Claude 3 and the Opus, Sonnet, Haiku Tiering

By Satwik · May 2, 2026

Anthropic released the Claude 3 family in March 2024 as three sized models: Opus (most capable), Sonnet (balanced), and Haiku (fast and cheap). The tiering made an explicit product argument that most tasks do not need the largest model, and that latency and cost are first-class design constraints. All three added vision, handling images, charts, and documents alongside text.

Why it mattered: Opus was, at launch, competitive with or ahead of the strongest models on many reasoning and knowledge benchmarks, and the family demonstrated that a single training recipe could be scaled up and down cleanly. The tiered approach became an industry pattern, with other labs shipping small, medium, and large variants under shared branding.

The safety framing is what we note. Anthropic shipped Claude 3 under its Responsible Scaling Policy, assigning the models an AI Safety Level and publishing evaluation results for dangerous-capability domains. Opus reportedly showed a marked drop in unnecessary refusals relative to prior versions, addressing the over-refusal failure mode that erodes user trust and pushes people toward less safe tools.

For practitioners, the tiering has a security dimension too: routing sensitive or high-stakes tasks to the more capable, better-evaluated tier while using smaller models for bulk work is a defensible design, but it means the cheaper models see the long tail of untrusted input. Haiku's speed makes it attractive as a first-line filter or classifier in agent pipelines, a role where prompt-injection resistance matters as much as raw capability. Claude 3 set a template the whole field followed.