Anthropic brings Claude 3.5 Haiku to web and mobile apps

The gap between lightweight speed demons and yesterday's flagship models is officially shrinking to zero.

Anthropic just rolled out Claude 3.5 Haiku to all users across Claude on the web and mobile apps. Reports started trickling across social media on Thursday morning before being confirmed, giving everyday users direct access to the company's newest lightweight model.

Fast and sharp, but missing its vision

While Anthropic initially unveiled 3.5 Haiku in November and brought it to their API last month, the consumer-facing chatbot was left waiting until now.

The headlining stat here is raw intelligence for its weight class. On specific industry benchmarks, Anthropic says 3.5 Haiku matches or even beats Claude 3 Opus, which was their top-tier flagship model just earlier this year.

Anthropic points to coding recommendations, data extraction, labeling, and content moderation as its sweet spots. Compared to the older 3 Haiku, it outputs longer chunks of text and features a more recent knowledge cutoff date, meaning it actually knows about recent events.

There is one glaring catch, though. The model does not support image analysis at all. That leaves it behind not just the powerhouse 3.5 Sonnet, but strangely behind the original 3 Haiku as well.

The ghost of the API price hike

This web and mobile debut follows a rocky introduction for the model among developers.

When 3.5 Haiku hit the API early last month, it sparked controversy across developer channels. Anthropic had initially hinted that it would cost the same as the rock-bottom 3 Haiku, but later walked that back and bumped the price up, arguing that the massive jump in intelligence justified a steeper bill.

While developers had to pull out their calculators, everyday Claude chat users do not have to worry about per-token billing, making its arrival on web and mobile a straightforward speed upgrade.

Why it matters

For regular users and developers looking for a fast brainstorming partner, 3.5 Haiku is immediately worth playing with. If your daily workflow revolves around fast code reviews, drafting messages, or parsing dense walls of text, the instantaneous latency is tough to beat. But if you rely on dropping screenshots, receipts, or diagrams into Claude, you will need to stick with 3.5 Sonnet.

Across the industry, this release proves that last season's flagship reasoning is now the baseline for budget models. Every lab is racing to make sub-second models competent enough for complex agent workflows, and matching 3 Opus performance in a Haiku-sized package puts serious pressure on rival small models.

That said, we think stripping image capabilities feels like a clumsy compromise. In an era where multimodal reasoning is table stakes, shipping a text-only model in late 2024 looks like a step backward, even if the speed and text reasoning are exceptional.