Claude 3 Sonnet
Last reviewed
Sources
14 citations
Review status
Source-backed
Revision
v2 · 1,833 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
14 citations
Review status
Source-backed
Revision
v2 · 1,833 words
Add missing citations, update stale details, or suggest a clearer explanation.
Claude 3 Sonnet is the balanced mid-tier large language model in Anthropic's Claude 3 family, released on March 4, 2024 alongside the more powerful Claude 3 Opus and the smaller, faster Claude 3 Haiku. It offers a 200,000-token context window, image (vision) input, an August 2023 knowledge cutoff, and pricing of $3 per million input tokens and $15 per million output tokens, and it powered the free tier of Claude.ai at launch. [1][2][3] Positioned as the workhorse of the three-model lineup, Sonnet sits between Opus at the top and Haiku at the bottom, trading some of Opus's peak reasoning for roughly twice the speed at one-fifth the price. [2]
Claude 3 Sonnet is a multimodal model that accepts text and image input and returns text, built to balance intelligence, speed, and cost for production and enterprise use. Anthropic described it as the practical default of the Claude 3 generation: "Claude 3 Sonnet strikes the ideal balance between intelligence and speed, particularly for enterprise workloads." [1] It was the first Claude model many users actually touched, because it ran the free experience on Claude.ai while the more expensive Opus stayed behind a Claude Pro subscription. [1][2]
Anthropic announced the Claude 3 generation on March 4, 2024, presenting three models named, in ascending order of capability, Haiku, Sonnet, and Opus. Opus and Sonnet became available the same day through the company's API and the Claude consumer interface, while Haiku was described at launch as arriving later; it reached general availability on March 13, 2024. [1][3] The API model identifier for Sonnet carries the snapshot date February 29, 2024, written as claude-3-sonnet-20240229, although the public release came on March 4. [4]
The Claude 3 models succeeded Anthropic's earlier Claude 2 and Claude 2.1 systems. Anthropic said "for the vast majority of workloads, Sonnet is 2x faster than Claude 2 and Claude 2.1 with higher levels of intelligence," a combination the company framed as the practical default for production use. [1][2] The whole family was trained under the company's constitutional AI approach, in which model behavior is shaped against a written set of principles rather than relying solely on human preference labels. Claude 3 was also Anthropic's first multimodal release, adding image understanding to a product line that had been text-only. [3]
The naming convention used poetic forms to signal size: haiku for the smallest and quickest, sonnet for the middle tier, and opus for the most capable. Sonnet denoted the workhorse tier intended for high-throughput tasks where a balance of quality, latency, and cost mattered more than maximum reasoning power. [2]
Claude 3 Sonnet is a multimodal model that accepts text and image input and returns text. Anthropic described its vision features as able to process photographs, charts, graphs, and technical diagrams, including material drawn from PDFs and slide decks, and the model could take up to 20 images in a single request for comparative analysis. The company noted limits: the model would not identify specific people in images, performed poorly on low-resolution images below roughly 200 pixels, and was weaker at spatial reasoning and object counting. [1][3]
The model has a context window of 200,000 tokens, which Anthropic put at about 150,000 words, and a maximum output of 4,096 tokens. Its training knowledge cutoff is August 2023. [3][5] Anthropic also highlighted improvements over the prior generation in following multi-step instructions, producing structured output such as JSON, and reducing unnecessary refusals, along with what it described as fewer hallucinations and higher accuracy on factual questions. [1]
Anthropic said Sonnet "excels at tasks demanding rapid responses, like knowledge retrieval or sales automation," and aimed it at data processing, retrieval-augmented search, sales tasks such as forecasting and recommendation, and code generation, framing it as the model for large-scale deployments that needed quick responses at a reasonable price. [1] Through Anthropic's business products, including Claude for Work, the model was offered to organizations building on these workloads.
In the Claude 3 model card, Anthropic reported Sonnet's results on standard public evaluations. Sonnet landed between Haiku and Opus on most tests, trailing the flagship while clearly improving on the smallest model. For example, Sonnet scored 79.0% on MMLU (5-shot) against Opus's 86.8% and Haiku's 75.2%, and 73.0% on the HumanEval coding benchmark against Opus's 84.9%. The figures below are Anthropic's reported numbers, shown next to Opus for context. [6]
| Benchmark | Setup | Claude 3 Sonnet | Claude 3 Opus |
|---|---|---|---|
| MMLU (general knowledge) | 5-shot | 79.0% | 86.8% |
| GPQA (graduate-level reasoning, Diamond) | 0-shot chain of thought | 40.4% | 50.4% |
| GSM8K (grade-school math) | 0-shot chain of thought | 92.3% | 95.0% |
| MATH (competition math) | 0-shot | 43.1% | 60.1% |
| HumanEval (code) | 0-shot | 73.0% | 84.9% |
| MMMU (multimodal) | validation | 53.1% | 59.4% |
Anthropic presented these scores to show the trade-off across the family, where each step up the tier raised accuracy at the cost of speed and price. [6] At launch Anthropic emphasized that the family as a whole, led by Opus, exceeded OpenAI's GPT-4 and Google's Gemini 1.0 Ultra on a range of academic benchmarks, though the comparison did not include Google's then-recent Gemini 1.5 Pro. [3][2] Independent benchmarking services later ranked the original Sonnet toward the lower end of contemporary models as stronger systems appeared, consistent with its early-2024 release. [7]
The Claude 3 lineup was designed as a three-tier ladder where each step traded speed and cost for capability. Opus is the most capable and the most expensive, aimed at the hardest reasoning and analysis tasks; Haiku is the smallest, fastest, and cheapest, built for near-instant responses; and Sonnet sits in the middle as the balanced workhorse. [1][2] At launch, Sonnet cost $3 per million input tokens and $15 per million output tokens, compared with Opus at $15 input and $75 output, and Haiku at $0.25 input and $1.25 output. [1] The table below summarizes the launch positioning across the three models.
| Model | Tier | Input price (per M tokens) | Output price (per M tokens) | Context window | MMLU (5-shot) |
|---|---|---|---|---|---|
| Claude 3 Haiku | Fastest / cheapest | $0.25 | $1.25 | 200,000 | 75.2% |
| Claude 3 Sonnet | Balanced | $3 | $15 | 200,000 | 79.0% |
| Claude 3 Opus | Most capable | $15 | $75 | 200,000 | 86.8% |
Sonnet was available on launch day through the first-party API, with general availability stated for 159 countries, and on Claude.ai. It also appeared on Amazon Bedrock the same day and in private preview on Google Cloud's Vertex AI Model Garden, with wider access following. [1][8] On the consumer side, Sonnet powered the free tier of Claude.ai, while Opus was reserved for Claude Pro subscribers, an arrangement Anthropic kept for later mid-tier releases. [1][2] Pricing held to the generation's 1:5 ratio between input and output tokens: $3 per million input tokens and $15 per million output tokens. [1]
| Attribute | Value |
|---|---|
| Input price | $3 per million tokens |
| Output price | $15 per million tokens |
| Context window | 200,000 tokens |
| Maximum output | 4,096 tokens |
| Input modalities | Text, image |
| Output modality | Text |
| Knowledge cutoff | August 2023 |
| API model ID | claude-3-sonnet-20240229 |
| Release date | March 4, 2024 |
Coverage of the launch focused mainly on Opus and Anthropic's claim that it had overtaken GPT-4 on standard benchmarks, a milestone the company had not previously reached. [9] Within days of release, the larger model rose to the top of the public Chatbot Arena leaderboard, the first time a non-OpenAI system had led that ranking. [10] Sonnet drew attention as the model most users would actually touch, since it ran the free Claude.ai experience and undercut comparable offerings on price while keeping strong throughput. Reporting noted that the family's addition of vision brought Anthropic in line with multimodal rivals, and that the 200,000-token context window was among the larger ones generally available at the time. [3][9] Some commentators cautioned that benchmark wins did not always translate into better day-to-day results and advised testing models directly rather than relying on headline scores. [11]
Anthropic announced Claude 3.5 Sonnet on June 20, 2024, with availability the following day. The company said the new model ran at twice the speed of Claude 3 Opus while keeping the speed and cost of Claude 3 Sonnet, and that it outperformed Opus across a wide range of evaluations. On an internal coding test it solved 64 percent of problems against Opus's 38 percent, and it surpassed Opus on vision benchmarks. Pricing stayed at $3 per million input tokens and $15 per million output tokens, matching the original Sonnet. [12] Claude 3.5 Sonnet also introduced the Artifacts feature, which renders generated content such as code or documents in a side panel, and an upgraded version released on October 22, 2024 added a computer-use capability. [13] Anthropic later extended the Sonnet line with Claude 3.5 Haiku and subsequent generations.
Anthropic placed the original Claude 3 Sonnet on its deprecation schedule, notifying developers on January 21, 2025, and retired the claude-3-sonnet-20240229 endpoint on July 21, 2025, alongside the older Claude 2.0 and Claude 2.1 models. After retirement, calls to the identifier returned errors, and Anthropic directed users to its newer Sonnet models as the recommended replacement. [14]