Claude 3 Family

Anthropic released Claude 3 on March 4, 2024 as a three-tier model family — Opus, Sonnet, Haiku — that briefly took the top of every public LLM benchmark. The release marked the first time a non-OpenAI lab held the public benchmark lead since GPT-4's March 2023 launch. Claude 3.5 Sonnet (June 2024) and Claude 3.5 Sonnet (new, October 2024) extended the line and moved Anthropic's focus from raw benchmarks to agentic capabilities — coding, computer use, long-running tasks.

The Opus / Sonnet / Haiku tier system

Claude 3's three-tier release was a distinct marketing-and-engineering choice from the GPT line:

• Claude 3 Opus — the frontier model, slowest, most capable, highest cost.
• Claude 3 Sonnet — mid-tier, intended as the daily-driver model. Cheaper, faster, slightly less capable.
• Claude 3 Haiku — fast and cheap, intended for high-volume inference.

The naming pattern is now common ("flagship / daily / fast") and signals to developers what to use where. Most production deployments use Sonnet or Haiku for cost-sensitive volume traffic, with Opus reserved for the hardest queries.

Capabilities at launch

Claude 3 Opus matched or beat GPT-4 (the then-incumbent) on most public benchmarks: MMLU, GSM8K, MATH, HumanEval, GPQA, MGSM, HellaSwag. Specific differentiators:

• Vision input. Claude 3 was Anthropic's first model with image input, comparable to GPT-4V at launch.
• 200K-token context. Up from Claude 2's 100K, with strong needle-in-haystack performance.
• Lower refusal rate. Anthropic explicitly tuned Claude 3 to refuse less of the time on benign queries — addressing the over-refusal complaints that had built up around Claude 2.
• Better calibration. Claude 3 was reported to express uncertainty more reliably than its predecessor (whether this carried through to subjective use was contested).

Claude 3.5 Sonnet (June 2024)

The Claude 3.5 Sonnet release (June 20, 2024) was strategically unusual: a Sonnet-tier model that beat the previous Opus-tier flagship on most benchmarks at a fraction of the cost. Anthropic shipped it with the framing "we're building toward Claude 3.5 Opus" — but Opus 3.5 didn't ship that year.

Claude 3.5 Sonnet became the default Anthropic model for most users and the strongest generally-available LLM on coding benchmarks (SWE-bench Verified, HumanEval, LiveCodeBench) for a substantial part of 2024.

"Claude 3.5 Sonnet (new)" — October 2024

Anthropic released a refreshed Claude 3.5 Sonnet on October 22, 2024 (confusingly under the same model name). The October version introduced:

• Computer Use — beta API allowing Claude to control a virtual computer (mouse, keyboard, screenshots). See computer use.
• Improved coding — substantial gains on SWE-bench, real-world IDE tasks.
• Stronger agentic capability — multi-turn tool-use trajectories with better long-horizon coherence.

The "new Sonnet" framing was the start of Anthropic's pivot from benchmark capability to practical agentic capability — not just "knows things" but "does tasks".

Architecture (mostly speculation)

Claude 3 Opus is widely believed to be a Mixture-of-Experts model in the GPT-4-class scale (1-2T total parameters). Sonnet and Haiku are presumably distilled or smaller MoE variants. None of this is confirmed; Anthropic followed the post-GPT-4 norm of withholding architectural details.

Training compute, dataset, and recipe were not disclosed. The model card (Anthropic, 2024) discusses safety evaluations and Responsible Scaling Policy commitments but not technical details.

What Claude 3 established

• Three-tier releases as default. OpenAI's GPT-4o / GPT-4o-mini and Google's Gemini Pro / Flash followed.
• Anthropic as a serious commercial frontier provider. Pre-Claude-3, Anthropic had a small market share. Post-Claude-3 (and especially post-3.5 Sonnet), it became one of the three primary frontier-API providers alongside OpenAI and Google.
• The "agentic capability" framing. Claude 3.5 Sonnet (new) explicitly positioned the model as a task-doer, not just a question-answerer.

Subsequent line

• Claude 3.5 Haiku (Oct 2024) — refreshed fast tier.
• Claude 3.7 Sonnet (Feb 2025) — extended thinking mode (visible chain-of-thought), competitive with o1-class reasoning models.
• Claude 4 family (May 2025) — next-generation flagship, with Sonnet 4, Opus 4 variants. See frontier models.

What to read next

• Anthropic Claude (1, 2) — predecessors.
• GPT-4 — the contemporary OpenAI frontier.
• Frontier Models — the post-Claude-3 frontier landscape.