Claude Opus 5: Same $5/$25 Price, Cache-Safe Tool Swaps for Agents

Anthropic released Claude Opus 5 on July 24, 2026, and held the price exactly where Opus 4.8 left it: $5 per million input tokens and $25 per million output tokens, the same rate Anthropic has charged since Opus 4.5. A flat price on a new flagship isn’t really the story here. The story is a beta feature that lets you change what tools an agent can call mid-conversation without blowing away the prompt cache it’s been building.

Tool Changes That Don’t Cost You the Cache

On the Claude Platform, Opus 5 ships a beta that lets a running conversation change its available tools without invalidating the prompt cache built up to that point. Prompt caching is what keeps a long agent session affordable: the model doesn’t re-read the whole system prompt and tool list on every turn, it reads the cached version. Changing the tool set used to be one of the things that broke that cache and forced an expensive cold re-read.

Reality check: this is the kind of change that shows up in your bill, not your benchmark chart. An agent that starts with read-only search access and earns write or deploy access partway through a run is a normal pattern once you’re building anything more than a single-shot assistant, and until now, that handoff meant paying to rewarm the cache every time.

Automatic Fallbacks Instead of a Hard Block

The second beta lands on the API: when a request trips a safety classifier, Opus 5 can route it to an alternative model instead of refusing it outright. For a multi-step agent pipeline, a hard stop mid-chain is often worse than a redirected one, so a request that used to dead-end now has a documented path to still get an answer.

The Benchmarks Anthropic Put on the Page

Anthropic’s own comparisons run against Claude Fable 5, its Mythos-class model, and the prior Opus generation. On CursorBench 3.2, Anthropic says Opus 5 at max effort lands “within 0.5% of Fable 5’s peak score” at half the cost per task. On Frontier-Bench v0.1, Anthropic says it “more than doubles Opus 4.8’s performance.” On ARC-AGI 3, the score is “three times as high as the next-best model.” On OSWorld 2.0, a computer-use benchmark, Opus 5 beats Fable 5’s best result “at just over a third of the cost.”

Anthropic hasn’t published a context window figure for Opus 5. The prior flagship, Claude Opus 4.8, shipped with a 1-million-token window billed at the standard per-token rate; treat Opus 5’s window as unstated until Anthropic says otherwise. Fable 5 remains the more capable Mythos-class model overall; Opus 5 is the frontier-class agentic model at unchanged Opus pricing.

Pricing, Fast Mode, and Availability

Standard pricing stays at $5 per million input tokens and $25 per million output tokens. Opus 5 also keeps an optional fast mode, now priced at twice the standard rate for roughly 2.5x the speed, a straightforward trade for anyone who needs latency more than they need the cheapest possible token. Opus 5 is available as claude-opus-5 on the Claude API, and across claude.ai, Claude Code, and Claude Cowork.

What It Means for Developers

Same move Anthropic made with Opus 4.8: the marquee changes are operational, not a bigger context window or a leaderboard jump. If you’re already running agents that reconfigure their own tool access mid-run (which is most agents doing anything beyond a single tool call), the cache-preserving swap is worth building around immediately. It arrived a few weeks after Claude Sonnet 5, so Anthropic now has a real spread between a cheaper agentic default and a flagship that costs the same as it did two releases ago.