both default and maximum, plus up to 128k output tokens, enabling multi-document synthesis, cross-file code refactors, and long-lived agent threads without frequent truncation. Pricing holds at $5/M input and $25/M output tokens, with a fast mode (Claude API only) providing higher output TPS at a premium. The model prioritizes deep reasoning, stable tool-use chains, and lower false-positive rates in code review. Lowering the prompt-cache minimum to 512 tokens unlocks more cache hits, reducing cold starts for shorter system prompts and templates you previously couldn’t cache.

Opus 5 activates thinking automatically and maps quality to an effort ladder: low, medium, high, xhigh, max. Because thinking consumes part of the total output budget, revisit max_tokens for tasks that were tuned on 4.8 without thinking. A good baseline is high effort with a generous max_tokens, then A/B against medium to reclaim throughput where quality holds. For hardest problems, xhigh or max often improves reliability, but only if you enlarge max_tokens to allow multi-step reasoning, subagent delegation, and multiple tool calls without truncation.

Opus 5 narrates progress more often, delegates to subagents more readily, and performs implicit verification. Remove legacy instructions like “add a final verification step” or mandatory verifier subagents that can now cause over-verification and latency spikes. Introduce task budgets to bound exploratory loops and help the model pace work across retrieval, planning, and tool use. For coding, expect higher completion rates on multi-file features and fewer placeholders. Instrument your tool calls to log depth, retries, and failure modes so you can correlate effort settings with stability and cost.

Mid-conversation tool changes (beta) let you update the available tool set without resending a fixed list each turn, preserving cache gains in long sessions. Use the specified beta header and test tool-add/remove flows that previously forced session resets. The default fallbacks mode (beta) selects Anthropic-recommended models by refusal category, reducing maintenance of brittle fallback lists. Pair this with monitoring to track refusal distribution and fallback hit rates. Together, these features shrink your orchestration surface area and improve reliability in helpdesk, RAG+actions, and build pipelines that evolve tools over time.

If your 4.8 workloads disabled thinking at high efforts, you must choose: keep effort at high or below, or re-enable thinking without the disable flag. With thinking disabled, expect rare tool-call text leakage or internal tags—sanitize outputs and consider post-parsing validation. Update your model ID, strip redundant verification boilerplate, and rerun evals with varied effort and max_tokens. Track content length growth (Opus 5 tends to write longer) and implement caps per channel. Finally, document which flows may use mid-conversation tool changes and fallbacks, and add observability to prove reliability before scaling.