Claude Haiku 5.5, Mistral Large 4, and Local Models for Mac and Windows
The Batch Analytics DeepLearning.AI
Anthropic refreshed Claude Haiku and Sonnet 5.5, and Mistral previewed Large 4. The catch: the cheapest and fastest model isn’t always the cheapest per task, and reasoning can backfire.
Based on reporting by The Batch, Analytics DeepLearning.AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Andrew Ng’s main point lands cleanly: AI is improving much faster on tasks with a checkable right answer than on tasks that depend on judgment. Coding and math can be scored automatically. Customer complaints, product decisions, and business strategy can’t. That gap matters because the industry keeps acting as if success on one kind of task proves mastery of the other.
Anthropic’s new Claude Haiku 5.5 and Claude Sonnet 5.5 fit that split. Haiku 5.5 replaces Haiku 4.5 and is now the company’s fastest model, while Sonnet 5.5 replaces Sonnet 5 at the same per-token prices. Anthropic says Sonnet 5.5 is aimed at routine coding and office work, and it ships with adaptive thinking across five reasoning levels: low, medium, high, xhigh, and max. In practice, that knob matters a lot.
On Artificial Analysis’ Intelligence Index, Haiku 5.5 scored 43, putting it 13th among available models, and Sonnet 5.5 hit 56 at max reasoning, good for second place. But more reasoning is not a free lunch. On some tests, Sonnet 5.5 at max reasoning uses about 193,000 output tokens per task, more than any other model Artificial Analysis has measured, and Anthropic says it can cost more per task than Claude Opus 5.5 at max reasoning. On FrontierCode, Sonnet actually does worse at max than at xhigh.
Anthropic is also leaning harder into safety controls. The models are screened for risks tied to cybersecurity, biology, weapons, and frontier AI work. For cybersecurity, the block decision uses three pieces: a probe of internal activations, a classifier running on Sonnet 5.5 itself, and a separate language model trained for classification. Anthropic says Sonnet 5.5 is the first Sonnet model to launch with classifiers that block attempts to extract its reasoning without falling back to another model.
Mistral’s Large 4 is the other big release here. It is Mistral’s largest model so far, the most powerful released by a European company, and Mistral says the weights will arrive by the end of this month. The company says it was trained from scratch on 3,800 GPUs, with data spanning more than 160 languages, including all European Union official languages. It also says the model leads in finding and repairing a class of memory-safety bugs, which is the sort of claim that makes security teams sit up straight.
And the broader lesson is the same one Ng is pushing: benchmark wins on tidy tasks don’t magically turn into judgment. The useful habit for builders is boring but necessary—test each effort level, each cost point, and each task separately. The models are getting sharper. Human overgeneralization is still winning by a mile.
My take — AI-written commentary, not fact-checked reporting
The industry still loves pretending that a better math score means better judgment. It doesn’t. Anthropic and Mistral are both showing that the real product is control, not just capability—and that usually means more knobs, more caveats, and more honest benchmarking, which is less glamorous than the AGI pitch deck but far more useful.
Read more about this at: The Batch