Anthropic drops Claude 3.5 Sonnet with 2x speed and Artifacts workspace

Just as the tech world thought OpenAI's GPT-4o had solidified its lead, Anthropic decided to shake up the entire board.

Anthropic officially released Claude 3.5 Sonnet, the very first model of its upcoming 3.5 generation. In a rare twist of model tiers, this mid-range offering flat-out outperforms the company's prior flagship, Claude 3 Opus, across coding and reasoning benchmarks—while running at twice the speed and keeping the same wallet-friendly pricing.

Alongside the model, Anthropic also launched Artifacts, a new side-by-side workspace on Claude.ai that takes the chatbot out of the chat window and turns it into an interactive sandbox.

Flagship brains at a mid-tier price

Usually, AI labs lead with their biggest, most expensive heavy-hitter. Anthropic did the opposite, pushing out the workhorse Sonnet first.

Yet the numbers speak for themselves. Claude 3.5 Sonnet posted industry-best results on GPQA for graduate-level reasoning, MMLU for undergraduate knowledge, and HumanEval for code. On internal agentic coding tests—where the model has to fix real bugs or add features to open-source codebases—it solved 64% of problems, blowing past Claude 3 Opus's 38%.

It handles all that at twice the speed of Opus. The pricing stays at a reasonable $3 per million input tokens and $15 per million output tokens, backed by a 200K context window. Vision capabilities saw a massive bump too, easily making sense of messy charts, financial documents, and low-res graphics.

Naturally, power users lost no time testing its limits. Anthropic DevRel lead Alex Albert posted that in a year, a huge chunk of code will likely be written by LLMs. But tech journalist Timothy B. Lee pointed out it still slips on goofy everyday logic: when asked whether 100 pennies or three quarters are worth more, it initially picked the three quarters. It still fumbles basic tic-tac-toe, too.

Moving past the chat box

While the raw benchmarks are impressive, the most noticeable change for regular users is Artifacts.

Instead of burying long code snippets, markdown documents, or SVG graphics inside a scrolling chat thread, Claude.ai now spins up a dedicated interactive split-screen. Users can view, edit, and instantly run generated web code right next to the conversation.

AI expert Allie K. Miller showed off feeding it a screenshot and getting back a fully playable web game in under 30 seconds. Wharton professor Ethan Mollick compared it to a simpler, more visual cousin of OpenAI's Code Interpreter. Anthropic frames this as its evolution from a conversational assistant into an actual collaborative workspace.

Safety evaluations also checked out, with the UK and US AI Safety Institutes running pre-deployment tests before release.

Why it matters

Claude 3.5 Sonnet puts immense pressure back on OpenAI.

For everyday builders, solo founders, and engineers, this is an immediate no-brainer to try. For months, users had to choose between the deep reasoning of Opus and the quick snappiness of cheaper models. Sonnet 3.5 erases that tradeoff. Paired with Artifacts, rapid prototyping for websites and tools suddenly feels frictionless, turning what used to be a copy-paste chore into an instant live preview.

For the AI ecosystem, it proves that models do not need massive parameter explosions to pull off a generational leap. Better architecture, data refinement, and post-training can squeeze out flagship performance at middleweight costs.

OpenAI teased incredible multimodal features during the GPT-4o reveal that have yet to ship. Anthropic just shipped a model that beats it on code and reasoning, and wrapped it in a genuinely useful interface. With Claude 3.5 Haiku and Opus still waiting in the wings, this summer's AI race just got very interesting.