February 17, 2026
Pelicans, power bills & polite refusals
Claude Sonnet 4.6 System Card
Bigger brain, smaller bill? And why more 'no' in Hindi
TLDR: Anthropic says Claude Sonnet 4.6 is smarter and safer, shipped with mid-level safety guardrails. Commenters are grilling the team on speed and price, the energy used to train it, and why the AI refuses more in Hindi—highlighting that utility, costs, and fairness may matter more than shiny test scores.
Anthropic just dropped the system card for Claude Sonnet 4.6, boasting it’s smarter, safer, and inching close to their top model. But the crowd? They’re not clapping yet—they’re asking for receipts. One camp is chanting, “Benchmarks are cute, but is it faster and cheaper?” echoing a top comment that it’s “as good as the fancy one” but only if it runs quicker and costs less.
Another flashpoint: power. One user bluntly asked how much electricity it took to train this thing, sparking a mini-storm over energy bills and carbon guilt. Then came the curveball: a sharp-eyed commenter noticed the model says “no” more often in Hindi than other languages. Cue debates about whether some languages are more ambiguous—or whether the safety filters are just English-first. Translation for non-nerds: “refusal rate” is how often the AI declines a request; people are asking why that changes by language.
Amid talk of its ASL-3 “safety tier” (think: guardrails to avoid harmful stuff) and claims of the “best alignment yet,” the thread devolved delightfully when someone yelled, “Come on, pelican!” because of course it did. And yes, someone dropped the ever-handy models table like a scoreboard. Verdict: big upgrade, bigger questions.
Key Points
- •Anthropic’s system card details Claude Sonnet 4.6’s capability and safety evaluations across a wide range of tasks and domains.
- •Sonnet 4.6 shows substantial improvements over Sonnet 4.5 and approaches or matches Claude Opus 4.6 on several benchmarks.
- •Safety assessments indicate low overall misaligned behavior, with some alignment measures being the best observed in any Claude model.
- •The model was deployed under Anthropic’s Responsible Scaling Policy at AI Safety Level 3 (ASL-3).
- •Evaluations span coding, reasoning, multimodal, computer use, web/agentic tasks, long-context, multilingual, finance, cybersecurity, and healthcare.