Claude Opus 5: Anthropic Halves the Cost of Frontier Coding Intelligence

What Anthropic shipped

Anthropic released Claude Opus 5 on July 24, 2026, available now as claude-opus-5 on the API, the new default model on Claude Max, and the strongest model on Claude Pro. The headline isn't a new price tier — it's the same one. Opus 5 costs exactly what Opus 4.8 did: $5 per million input tokens, $25 per million output tokens — while Anthropic says it closes much of the gap to Fable 5 (priced at $10/$50) on several evals. As Simon Willison summarized it, it's a "thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price."

Why the price/performance move matters for builders

For anyone running agent loops in production — code review bots, deployment agents, research pipelines — the token bill scales with every planning step, tool call, and retry. Holding Opus-tier capability at Opus 4.8's rate, rather than pricing the upgrade at Fable-5 rates, changes the cost math for teams that were staying on 4.8 purely for budget reasons. There's also a new Fast mode: roughly 2.5x the default inference speed at 2x the base price. That's a genuine trade-off knob rather than a free upgrade — worth reaching for on latency-bound steps in an agent pipeline, not as a default.

Where Opus 5 actually leads

Anthropic's own benchmark set is the source here, so treat it as vendor-reported, but the specifics are concrete:

• Frontier-Bench: Opus 5 "more than doubles Opus 4.8's performance at a lower cost per task," per Anthropic's announcement.
• OSWorld 2.0 (computer-use tasks): outperforms every other tested model at any given cost, including Fable 5's best result.
• CursorBench 3.2: near-Fable-5 results at roughly half the token price.
• Other cited evals include GDPval-AA v2, ARC-AGI 3, and Zapier AutomationBench — a spread that skews toward agentic and workflow-automation tasks rather than pure chat quality.

The vulnerability-finding vs. exploiting gap

The detail most relevant to security-conscious teams sits in the cybersecurity numbers. On OSS-Fuzz, Opus 5's ability to *find* vulnerabilities comes close to Anthropic's own Mythos 5, a substantial jump from earlier Opus models. But it remains "substantially behind" Mythos 5 at *exploiting* what it finds — Anthropic notes Opus 5 wasn't trained specifically on exploitation tasks. For anyone evaluating how much autonomy to grant a coding or infrastructure agent, that finding-versus-exploiting split is worth tracking release over release: a model that gets sharper at spotting weaknesses without a matching jump in weaponizing them is a narrow but real signal, not a guarantee, when you're deciding what an agent gets to touch unsupervised.

None of this changes the fundamentals of running LLM agents near anything that moves money or ships code: cost caps, human review gates, and fail-closed guardrails still do the actual safety work. What Opus 5 changes is the economics of running more of that agent work at Opus-tier quality without a Fable-tier bill.