Anthropic launches Claude 3.5 Sonnet, more capable than Opus and cheaper
Anthropic introduces Claude 3.5 Sonnet, a model that outperforms Claude 3 Opus on several tests despite costing far less. It also debuts Artifacts, a workspace for working with generated code and documents.
On June 20, 2024, Anthropic introduced Claude 3.5 Sonnet and Artifacts, a side panel for working with model outputs. Anthropic’s technical announcement publishes pricing, speed and evaluations against other models; these are vendor measurements and provide a starting point, not proof that the system wins on every task.
The launch also introduces Artifacts, a new Claude.ai feature that separates useful output from the conversation. Instead of leaving a document, webpage or block of code buried among messages, Claude displays it in its own workspace, where users can view, modify and reuse it.
More performance at Sonnet’s price
Claude 3.5 Sonnet retains the price of Claude 3 Sonnet: $3 per million input tokens and $15 per million output tokens. Tokens are the units of text these systems use to read instructions and produce a response.
Claude 3 Opus, by contrast, costs $15 per million input tokens and $75 per million output tokens. In other words, 3.5 Sonnet is a far more affordable alternative to Anthropic’s high-end model, without giving up its ability to compete with it on performance.
The company also says the new model runs at roughly twice Opus’s speed. For someone who uses a chatbot occasionally, that difference may seem secondary. For a company processing thousands of documents, serving customers or integrating the model into an application, price and speed determine whether a feature is viable beyond a demo.
In its internal evaluation of tool-based coding tasks, Anthropic says Claude 3.5 Sonnet solved 64% of the problems, compared with 38% for Claude 3 Opus. This type of test measures a model’s ability to tackle real-world software issues, where completing a few lines of code is not enough: it must locate files, interpret an error and propose coherent changes.
The comparison should be read with caution. Benchmarks make it possible to measure progress and compare models, but they do not replace testing in each user’s specific workflow. An assistant can perform very well on a battery of exercises and still make mistakes when interpreting incomplete data, applying an internal policy or generating code that looks correct but is not.
Artifacts takes work out of the chat window
The most visible new feature for Claude.ai users is Artifacts. When a conversation produces content that is sufficiently extensive and self-contained — such as code, a report, an SVG diagram or a small website — Claude can open it in a panel beside the chat.
The change may look like a UI update, but it affects how people use an assistant. Traditional chats mix instructions, responses, corrections and successive versions into a single stream. Finding the final result can be awkward, especially when working on a simple application or a document with multiple revisions.
With Artifacts, users can review the result while continuing to chat with the model. If Claude creates an HTML page, the workspace can display both the code and a preview of the result; if it drafts a document, the document remains separate from the explanations that led to its creation. The feature is available in preview to Claude.ai users.
Anthropic did not invent the idea of a workspace beside the chat. Microsoft and Google have already explored interfaces that combine text generation, editing and code execution. But bringing this format to Claude points to a competition that is no longer solely about answering better than the rival. The major labs want users to do the entire job inside their products.
A different strategy from reserving the best for the high end
Anthropic’s move has a clear commercial rationale. Instead of placing its most capable model at the highest price tier, the company is upgrading its mid-range option. That increases pressure on providers that draw a sharper line between their premium and cheaper models.
For developers and businesses, the promise is attractive: get reasoning performance close to or better than that of an elite model without multiplying the infrastructure bill. For Anthropic, the challenge will be proving that advantage holds up in everyday use, not just in its published evaluations.
Claude 3.5 Sonnet is already available through Claude.ai and Anthropic’s API. The company has said it is preparing new versions of Claude 3.5 Haiku and Claude 3.5 Opus, along with more features for Artifacts. For now, the launch shifts attention away from the largest model and toward a more practical question: how much useful performance can an assistant deliver before its cost stops making sense.
A vendor table is a working hypothesis
A useful comparison begins by copying the task, dataset, version, prompt, permitted tools and scoring rule. “Beats Opus” summarises a table; it does not say whether the margin survives private documents, Spanish, proprietary code or ambiguous requests. When a protocol detail is missing, record it as unknown instead of filling the gap with the model’s reputation.
Internal coding tests require particular care. Solving an issue may mean locating the file, changing it and passing checks prepared by the evaluator. Before buying the claim, repeat a sample on the team’s repositories with hidden tests and regression review. A solution that passes the exam while breaking another route is not useful.
Price per token is not task cost
Real cost includes input, output, retries, tool calls, cache, review time and the share of accepted jobs. A cheap model that needs three attempts may cost more than an expensive one that finishes once. Sustained latency matters too: stated speed and the duration of a session with files, network and queues are different measurements.
Keep twenty or thirty representative assignments and a reference result. Run them with comparable temperature and limits, then measure accuracy, human edits, seconds and total cost. That record reveals whether an update improves the average by creating rare but serious failures.
Artifacts creates an object to govern
Separating output from chat makes review and reuse easier, but it does not turn code or documents into validated material. Treat an executable artifact like any outside contribution: minimum permissions, inspected dependencies, tests, change history and a named approver. A preview displays an interface, not its security.
For documents, verify citations, figures and version differences. For code, keep secrets out of the conversation and run it in an isolated environment before production. The transferable skill is to judge a model with a local task portfolio and accepted-job cost rather than an aggregate table. A product improvement then becomes a reproducible decision instead of borrowed enthusiasm.
Review adoption after the pilot. If the team uses Artifacts to build small applications, count changes accepted without correction, defects found later and maintenance time. The value of an interface lies not in producing more drafts but in increasing the results someone can retain with clear responsibility.
The final decision fits in a short table: task, model and date, minimum quality, maximum cost, latency, risk and manual fallback. The team signs it before expanding use and measures it again after every version change. A strong first impression is not mistaken for stable capability, and an old comparison does not become permanent policy.
This article was produced with artificial intelligence under human editorial oversight.
