Claude Sonnet 4.6

Claude Sonnet 4.6 is the mid-tier model that started beating last generation's flagship in head-to-head coding work. In internal Claude Code testing, developers picked it over Sonnet 4.5 about 70% of the time. Against Opus 4.5, the previous frontier model from November 2025, testers still preferred 4.6 on 59% of sessions. That is a Sonnet model outscoring an Opus model on developer preference, priced at $3 input and $15 output per million tokens.

The other big jump shows up in computer use. And across enterprise document work, Sonnet 4.6 lands on par with its Opus sibling. Same price as before.

Key Specs

What Changed: The Coding Improvements

Long sessions are where the difference is loudest. Sonnet 4.6 reads context before it touches code, folds shared logic into one place instead of duplicating it, and backs off the over-eager refactors that made older models frustrating.

Better context comprehension. Before making a change, the model actually reads the code around it. House conventions get picked up. Redundant patterns are avoided. Edits land in a way that fits the codebase they are going into.

Reduced overengineering. Ask for a small fix and you get a small fix. Hallucinated changes and false "done" claims drop sharply compared to earlier models. You will see far fewer "I've refactored the entire module for you" answers when a one-line patch was all you wanted.

Stronger on complex tasks. Bug detection got good enough that Sonnet can stand in for Opus as a parallel reviewer on a team. Cross-file refactors, dependency chains, and hard multi-step fixes land more reliably across big codebases.

Improved design sensibility. Frontend generation looks more polished out of the gate. Layouts read cleaner, animations are less clumsy, and fewer rounds are needed to reach something shippable. Early testers said the model had "perfect design taste" for building frontend pages and data reports.

Long-horizon planning. On Vending-Bench Arena, a strategic simulation, Sonnet 4.6 beat Sonnet 4.5 by investing in capacity early and then pivoting to profit at the end. Branched, multi-step reasoning like that shows up directly in harder daily work.

Benchmark Results

Numbers that used to need an Opus-class model now come out of Sonnet:

For a working developer, the preference numbers matter most. A $3/$15 model picked over a $5/$25 model on most coding sessions. That changes how you spend your daily Claude budget.

Outside validation backs the internal results. Cursor's co-founder called it "a notable improvement over Sonnet 4.5 across the board, including long-horizon tasks and more difficult problems." GitHub reported "strong resolution rates and the kind of consistency developers need" on complex code fixes across large codebases. Cognition found it "meaningfully closed the gap with Opus on bug detection," letting them run more parallel reviewers without increasing cost.

Computer Use

This is where the biggest jump shows up. OSWorld scores across 16 months of Sonnet releases climbed steadily, and 4.6 is the largest single step on that chart.

Spreadsheets, multi-step web forms, and enterprise document intake all run with higher accuracy. On the submission intake and first notice of loss workflows at Pace, an insurance tech company, the model hit 94% accuracy. That is the best number any model has posted for Pace so far.

Adversarial content also has a harder time derailing a session. Prompt injection resistance on Sonnet 4.6 is roughly even with Opus 4.6, so computer use runs are harder to knock off course.

Safety Profile

The intelligence bump did not cost safety. Anthropic's evaluations describe Sonnet 4.6 as "broadly warm, honest, prosocial, and at times funny" in character. They also report "very strong safety behaviors and no signs of major concerns around high-stakes forms of misalignment."

Resistance to prompt injection is up versus Sonnet 4.5 and now sits with the current Opus tier. For teams shipping computer use or pointing the model at untrusted documents, that is a real improvement in how well it shrugs off manipulation.

New Platform Capabilities

A handful of platform pieces shipped alongside the model.

Claude in Excel with MCP. The Excel add-in now speaks MCP. You can read directly from financial sources like S&P Global, LSEG, Daloopa, PitchBook, Moody's, and FactSet. Pro, Max, Team, and Enterprise plans get access.

Free tier upgrades. File creation, connectors, skills, and context compaction are now on the Free plan. Features that used to sit behind Pro are becoming baseline.

Default on Free and Pro. On claude.ai and Claude Cowork, Sonnet 4.6 takes over from Sonnet 4.5 as the default for Free and Pro users.

Expanded tool access. Web search with dynamic filtering, code execution, memory, programmatic tool calling, and tool search are all GA with this release.

Pricing

No price change. Pricing is flat across the full 1M context window. There is no long-context premium. A 900K-token request bills the same per token as a 9K request:

Already on Sonnet 4.5 with your usage and cost in hand? Moving up is pure upside at the same rate.

How to Use Sonnet 4.6 in Claude Code

Switch your default model with one command:

claude config set model claude-sonnet-4-6

For per-session overrides without changing your default:

claude --model claude-sonnet-4-6

The model is live on every platform. That covers claude.ai (where it is the Free and Pro default), Claude Cowork, the Messages API, Claude Code, Amazon Bedrock, and Google Vertex AI. The API model identifier is claude-sonnet-4-6.

Sonnet 4.6 vs Sonnet 4.5: What Changed

The core gains are coding quality, computer use, and instruction following. Everything Sonnet 4.5 was already good at, like speed, cost, and agent behavior, carries over with a real intelligence step on top.

Model selection is simple. Make Sonnet 4.6 the default for fast iteration and for the 90%+ of daily coding work where speed and cost decide the tradeoff. Reach for Opus 4.6 on the deepest reasoning, big codebase refactors, multi-agent coordination, and work where precision cannot slip. Opus 4.6 still holds the top slot on Terminal-Bench 2.0 and Humanity's Last Exam.

Related Pages

• All Claude Models for the complete model timeline
• Opus 4.6 for the top-tier option in the same generation
• Sonnet 4.5 for the previous Sonnet release
• Model selection guide for strategic model switching
• Usage optimization for managing costs across models