Anthropic released Claude Opus 5 on July 24, 2026, and the pitch is agentic economics: a standardized 1M-token context window (default and maximum), a 128k-token maximum output, and Thinking enabled by default — all priced at $5 per million input tokens and $25 per million output tokens, consistent with Opus 4.8 and roughly 50% cheaper than Fable 5. The prompt-cache minimum drops to 512 tokens, and Anthropic says Opus 5 outscores Fable 5 on most major reasoning and coding benchmarks.
Opus 5 is positioned as a break from conversational iteration: Anthropic frames it as the transition point from “chatbots” to “autonomous agents,” engineered for deep reasoning and long-horizon task execution. The release targets developers who need one model to hold an entire codebase, case file, or documentation set in a single pass — and to act on it with tools rather than merely describe what to do.
The quietest but most operationally significant change is standardization: unlike earlier models with varying context lengths, every Opus 5 request runs against the same 1M-token window. For RAG pipelines and large codebase ingestion, that removes an entire class of capacity planning. Pair it with the 128k-token output ceiling and Opus 5 can emit complete technical documents or substantial code in a single generation, instead of stitched-together chunks.
Thinking is no longer an opt-in nicety — it is the default path, letting the model spend test-time compute working through problems before answering. Two integration-level changes come with it: support for mid-conversation tool changes, and a breaking change in when thinking can be disabled, which gives more granular control over model behavior but will require attention from anyone migrating existing pipelines.
⚡ Migration note: the prompt-cache minimum of 512 tokens means existing implementations with reused system prompts or large documentation context can see immediate cost benefits without code changes — but the thinking-disable breaking change should be checked before upgrading production agents.
Anthropic’s comparative testing puts Opus 5 ahead of Fable 5 on most major reasoning and coding benchmarks, at roughly half the price. Against Claude Opus 4.8, the gains concentrate where multi-step logic lives: SWE-bench-style tasks show Opus 5 completing multi-file features and large-scale refactors without leaving stubs or placeholders — a documented failure mode of earlier generations. Its code reviews carry minimal false positives, which Anthropic positions as making the model viable for automated CI/CD gates.
| Specification | Claude Opus 5 | Notes |
|---|---|---|
| Context window | 1M tokens | Standardized default & max |
| Max output | 128k tokens | Single-pass documents |
| Input price | $5 / MTok | Same as Opus 4.8 |
| Output price | $25 / MTok | Same as Opus 4.8 |
| vs Fable 5 cost | ~50% cheaper | Frontier tier, half price |
| Prompt cache minimum | 512 tokens | Lowered threshold |
For agent builders, the release notes emphasize stamina: Opus 5 stays on task across extended tool-use loops, delegates work to subagents, and narrates its progress as it goes — making the internal reasoning process transparent to the orchestrator. The vision stack supports UI replication and diagram analysis, and the model can turn a screenshot into functional, high-fidelity frontend code.
Multi-agent orchestration gets specific attention. In writer–verifier setups, Anthropic reports a marked reduction in cases where subagents overwrite each other’s work — the classic failure that turns two cooperating agents into one corrupted output. Coordinated outputs come out cleaner and more coherent as a result.
This technical summary was compiled exclusively using verified public data points from the following release documentation: