Claude 3 launched March 2024 with Opus, Sonnet, Haiku tiers, vision input, 200K context windows. Specs, benchmark scores, pricing, and the three-tier template.
Stop configuring. Start building.
SaaS builder templates with AI orchestration.
Claude 3 is the launch that pushed Anthropic out of the pure-research corner and into the frontier race. On March 4, 2024, the lineup arrived as three models at three price points, each pitched at a different workload. No AI lab had shipped a coherent family like that before, with clean tradeoffs between speed, cost, and raw capability built right into the release.
| Spec | Opus | Sonnet | Haiku |
|---|---|---|---|
| API ID | claude-3-opus-20240229 | claude-3-sonnet-20240229 | claude-3-haiku-20240307 |
| Context window | 200K tokens | 200K tokens | 200K tokens |
| Input pricing | $15 / 1M tokens | $3 / 1M tokens | $0.25 / 1M tokens |
| Output pricing | $75 / 1M tokens | $15 / 1M tokens | $1.25 / 1M tokens |
| Release date | March 4, 2024 | March 4, 2024 | March 7, 2024 |
A ladder, not one model. Before this release, you picked one Claude and ate the tradeoffs. Anthropic handed developers a hierarchy instead. Haiku sat at the bottom for cheap, latency-sensitive jobs. Sonnet was the everyday driver. Opus lived at the top for the hardest problems. Every lab copied the shape.
Claude learned to see. Vision was new here. Drop in a photo, a chart, a diagram, a handwritten note, or a screenshot, and the model would read it and talk back. Whole categories of work opened up on day one.
Stop configuring. Start building.
SaaS builder templates with AI orchestration.
Stop configuring. Start building.
SaaS builder templates with AI orchestration.
200K token windows across the board. At launch, 200K was close to the biggest context window on the commercial market. Entire codebases, long legal filings, and full research papers fit in one prompt without the chunking ceremony.
Opus swept the benchmarks. The top-tier model took the headline academic evals. MMLU hit 86.8%. GPQA landed at 50.4%. GSM8K came in at 95.0%. Those results cleared GPT-4 and Gemini 1.0 Ultra across most tests. No earlier Claude release had a claim on the best-in-class label.
Haiku moved fast. The small model chewed through a 10,000-token research paper, charts and graphs included, in under 3 seconds. Nothing else in the market ran that quick for high-volume, latency-sensitive jobs at that price.
Claude 2 was one model with one voice. Claude 3 replaced it with a product line. Reasoning, multilingual handling, and accuracy all stepped up in a way you noticed on daily work. The old version refused too much. The new generation engaged with hard requests while keeping its safety lines intact. Vision alone made this a different product.
| Model | Status |
|---|---|
| Claude 3 Opus | Retired (January 2026) |
| Claude 3 Sonnet | Retired (July 2025) |
| Claude 3 Haiku | Still available |
Haiku from this generation is still reachable on the API for lightweight jobs where cost wins over everything else. The 3.5 and 4.x families absorbed what Opus and Sonnet were doing, and those two slots are no longer available.