Anthropic Claude 3
Anthropic introduced Claude 3 as a three-tier family — Haiku (fast/cheap), Sonnet (balanced), and Opus (most capable) — and claimed Opus outperformed peers, including GPT-4, on most common benchmarks such as MMLU (undergraduate knowledge), GPQA (graduate reasoning), and GSM8K (math). It touted near-instant responses at the cheaper tiers and, for the first time across the family, vision capabilities able to read photos, charts, and technical diagrams. Anthropic also claimed reduced unnecessary refusals versus prior Claude models.
Opus and Sonnet shipped at launch on claude.ai and the API (with Sonnet powering the free tier and Opus behind the paid Claude Pro plan); Haiku followed shortly after. Opus was priced at the premium end at $15/$75 per million input/output tokens, with Sonnet and Haiku progressively cheaper.
Claude 3 was the moment Anthropic credibly arrived at the frontier: Opus was the first model from a rival to plausibly claim parity with or an edge over GPT-4 on headline benchmarks, ending OpenAI's roughly year-long uncontested lead. Real-world reception largely validated the claims for writing quality, long-context work, and nuanced reasoning, though benchmark wins were narrow and contested, and GPT-4-class competition remained close. The tiered Haiku/Sonnet/Opus structure proved strategically durable — it became Anthropic's lasting product template and pushed the industry toward a price-for-capability menu rather than one flagship. Opus's premium price limited its everyday use, but Sonnet and Haiku drove broad adoption.
This was the launch that turned the AI race from a one-leader story into a genuine contest, giving executives a real choice of frontier vendor — and a clear capability-vs-cost menu to buy against.