Just over a week ago, the CEOs of Anthropic and OpenAI voiced support for slowing down AI development. Yesterday, both companies announced their new models at the same time.
While OpenAI launched GPT-6 Sol and GPT-6 Luna, which are not intended to become its new flagship models, Anthropic introduced Claude Opus 5.5, the first model in its new 5.5 family. In just a few hours, it had already claimed the top spot as the highest-rated AI in the Intelligence Index from the independent analysis and benchmarking platform Artificial Analysis. Specifically, the model achieved a score of 58, the highest ever recorded in this index to date.
According to Anthropic, Claude Opus 5.5 is positioned as an advanced model that brings major improvements in key areas such as performance, safety, communication ability, speed, and cost. In addition, this AI delivers performance similar to Claude Fable 5.1 across most tasks while costing 40% less to run than Claude Opus 5. Taking all of this into account, we could say this is the company’s new crown jewel.
Claude Opus 5.5 has already been rolled out across all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. Meanwhile, developers can access the model through the Claude Platform API.
As has become customary, Anthropic shared the performance results of its new model across different benchmarks, along with comparisons against both its own AIs and competing ones. In tests related to automated coding, computer use, and knowledge work, Claude Opus 5.5 posted the highest scores, even outperforming Claude Fable 5.1.
Source: Anthropic
The developer describes Claude Opus 5.5 as a “reliable and capable researcher”, one that delivers the “best performance we have tested to date.” Likewise, to ensure that this model’s intelligence advances are matched by the appropriate safety standards, Anthropic has equipped Claude Opus 5.5 with “the safety measures we have developed for our most advanced models.”
This improvement will not arrive alone, as Anthropic has confirmed that “in the coming weeks, the Claude Sonnet 5.5 and Claude Haiku 5.5 models will also reach the market, incorporating many of the same improvements in performance, efficiency, and safety.”
Claude Opus 5.5 represents a qualitative leap over its predecessor, Opus 5. According to Anthropic, “it is the new leading model, and early evaluators observed major performance improvements on their most complex tasks. One evaluator completed a 680,000-line code migration in less than a day, work that would have taken an engineering team weeks.”
This model also stands out in detecting and correcting software inefficiencies, and it is highly effective when creating graphics for video games, carrying out complex migrations, or conducting source code audits.
Anthropic has also optimized the efficiency of Claude Opus 5.5, managing to cut costs by up to 40% compared with its predecessor. The company has adjusted its API pricing to $4 per million input tokens and $20 per million output tokens (20% less), while highlighting a 60% reduction in cache read pricing ($0.20/M).
In addition to being 30% faster than Opus 5, it improves the user experience by extending usage limits to 5 hours on Pro, Max, Team, and Enterprise plans, and by including the option to save the usage reset.
Claude Opus 5.5 marks a step forward in natural communication, prioritizing key information and offering a fluid, intuitive style that makes review and ongoing collaboration easier on long-running projects. According to the first users who have tested it, the model stands out for its clarity and for producing more “human” writing.
Anthropic has significantly strengthened the safety of Claude Opus 5.5, applying “a classifier that analyzes each action before execution, an open-source sandboxed environment that security teams can audit, and code review that detects vulnerabilities before they are integrated.”
Against prompt injection attacks, Opus 5.5 outperforms its predecessor across every scenario tested (programming, browsing, and tool use) and matches Claude Fable 5.1 with the lowest vulnerability rate of all models evaluated, according to AI security firm Gray Swan.
In internal audits covering nearly 2,000 scenarios, it is Anthropic’s most honest and safest model to date. It stands out for reducing attempts to bypass limits by 85% compared with Opus 5 or Mythos 5.1, eliminating biases and risky behaviors in simulated environments, a key factor for autonomous enterprise deployments.
However, Anthropic warns that the model shows signs of detecting when it is being evaluated, which makes it harder to measure its real behavior in uncontrolled environments. To mitigate this, the company combines this alignment advance with multilayer defenses and a stronger focus on AI interpretability.
Anthropic has shared feedback from the first professionals who were able to test Claude Opus 5.5. It is a way of demonstrating this AI’s effectiveness with greater credibility by letting users themselves assess it.
Aleksandar Mitic, Senior Engineer at Spotify, said that “with Claude Opus 5.5, we observed a clear improvement in token efficiency in our internal evaluations, as we were able to complete the same tasks more affordably and more quickly.”
Meanwhile, Fabian Hedin, CTO and co-founder of Lovable, stated: “For Lovable developers, Opus 5.5 means faster builds with the same quality, whether they are starting from scratch or working on a production application. It gathers context only once, makes fewer but more comprehensive edits, and does not get stuck retrying, finishing in one-third to one-half fewer steps while using far fewer tokens in the process.”
Photo: generated with ChatGPT Images 2.5
Your email address will not be published. Required fields are marked *
Δ