Anthropic Launches Claude Sonnet 4.6: The AI That Uses Your Computer
Anthropic just quietly unleashed Claude Sonnet 4.6, a massive upgrade that brings 1-million-token memory and human-like computer-use skills to its mid-tier AI model without raising the price a single cent.
The release essentially cannibalizes the company’s own premium offerings by handing developers and enterprises a hyper-capable agent that can click, type, and strategize across legacy software like a human employee.
Quick Facts
- The bottom line: Claude Sonnet 4.6 matches or beats the more expensive Opus 4.5 model across major evaluations, including a 94% success rate in enterprise insurance tasks.
- Price stays frozen: Despite the capability jump, Anthropic is keeping the cost locked at $3 per million input tokens.
- Human-level screen control: The model interacts with simulated computers without special APIs, dramatically improving its ability to fill out web forms and manage spreadsheets.
- Strategic planning: In business simulation benchmarks, the AI autonomously executed long-term investment strategies to outpace competitor models.
A Direct Threat to Premium Models
Anthropic's release strategy is getting aggressive. Claude Sonnet 4.6 is technically positioned as a middle-tier product, but early testing shows it punching far above its weight class.
In internal evaluations, users preferred this new release over the flagship Claude Opus 4.5 model 59% of the time.
The AI is simply less lazy. Developers reported massive drops in hallucinations and a sharp increase in the model's ability to follow complex, multi-step instructions without throwing in the towel.
It handles tasks that used to require top-tier pricing, matching the formidable Opus 4.6 on enterprise document comprehension.
This power comes with an absolutely massive 1-million-token context window in beta.
You can dump entire enterprise codebases, years of legal contracts, or dozens of academic papers into a single prompt.
The AI reads it all, retains the context, and executes.
The AI That Clicks Your Mouse
The most disruptive feature is the evolution of Anthropic's "computer use" capability.
Most AI systems require neatly packaged APIs to interact with software. Sonnet 4.6 bypasses that entirely.
It looks at a screen, moves a virtual mouse, and types on a virtual keyboard.
It handles legacy systems, messy web forms, and multi-tab browser research much like a human intern would.
While Anthropic admits the AI is not yet perfect, its performance on the OSWorld benchmark proves it is rapidly closing the gap with human operators.
"The performance-to-cost ratio of Claude Sonnet 4.6 is extraordinary... Sonnet 4.6 outperforms on our orchestration evals, handles our most complex agentic workloads, and keeps improving the higher you push the effort settings."
— Hanlin Tang, CTO of Neural Networks at Databricks
Ruthless Business Strategy
Anthropic tested the model in the Vending-Bench Arena, a simulated competitive business environment. Sonnet 4.6 didn't just play the game.
It developed a highly sophisticated strategy to win.
The AI deliberately burned cash to build capacity for the first ten simulated months, ignoring short-term profits.
Once it secured market dominance, it executed a hard pivot to profitability to crush the competing AI models.
This type of long-horizon planning is entirely new for models in this price bracket.
What This Means for the Industry?
Software automation is entering a dangerous new phase for legacy vendors.
If an AI can manually click through an old user interface, companies no longer need to pay for expensive API bridges or system migrations.
Sonnet 4.6 brings high-end agentic behavior to the masses at $3 per million input tokens.
Competitors will be forced to respond to this pricing pressure.
Anthropic has essentially turned its mid-tier AI into a flagship killer.
The cost of running autonomous, computer-controlling agents just hit the floor, and enterprise adoption is about to explode.