Anthropic launches lower-cost Claude Opus 5 AI model

WASHINGTON: Anthropic on Friday introduced Claude Opus 5, a new artificial intelligence model that the company says delivers performance close to its most powerful system at roughly half the price.

The San Francisco-based startup, a major competitor to OpenAI, said Opus 5 leads industry benchmarks in coding and knowledge work while using fewer computing resources than earlier versions. The company framed the launch as a response to concerns among business customers over the high cost of operating advanced AI systems.

Claude is positioned against OpenAI’s ChatGPT and Google’s Gemini, with customers billed according to the amount of work completed in tokens. Anthropic said Opus 5 is priced at $5 per million input tokens and $25 per million output tokens, which it described as half the cost of Fable 5, its most expensive publicly available model. According to the company, the new model performs near Fable 5 across a range of tasks despite the lower price point.

The issue of cost has drawn greater attention after both Anthropic and OpenAI, the maker of ChatGPT, filed for stock market listings in June. The filings have put their commercial performance under sharper scrutiny from Wall Street ahead of their expected market debuts.

Safety and cybersecurity limits

On safety, Anthropic said it intentionally did not train Opus 5 on cyber tasks. It added that the model remains less capable than Mythos 5 — its most advanced and most tightly controlled system, which is available only to select partners — in both offensive cyber work and biological research capabilities that could be misused.

Anthropic described Mythos as its most advanced family of AI systems. The company said the series’ ability to identify weaknesses in computer systems and convert those weaknesses into functioning attacks has stirred controversy and led it to impose strict safeguards.

According to Anthropic, Opus 5 is about as effective as Mythos 5 at finding software vulnerabilities, but is far less capable of turning those flaws into attacks that could be weaponised. Citing that lower level of risk, the company said its automated safety filters will step in around 85 per cent less often on Opus 5 than on Fable 5, the Mythos-class model it released last month.

Fable 5 and Mythos 5 were temporarily removed from public access in June after the US Commerce Department introduced export controls. Those restrictions were lifted later that month.