Anthropic Unveiled Claude Opus 4.5 — the “Most Powerful AI” for Coding and Agent Systems

• Anthropic presented Claude Opus 4.5, a model that outperformed all the people in the technical exam.
• It showed record results and reduced the cost of use.
• The company also updated the developer platform.

Anthropic has officially unveiled the new flagship generation of its artificial intelligence (AI) model, Claude Opus 4.5, which is already available in apps, via APIs, and on three major cloud platforms.

The company claims that it is the “best model in the world” for coding, agent systems, and complex workflows, as well as a significantly improved tool for everyday tasks, research, and working with presentations or spreadsheets.

The model is priced at $5/$25 per million tokens, making Opus “level capabilities accessible to even more users, teams, and enterprises,” the statement said.

Anthropic claims that Claude Opus 4.5 is the most consistent and secure model the company has ever released. The model is significantly more resistant to prompt injection attacks and is able to better resist malicious attempts to manipulate content.

According to the team, internal testers provided “impressively consistent” feedback before the release.

The company notes that Claude Opus 4.5 works effectively in conditions of uncertainty, is able to find solutions to complex technical multi-system problems on its own, and performs tasks that were unattainable for the previous Sonnet 4.5 model just a few weeks ago.

The latter was released in September 2025.

Early access customers also called the update a breakthrough. Windsurf CEO Jeff Wang said:

GitHub CPO Mario Rodriguez added:

Anthropic claims that Claude Opus 4.5 passed the company’s toughest internal test for performance engineer candidates, performing better than any human candidate within the two-hour time limit.

The company acknowledges that the test only measures technical skills, but the result has become an important signal of how AI can change the engineering profession. Anthropic’s Societal Impacts and Economic Futures research group continues to study the impact of AI on various fields.

Opus 4.5 demonstrates record-breaking performance in a number of benchmarks, significantly outperforming previous Anthropic models and competitors.

In some scenarios, the model demonstrates non-standard but correct solutions. For example, in the τ2-bench, where models are supposed to refuse to change an economy class ticket, Opus 4.5 instead found a legitimate way to improve the class of service first and then change the reservation. The benchmark counted this as a mistake, but the company called the behavior an example of “creative problem-solving.”

It is worth noting that a week ago, Google launched a new AI model Gemini 3.0 Pro, calling it “the most advanced model for complex tasks.”

Along with the launch of the model, Anthropic has updated its developer platform:

• The new effort parameter allows you to choose between speed, cost, or maximum power
• Opus 4.5 with medium effort is equal to Sonnet 4.5 on the SWE-bench, but uses 76% less tokens
• At maximum effort, it outperforms Sonnet 4.5 by 4.3 percentage points, using 48% fewer tokens
• Improved context and memory management increases the efficiency of agent tasks by ~15 percentage points

The model is also better at coordinating multiple sub-agents and can work in complex multi-step systems.

Anthropic claims that the updates make Opus 4.5 “a model for everyday work” in particular:

• Claude Code has an improved Plan Mode and is now available in the desktop app;
• Long dialogues in the Claude App are no longer “cut off” — the previous parts are automatically compressed;
• Claude for Chrome is now available to all Max users;
• Claude for Excel extended beta access to Max, Team, and Enterprise;
• Opus 4.5 usage limits for Max and Team Premium have been increased.

Earlier, we wrote that Claude 2.1 chatbot was released in November 2023, and the Claude app for iPhone and iPad was launched in May 2024.

In Europe, Claude was initially made available to certain users and businesses in May last year.