AI: Argon, Haiku 5.5, and Large 4 Between Limited Access and Cost

The latest versions from Google, Anthropic, and Mistral are not playing the same game. On one side, there’s selective access or a controlled public preview; on the other, promises of significantly lower costs and better safeguards. All of this is accompanied by targeted performance increases, particularly in coding and multimodal reasoning.

Restricted Access at Google, Public Preview at Mistral

Last week, Google announced Gemini 4 Argon, a cutting-edge model currently reserved for a limited number of users. This model imposes an output limit of 1 million tokens and is only accessible to trusted cybersecurity workers through the Fairwind program. In contrast, Mistral AI launched Mistral Large 4 in public preview earlier this week, with an API available for testing. The public preview is set to conclude with an official deployment on October 27. These launches illustrate access strategies ranging from highly exclusive versions to models widely testable before general release.

Cost and Throughput: Haiku 5.5 Aims for Volume at Lower Prices

Anthropic is updating its entry-level range with Claude Haiku 5.5, the first major evolution in a year for its cheapest and fastest model. The company announces an operating cost approximately 75% lower than that of the previous version. Haiku 5.5 is designed for high-volume, cost-sensitive tasks, capable of reliably handling fast and repetitive workloads. It introduces an adjustable effort tuning to prioritize either cost or intelligence. This approach aligns with the so-called incremental updates focused on speed and cost.

Enhanced Safeguards and Protection Against Distillation

Anthropic reports improvements in alignment and cybersecurity for Haiku 5.5, with strengthened protections. Sonnet 5.5 also benefits from optimizations comparable to those of Opus 5 in these areas and includes better defenses against distillation attacks. These attacks involve extracting a model's capabilities to build third-party systems that lack the protections implemented by the publisher. On Mistral's side, Large 4 is presented as specialized in cybersecurity, geospatial and anchoring capabilities, as well as agentic AI.

Targeted Performance: Coding, Multimodal, and Open Weight Ranking

Regarding performance, Anthropic indicates that Haiku 5.5 shows notable improvements in computational tasks, knowledge management, agentic coding, and visual reasoning. Sonnet 5.5 is presented as a significant upgrade over 5.0, with notable increases in agentic coding and a particularly noticeable benefit in programming, as well as better collaboration capabilities. On Google’s side, Gemini 4 Argon also incorporates advancements in reasoning, coding, and multimodality, emphasizing enhanced deep reasoning abilities. According to Mistral, Large 4 ranks among the highest-performing open-weight models globally. Several of these advancements are significant and tend to approach the most advanced technological level.