Three AI Labs, Four New Models: Anthropic, Google and Mistral Move

Anthropic says its new Claude Haiku 5.5 costs about 75% less to run than its predecessor, and that single figure frames a week in which three major AI labs shipped new models within days of each other. Google announced its frontier Gemini 4 Argon to a tiny group of trusted users. Mistral AI put its flagship Mistral Large 4 into public preview. The releases span the full spectrum, from bargain-tier workhorses to models most people can't touch yet.

Why does it matter? Because the competition has shifted. The fight is no longer only about benchmark bragging rights. It's about who can deliver speed, cost efficiency and security at scale, and who controls access to the most capable systems.

What's new in Anthropic's Haiku 5.5?

Haiku 5.5 is the first major update in a year for Anthropic's cheapest and fastest Claude model. Anthropic says it's designed for high-volume, cost-sensitive tasks and can reliably handle quick, repetitive workloads.

The headline number is the price. Compared with the previous Haiku model, Anthropic says the new version costs about 75% less to run — a steep drop for a model family already positioned as the budget option.

Anthropic also reports substantial gains across a series of benchmarks:

• Computer-use tasks
• Knowledge work
• Agentic coding
• Visual reasoning

Haiku 5.5 is the first Haiku-level model to ship with an adjustable effort setting, which lets users optimize for cost or intelligence depending on the job. Anthropic says the model has seen significant improvements in alignment and cybersecurity, with stronger safeguards built in.

How far does Sonnet 5.5 go?

One step up the ladder, Claude Sonnet 5.5 marks a clear upgrade over its 5.0 version. Like its smaller sibling, it's optimized for speed and affordability. Anthropic says there have been significant improvements to the model's performance and collaboration capabilities.

The security story is notable. Anthropic said Sonnet 5.5 has seen enhancements in both alignment and cybersecurity capabilities, making it comparable to those of Opus 5, a previous iteration of its big sibling. In plain terms: the mid-tier model now carries safety credentials once reserved for the flagship.

Sonnet 5.5 also posted fairly sizable boosts in benchmarks for agentic coding and knowledge work, and Anthropic says the performance jump is particularly noticeable in coding.

There's a defensive upgrade too. The model includes better protection against distillation attacks — attempts by bad actors to extract a model's capabilities and rebuild them without Anthropic's built-in safeguards. As frontier models get more capable and more valuable, protecting the weights and behaviors inside them has become a competitive concern in its own right.

Who gets access to Gemini 4 Argon?

Very few people, for now. Google announced Gemini 4 Argon last week as its latest frontier model, but it's currently limited to a small number of users — specifically, trusted cybersecurity workers enrolled in Google's Fairwind program.

The model ships with a 1 million-token output limit, enhanced deep reasoning capabilities and support for bigger projects. Like other Gemini models, it also brings improvements in reasoning, coding and multimodal capabilities.

The restricted rollout follows a pattern. Competitors like Anthropic and OpenAI have done the same with major model updates, releasing carefully to vetted users first before a broader launch trickles down to everyday users. For Google, gating a frontier model behind a cybersecurity trust program signals how seriously labs now take the misuse risks of their most capable systems.

What is Mistral's 'Le Chonk'?

Mistral AI took the opposite approach. Its latest flagship model, Mistral Large 4 — dubbed "Le Chonk" because of its size — went into public preview earlier this week, and anyone can try the preview API now.

According to Mistral AI, the new model ranks among the top open-weight models globally. It specializes in:

• Cybersecurity
• Geospatial and grounding capabilities
• Agentic AI

The public preview ends with an official rollout on Oct. 27, giving developers a fixed date to plan around.

What does the week tell us about the AI market?

Three labs, four models, three distinct strategies. Anthropic is pushing down the cost curve — a 75% running-cost cut on Haiku 5.5 — while pulling flagship-grade security into cheaper tiers with Sonnet 5.5. Google is holding its frontier model close, releasing Gemini 4 Argon only to trusted cybersecurity workers through Fairwind. Mistral is betting on openness, putting a top-ranked open-weight model into public preview ahead of its Oct. 27 launch.

Some of this week's releases are incremental updates in model speed and price. Others are more substantial upgrades that push close to the frontier. The common thread is that raw capability alone no longer wins the argument. Cost efficiency, safety hardening and access policy now carry equal weight in how these models compete — and in which ones deserve your attention.