Anthropic Releases Opus 5.5: Cheaper, Faster, but Signs of an AI Slowdown Emerge

Anthropic officially released Opus 5.5 on Tuesday, claiming its latest model sets a new record in coding performance and knowledge-based work. The model is also sold cheaper and faster than its predecessor, marking a strategic shift amid increasingly tight artificial intelligence competition.

Opus is the most advanced and most expensive tier in the three-level Claude lineup. Sonnet occupies the middle position, while Haiku is the fastest and cheapest option. According to Anthropic, Opus 5.5 is able to outperform the larger Fable model on a number of benchmarks, even completing informal tasks that Fable previously failed to finish.

On cost, Opus 5.5 output tokens are priced at US$20 per million tokens, down from US$25 in the previous version. A similar decrease applies to other metrics. The model also runs faster because the overall compute needed to serve it is reduced.

A significant change also occurred in how Opus communicates. Version 5.5 uses jargon less often and tends to place important information at the start of a message. This step answers a common user complaint about wanting direct answers without beating around the bush.

This launch is the first since CEO Dario Amodei openly backed the idea of slowing the pace of AI capability development to keep it in step with progress in alignment. In a piece earlier this month, Amodei stated his belief that fully addressing risk requires extra caution, not just investing in risk prevention, but also regulating the pace of progress so that mitigation efforts have time to catch up.

"I have become convinced that fully addressing risk requires greater caution, not only investing in risk prevention, but also regulating the pace of capability progress so that risk prevention has time to catch up." — Dario Amodei, CEO of Anthropic

Opus 5.5's safety training is broadly similar to its predecessor, covering alignment testing and pre-release evaluations by external organizations such as METR and Frontier Design. However, Anthropic stressed that more sophisticated training and evaluation systems are being prepared for future models, including improved security and monitoring.

The company also emphasized that Opus 5.5 has biological and cybersecurity capabilities equivalent to Mythos, so it is subject to the same safeguards as the Fable model. Those safeguards restrict the model's use for finding exploit vulnerabilities in compiled programs or developing recognizable biological weapons, among other tasks.

Indonesian Context

For Indonesia, the release of Opus 5.5 and Anthropic's strategic shift offer important lessons. On one hand, lower prices and increased inference speed can broaden access for local businesses, especially technology-based startups and MSMEs, to adopt advanced AI in their services. However, the "slow the pace" policy championed by Amodei is also a signal that the era of uncontrolled capability explosion may be starting to subside.

The Indonesian government, through the AI regulation currently being drafted, needs to examine the balance between encouraging innovation and ensuring safety. Anthropic's move to engage external organizations for independent evaluation could serve as an example of governance relevant to the national AI ecosystem, especially regarding model transparency and accountability.

Moreover, dependence on foreign models such as Claude, GPT, or Gemini remains a major piece of unfinished business. Without developing equivalent local models, Indonesia risks becoming merely a user market without control over the AI infrastructure and policies that affect its digital space.

Going forward, the question is no longer how quickly new models appear, but how ready the ecosystem and regulation are to absorb them. Will the acceleration of AI adoption in Indonesia be matched by an adequate safety framework, or will it fall behind innovation that keeps running ahead?