Claude 3.5 Sonnet Explained

What it is

Claude 3.5 Sonnet is Anthropic's mid-tier large language model, first released on June 20, 2024. It made an unusual impression at launch: on many benchmarks it outperformed Claude 3 Opus, Anthropic's own larger flagship model, and it was widely regarded as the leading model for coding and vision tasks through the middle of 2024.

That inversion was the story. The industry expected the biggest model to be the best model, and Anthropic's middle child broke the pattern. Developers quickly adopted it as the default for code generation, and for months it set the bar every competing lab tried to clear.

Anthropic released an upgraded version on October 22, 2024, alongside Claude 3.5 Haiku and a public beta of "computer use," a feature that let the model operate a computer interface by looking at the screen and moving the cursor. By 2026 the model was discontinued, superseded by the 3.7, 4, and 5 generations, but its release marked the moment Anthropic's mid-tier line started beating flagship-class models from every lab.

How it works

Claude 3.5 Sonnet was trained with reinforcement learning from human feedback and Anthropic's Constitutional AI alignment approach, in which the model is trained against a written set of principles rather than purely against human preference votes. The architecture itself was never published, so the internals remain undisclosed.

What Anthropic documented was the capability profile: strong reasoning, strong vision, and tool use. The Artifacts feature launched alongside the model and changed how people used it. Instead of code or documents appearing inline in the chat, Artifacts opened a dedicated workspace window where the model could render substantial content separately: working code, SVG graphics, documents, simple web pages. It turned the chat interface into something closer to a workbench.

It takes text and image input and produces text output. There was no image or audio generation; vision was input-only.

Key details

The context window is 200,000 tokens, with up to 8,192 tokens of output. The knowledge cutoff is April 2024. Anthropic's published benchmarks for the original June version included 59.4 percent on GPQA Diamond, 88.7 percent on MMLU, and 92.0 percent on HumanEval. The October upgrade pushed GPQA Diamond to 65.0 percent and HumanEval to 93.7, while MMLU held at 88.7. On Anthropic's internal agentic coding evaluation, the model solved 64 percent of problems compared with 38 percent for Claude 3 Opus, which explains why developers noticed the difference in practice and not just on paper. These are Anthropic's own figures.

Pricing and access

The API was priced at $3 per million input tokens and $15 per million output tokens, unchanged for the upgraded version. That pricing undercut larger flagship models while delivering better benchmark results, which was part of why adoption was so fast.

The model was free on claude.ai and the Claude iOS app with rate limits, while Claude Pro at around $20 per month and Team plans offered much higher limits. It was also available through Amazon Bedrock and Google Cloud's Vertex AI. Subscription prices come from pricing trackers, so treat them as approximate.

Limitations

The 8,192-token output cap was short for long-form generation work; anything long had to be produced in pieces. Knowledge stopped in April 2024. It hallucinated and failed at instruction following like any model of its generation, and it could not generate images or audio.

As a discontinued model, it is now mainly of historical interest. The interesting question it raised, whether a mid-tier model can outrank a flagship, was answered in the affirmative, and every lab has chased that efficiency curve since. The models that followed it inherited its training recipe and its product features, and the coding-assistant market it defined is still organized around the bar it set.

Why the mid-tier win mattered

The most consequential thing about Claude 3.5 Sonnet was not any single capability but the market structure it revealed. Before June 2024, the industry assumed a clean ladder: the biggest model was the best model, and you paid flagship prices for flagship performance. Sonnet broke the ladder. A mid-tier model beat the flagship above it on the benchmarks people cared about, at a lower price, and developers switched in large numbers.

Every lab internalized the lesson. The efficiency curve became the competition: who could deliver the most capability per dollar of inference. Anthropic's own later releases followed the pattern Sonnet established, with each generation's mid-tier model absorbing the previous flagship's role. The company's pricing stayed aggressive relative to competitors, which kept the pressure on across the industry.

Two product ideas from the Sonnet era outlived the model. Artifacts, the dedicated workspace window, was widely imitated; the notion that a chat interface should have a separate surface for substantial output is now standard. The computer-use beta, which let the model operate a graphical interface, was a direct ancestor of the computer-operating agents every lab shipped in the following years. Sonnet itself is discontinued, but its fingerprints are on everything that came after.