Anthropic launched Claude Sonnet 5.5 on Monday, introducing a model that generates output more than 30% faster than its predecessor. The release targets routine tasks requiring substantial reasoning, aiming to reduce operational costs for software development and workplace applications.

Efficiency gains without price cuts

The new model achieves cost savings of up to 30% per task not by lowering API rates, but by optimizing token usage. Input pricing remains at $2 per million tokens and output at $10 per million tokens. The reduction in expense stems from the model’s ability to complete tasks using fewer tokens and tool calls.

Performance benchmarks and capabilities

Sonnet 5.5 demonstrated significant improvements in coding and general reasoning. Anthropic reported a score of 1,811 on AA-Briefcase v1.1 and 61.6% on Chartography without tools. The company also noted that this is the first Sonnet version capable of beating "Pokémon Red" using screenshots alone.

Early testers included Daniel Vogel, chief operating officer at Epic Games, who highlighted the model’s proficiency in handling tens of thousands of lines of gameplay-system code. The launch follows the introduction of Claude Opus 5.5 on September 22, which was designed for more complex work and claimed to be 40% cheaper to run than its predecessor.

Security enhancements and future lineup

Anthropic integrated stronger cybersecurity safeguards into Sonnet 5.5, marking it as the first Sonnet model to launch with cyber protections and fallback measures similar to those in higher-end systems. This move aligns with the industry shift toward competing on the cost and safety of running long, multi-step coding agents.

The company plans to release Claude Haiku 5.5 in the coming weeks. This will complete a three-model 5.5 lineup spanning high-end reasoning, general-purpose work, and high-volume applications.

What the numbers show

The data reveals a strategic divergence between list price and effective cost. While API rates remain static, the 30% reduction in per-task cost indicates that Anthropic is prioritizing inference efficiency over price wars. This suggests that for enterprise clients with high-volume workloads, the total cost of ownership may decrease significantly even without explicit discounting, provided the model’s token efficiency translates directly to reduced billable usage.

A computational biologist has challenged Anthropic’s claim that its AI model, Claude, autonomously discovered a novel enzyme system. Mario Rodríguez Mestre stated that he and his colleagues had been studying the relevant ART enzymes for four years and shared key findings with Anthropic during consultations.

The dispute centers on whether Claude independently reasoned its way to the discovery or was guided by information derived from Mestre’s team’s unpublished research. The New York Times reported on Sunday that Mestre identified the enzymes in jumbo phages in 2022 while researching reverse transcriptases. He later consulted for ReNegade Therapeutics, which filed a 2023 patent covering several reverse transcriptases, including those he discovered, with Mestre listed as a co-inventor.

The core of the controversy

Mestre emphasized that the primary question is not whether ARTs were already known, but rather the origin of the specific insights attributed to the AI. "So, for me, the most important question is not ‘Were ARTs already known?’ They were," he told the publication. This raises significant questions about intellectual property and the definition of autonomous discovery in AI-driven science.

Anthropic did not immediately respond to requests for comment regarding these allegations. The company maintains that Claude’s discovery of a novel enzyme system had properties resembling CRISPR and was guided only at a high level by its scientists.

Anthropic’s account of the discovery

On Wednesday, Anthropic announced that Claude autonomously discovered a previously uncharacterized enzyme system in bacteriophages after analyzing a massive DNA database. The process involved nearly 950 Claude agents processing 210 million tokens over 21 hours before one flagged an unusual repeating pattern for human review. Scientists provided initial guidance and conducted subsequent laboratory testing.

While the function of the enzyme system is still under investigation, Anthropic stated it felt it was crucial to share these early findings to showcase Claude’s capabilities and provide insight into their ongoing work. Feng Zhang, one of the pioneers of CRISPR genome editing, called the discovery "genuinely intriguing" and said that it merits further investigation.

Broader life sciences expansion

This incident occurs as Anthropic aggressively expands its presence in the life sciences sector. Earlier this month, the company established a physical wet lab in the San Francisco Bay Area, expanding its AI-driven life sciences research into hands-on experimentation. Life sciences leader Eric Kauderer-Abrams said real-world lab work remains essential for biology, while the company combines internal research with external collaborations. Anthropic clarified that the facility is not specifically focused on drug discovery.

In the same week, Anthropic announced a collaboration with Novo Nordisk (NYSE: NVO) to accelerate the development of new medicines. Novo would use Anthropic’s advanced AI models to enhance software development and expand AI adoption, with strong data governance and human oversight to ensure responsible use.

In July, Anthropic launched Claude Science, an AI workbench designed for scientists, marking its significant expansion into life sciences. The platform supports fields including genomics, proteomics, structural biology and cheminformatics, integrates more than 60 scientific databases and tools, and connects with Nvidia Corp.’s (NASDAQ: NVDA) BioNeMo Agent Toolkit for access to specialized biology models.