Sonnet 4.6: The Smartest AI Model for Engineering

Anthropic's latest release is our new daily 'workhorse'

Once again, the rumors were wrong. If you've been on X and Reddit expecting a "Sonnet 5" this week, you might be feeling a pang of disappointment. Don't be.

Anthropic has just released Sonnet 4.6, and while the version number seems like a small incremental jump, the performance feels like a major leap forward.

We know that, despite the price, Opus and Sonnet remain absolute favorites for users. You're not optimizing for the cheapest tokens; you're optimizing for code that actually compiles, agents that don't get stuck in loops, and PRs that pass review on the first try.

You're optimizing for models that fit your vision and — you know the truth — your agentic lifestyle (when AI performs tasks for you).

Sonnet 4.6 is the new daily 'workhorse' for that workflow. It's arguably the smartest and most effective model we've tested yet. Here's what you need to know to get the most out of it.

The Highlight: Cutting-Edge Performance, Sonnet Speed

Anthropic calls Sonnet 4.6 "our most capable Sonnet model yet," but that marketing language undersells it. In our initial tests, this model is showing cutting-edge performance across the board — specifically in coding, agentic workflows, and complex project management.

Sonnet 4.6 achieved an impressive 74.7% on the BrowseComp benchmark and 79.6% — nearly matching Opus 4.6's leading 80.9% — on SWE-bench Verified. In other words, it's here to power any agentic workflow you throw at it.

This isn't just about writing a Python function; it's about iterative development. Sonnet 4.6 excels at navigating complex codebases, managing end-to-end projects with memory, and handling reliable computer use for things like web QA and workflow automation.

New Controls: Effort and Thought

This might also be the closest to an Opus-like improvement we've ever seen from a new Sonnet model.

Just like with Opus 4.6 (released only a few weeks ago), with Sonnet 4.6, we're gaining granular control over how the model applies its intelligence. If you're used to just clicking 'generate,' you'll want to pay attention to these changes to get your money's worth.

Sonnet 4.6 offers strong performance across all thought efforts, even with extended thought turned off, and introduces three distinct 'thought' modes. This is where the magic happens:

• Thought Disabled: The classic experience. Fast and straightforward.

• Extended Thought: The model takes its time to reason about the problem before outputting code.

• Adaptive Thought: A middle ground that adjusts based on query complexity.

For most heavy coding tasks, we're seeing the best results with Extended Thought on 'Medium' effort. The reasoning capability here is surprisingly good at catching edge cases before writing a single line of code.

However, if you're migrating existing Sonnet 4.5 workflows or prompts and want 'just works' reliability, Thought Disabled is your safest bet. It mimics the 4.5 behavior, but with the 4.6 intelligence upgrade.

1 Million Context Tokens (Beta)

This is the big one for enterprise codebases. Sonnet 4.6 supports a 1 million token context window in beta.

If you've ever hit the context limit while trying to feed a massive documentation file into your prompt, this is the solution.

What does this mean in practice? Essentially, you can ask the model to analyze enormous documents without needing to break them into smaller pieces. It's like having a mega-memory capable of remembering everything at once.

Feature Release: Now GA (Generally Available)

Alongside the model, several critical API features have moved to General Availability (GA).

• Code Execution and Web Search: The agent can execute code and browse the web more reliably.

• Tool Use and Programmatic Tool Calling: This makes agentic capabilities significantly faster.

• Memory: Improved retention of project details across chat turns, sessions, and modes.

The Verdict

Sonnet 4.6 is a huge improvement that almost matches Opus's performance, but at Sonnet's more accessible price point. If you want an all-rounder model — coding, analyzing huge documents, task automation — without paying top dollar for Opus, now is the time to try Sonnet 4.6.

Takeaway: Sonnet 4.6 proves we don't need to wait for big numerical jumps to get significant advancements. With refined control over the reasoning process and expanded context capacity, this version represents the most balanced model for serious development work — becoming not just an upgrade, but the new gold standard for engineering teams.

Glossary of Technical Terms

• Benchmark: A standardized test used to measure the performance of an AI model on specific tasks. It's like a school exam for computers.

• Token: The smallest unit of text that an AI model processes. Think of it as a "syllable" or fragmented word. The more tokens, the more information can fit.

• Context Window: The amount of information the model can "remember" during a conversation. The larger it is, the more context it can process at once.

• Agentic: When an AI not only answers questions but performs tasks autonomously, like a personal assistant doing work for you.

• Code Execution: The ability to actually run code, not just talk about code. The model can create and execute programs.

• Computer Use: When AI can use the computer as if it were a person (click, type, navigate websites).

• PR (Pull Request): A request to include your code in a project. It's like submitting your work for review before it gets approved.

• Infinite Loop: When a program gets stuck repeating the same thing endlessly and never terminates.

• Workhorse: A super reliable tool you use every day for heavy-duty work.

• Output: The text that the model generates as a response.

• Iterative: Doing something in stages, gradually improving each time.

• SWE-bench Verified: A specific benchmark that tests how AI solves real programming problems.

• BrowseComp: A benchmark that tests the ability to search for and analyze information on the web.

• Memory (in AI context): The model's ability to remember things you've said earlier in the conversation.

• Chat Turn: Each time you send a message and the AI responds, it's a "turn".