Claude Opus 5 Leads on Computer Use and Undercuts Fable 5 on Price

Anthropic released Claude Opus 5, a new top-tier model built to run long, multi-step agents, and it lands at the same price as the previous Opus model. Opus 5 is available now across Claude.ai, the Claude API, Claude Code, and Claude Cowork, under the API name claude-opus-5.

The pitch is agentic work at a lower cost than Anthropic's frontier tier. The company says Opus 5 comes close to the intelligence of its flagship Claude Fable 5 at half the price, and posts standard rates of $5 per million input tokens and $25 per million output tokens, matching Opus 4.8. A fast mode runs at 2.5 times the speed for twice the base price.

Anthropic says Opus 5 beats every other model on computer use

Anthropic positions Opus 5 as a step up for agents that operate software on their own. The company reports the model outperforms every other model on OSWorld 2.0, a benchmark for computer use, at just over a third of Fable 5's cost per task. It describes Opus 5 as a proactive model that verifies its own work and iterates before returning a result, the behavior that matters most when an agent runs unattended for long stretches. VP Land readers have followed this thread since Anthropic first gave Claude the ability to control a computer.

On agent-style evaluations, Anthropic's claims are comparative and should be read as the company's own numbers. Anthropic says Opus 5 more than doubles Opus 4.8's performance at a lower cost per task on its internal Frontier-Bench v0.1, scores three times higher than the next-best model on ARC-AGI 3, and lands within 0.5% of Fable 5's peak score on CursorBench 3.2 at maximum effort, at half the cost per task. On science tasks, the company reports gains of 10.2 percentage points on organic chemistry and 7.7 points on protein sequence prediction.

Customers cited in the announcement echo the cost-to-capability framing. Zapier CEO Wade Foster said Opus 5 topped the company's AutomationBench leaderboard. Lovable co-founder Fabian Hedin reported a 22% gain over Opus 4.7 on agentic coding, with less variance run to run.

New agent controls: mid-conversation tool swaps and automatic fallbacks

Two beta features target developers building agents. Opus 5 can change the tools available to it mid-conversation without invalidating the prompt cache, so an agent can pick up new capabilities without paying to reprocess its context. A separate beta on the API automatically routes flagged requests to an alternate model. Anthropic also says the model can manage parts of its own memory in production, aimed at monitoring agents that run over long horizons.

Anthropic calls it its most aligned model to date

Anthropic describes Opus 5 as its most aligned model so far, citing the lowest rates of deceptive behavior on an automated behavioral audit and a score of 2.3 on overall misaligned behavior, the lowest of its recent models. The company says its cybersecurity classifiers intervene 85% less often than with Fable 5, and that flagged requests on Claude.ai fall back to Opus 4.8 by default. Anthropic notes Opus 5 stays behind its Mythos 5 model on offensive cybersecurity and biology tasks, while calling it the most capable generally available model for scientific research.

What Opus 5 means for creative pipelines built on Claude

Opus is the tier that already sits under agentic creative tooling, so the practical read for production teams is cost and reliability, not a new interface. We covered how a single image gets turned into a full 3D scene with Claude Code as the orchestration layer coordinating other services, the kind of multi-step job where a cheaper, more consistent agent model changes the math.

The same logic applies to the growing set of Claude-plus-MCP integrations inside creative software. At NAB, we reported on Claude's push into creative tools through agentic workflows and MCP servers. If Opus 5 holds up to Anthropic's cost and consistency claims outside its own benchmarks, the teams wiring Claude into editing and asset pipelines get more runway per dollar without swapping models.