Claude Sonnet 5.5 Announced: Faster, More Affordable and Top of the Class in Coding

As competition in autonomous software development and reasoning continues unabated in the AI ecosystem, Anthropic has introduced its eagerly awaited new-generation language model. Claude Sonnet 5.5, the second member of the Claude 5.5 model family, makes a huge leap in performance compared to the previous-generation Sonnet 5 while promising developers higher speed and a clear cost advantage.

Positioned alongside Claude Opus 5.5, the series' most capable flagship with the deepest reasoning, Sonnet 5.5 was optimized for daily professional workflows, codebase maintenance, debugging and multi-step autonomous tasks. Announced ahead of Claude Haiku 5.5, which will join the series in the coming weeks, the model delivers results that surpass even flagship models, especially in software engineering benchmarks.

A Record Leap in Coding and Agent Skills

The most striking aspect of Claude Sonnet 5.5 was the results it achieved in terminal and agent-based coding tests. On the Terminal-Bench 4.0 benchmark, which measures the ability to complete complex multi-step tasks in a command-line interface (CLI), the model reached a success rate of exactly 70.6%. Considering that its predecessor Sonnet 5 stayed at 10.3% on this test and the family's largest model Opus 5.5 scored 66.4%, the maturity Sonnet 5.5 has reached in autonomous terminal management stands out.

The model's success in software development is not limited to synthetic command-line tests:

• FrontierCode 1.1: On this test, which measures whether code written by an AI agent can be merged directly into a real project, the model achieved 52.1% at the Xhigh effort level and 46.2% at the Max level.
• CursorBench 4.0: On multi-file tasks with ambiguity compiled from real Cursor development sessions, it scored 55.5% and took second place, right behind Opus 5.5.
• OSWorld 2.1: In autonomous scenarios involving computer and operating system use, it reached a partial success level of 80.1%, setting a new standard in interface interaction.
• Visual Perception and Gaming Achievement: Proving its long-horizon task tracking and multimodal capabilities, the model became the first Sonnet model to successfully complete the popular game Pokémon Red from start to finish by analyzing only screenshots.

Teams that took part in early access testing, such as Epic Games, Unity, Slack, CodeRabbit and Lovable, stated that the model's speed in understanding codebases and its architectural design ability are on par with top-tier models. In particular, the model's habit of batching tool calls together to avoid unnecessary steps significantly reduces waiting times in software development workflows.

30% Faster and 30% Lower Cost

With Claude Sonnet 5.5, Anthropic managed to combine the performance gain with cost savings. The model's list token prices were kept exactly the same as Sonnet 5:

• Input Tokens: $2.00 per million tokens
• Output Tokens: $10.00 per million tokens
• Cache Reads: $0.20 per million tokens
• Cache Writes: $2.50 per million tokens

Although the price table stays the same, the efficiency in Claude Sonnet 5.5's thinking and problem-solving architecture lets it complete the same task using far fewer tokens than Sonnet 5. In measurements, the model's cost per task comes out about 30% lower. On top of that, the more than 30% speed increase in output generation provides a smooth experience in real-time developer interactions.

Users can adjust the effort level in Claude Code and the API to customize the balance of speed, cost and deep thinking for their projects. Even at low and medium effort levels, Sonnet 5.5 manages to beat Sonnet 5's scores at its highest effort level at one tenth of the cost.

Knowledge Work and Design Competence

The model's capabilities go far beyond coding and extend into multidisciplinary analytical tasks. Scoring 1844 Elo on the GDPval-AA v2.1 evaluation, which tests real-world scenarios from 44 different occupational groups, Sonnet 5.5 sits right behind Opus 5.5 (1846 Elo) while beating Sonnet 5 (1449 Elo) by about 400 points. Similarly, it reached 1811 Elo on the AA-Briefcase test, which measures long-horizon workflows.

There are also clear improvements in text quality and presentation design. The model can analyze complex financial reports, earnings tables and call transcripts and directly produce presentation templates ready for sharing. Successfully capturing aesthetic details in interface design, the model builds a cleaner component hierarchy in front-end development.

Security, Cyber Safeguards and Preserved Thinking

Because of the high reasoning capability it offers, Sonnet 5.5 became the first Sonnet model equipped with the advanced Cyber Safeguards previously applied only to Opus-class models. While routine software development and vulnerability patching work runs without interruption, for high-risk cybersecurity requests the model falls back to Sonnet 5 mechanisms in a controlled way.

In addition, the Preserved Thinking infrastructure was expanded to prevent the model's reasoning chain from being stolen and copied by malicious actors using fake accounts. Thanks to this technology, the thinking steps Claude produces are securely associated with the original user account that called the model.

Availability and Integration

Claude Sonnet 5.5 is available to developers and enterprises starting today:

• Claude Platform: Accessible through the API directly with the claude-sonnet-5-5 model identifier.
• Cloud Providers: It has taken its place in the Amazon Web Services (AWS Bedrock), Google Cloud (Vertex AI) and Microsoft Azure cloud catalogs.
• Zero Data Retention: In line with enterprise privacy standards, data is processed without being stored.
• Developer Settings: For developers who use it with thinking mode turned off, a new between_tools configuration was introduced that optimizes thinking between tool calls.

Claude Sonnet 5.5 stands out as 2026's most balanced and powerful model for software teams that want access to flagship-class AI power without straining speed and budget constraints.