Claude Sonnet 4.6: Opus-Level Power at Sonnet Pricing

Here's the thing about the Sonnet tier: it's where most of the actual work happens. Opus gets the headlines, but Sonnet is what developers hit millions of times a day in production. It's the model powering the majority of Claude API calls, the default in Claude Code, and the go-to for teams that need speed and capability without burning through budgets.

Sonnet 4.6 changes the calculus for when you need to reach for Opus. Developers with early access have been preferring Sonnet 4.6 over its predecessor by a wide margin. The surprise: many preferred it to Claude Opus 4.5, the smartest model from November 2025. Performance that previously required an Opus-class model is now available at Sonnet pricing.

Sonnet 4.6 doubles the previous Sonnet context limit to 1 million tokens in beta. That's enough to hold entire codebases, lengthy contracts, or dozens of research papers in a single request. When Opus 4.6 launched with 1M context two weeks ago, many developers wondered when Sonnet would follow. The answer: now.

For practical use, this means you can analyze a full Next.js application's source code, review an entire quarter of SEC filings, or process a complete technical documentation site - all in one conversation without chunking or RAG workarounds.

Sonnet 4.6 brings much-improved coding skills to the Sonnet tier. The model excels at complex code fixes, especially when searching across large codebases is essential. It handles multi-file refactors, understands project-level architecture, and catches its own mistakes more reliably than Sonnet 4.5.

The numbers: 79.6% on SWE-bench Verified, up from Sonnet 4.5's 77.2% and within 1.2 points of Opus 4.6's 80.8%. That's the narrowest Sonnet-to-Opus gap in any Claude generation. With a prompt modification, Anthropic saw it hit 80.2%. The SWE-bench score was averaged over 10 trials with thinking turned off - meaning this is the model's baseline coding ability, not its ceiling.

OSWorld is the standard benchmark for AI computer use, and Sonnet models have been making steady gains across sixteen months. Early Sonnet 4.6 users are seeing human-level capability in tasks like navigating complex spreadsheets, filling out multi-step web forms, and coordinating actions across multiple browser tabs.

Sonnet 4.6 hit 72.5% on OSWorld-Verified - within 0.2 points of Opus 4.6's 72.7%, and nearly double GPT-5.2's 38.2%. In real-world testing, insurance company Pace reported 94% accuracy on their computer use tasks. For anyone building browser automation or desktop agent workflows, this is the most capable Sonnet model for the job.

On OfficeQA - which measures how well a model can read enterprise documents like charts, PDFs, and tables - Sonnet 4.6 matches Opus 4.6 performance. That's a meaningful upgrade for document comprehension workloads. If you're building document processing pipelines, you can now get Opus-quality extraction at one-third the cost.

Early access customers independently described visual outputs from Sonnet 4.6 as notably more polished, with better layouts, animations, and design sensibility. The model also needed fewer rounds of iteration to reach production-quality results. For frontend developers and designers using Claude to generate UI code, this means less back-and-forth to get something that actually looks good.

Read the original source on anthropic.com

The numbers paint a clear picture. Sonnet 4.6 posts improvements across the board, with standout results on reasoning and coding benchmarks.

The story across these benchmarks is consistent: Sonnet 4.6 is within striking distance of Opus on coding and computer use, while posting major gains in reasoning and math. The 1.2-point SWE-bench gap and 0.2-point OSWorld gap are essentially rounding errors. Where Opus still leads meaningfully is scientific reasoning (GPQA Diamond: 91.3% vs 74.1%) and the hardest novel reasoning tasks. For a model at $3 per million input tokens, these numbers rewrite what's possible at the Sonnet price point.

As of today, Opus 4.6 sits at the top of the Arena leaderboard (formerly LMArena/LMSYS Chatbot Arena) with a 1,506 Elo score in the thinking variant. Sonnet 4.5 previously held positions #11 and #12 on the text leaderboard. Sonnet 4.6 is brand new - expect Arena rankings to update as users cast votes over the coming days.

Given the benchmark improvements over Sonnet 4.5 and the early-access preference data showing it competing with Opus 4.5, there's reason to expect Sonnet 4.6 will climb significantly on the Arena leaderboard.

Five days before Sonnet 4.6 launched, Google released a major upgrade to Gemini 3 Deep Think on February 12, 2026. It's the clearest competitor to Claude's reasoning capabilities - but the two models are solving different problems.

On pure reasoning benchmarks, Deep Think is ahead. Its 84.6% on ARC-AGI-2 (verified by the ARC Prize Foundation) towers over Sonnet 4.6's 60.4% and even Opus 4.6's ~65%. On Humanity's Last Exam - the benchmark designed to be unsolvable by current AI - Deep Think scored 48.4%, blowing past GPT-5 Pro's ~31.6%. Its Codeforces Elo of 3,455 puts it at elite competitive programming levels. Google also reports gold medal-level performance on the International Physics and Chemistry Olympiads.

But here's the tradeoff: Deep Think is a specialized "slow reasoner." It's designed for asynchronous scientific discovery and mathematical proof verification - problems where you're willing to wait minutes for an answer. Sonnet 4.6 is a general-purpose workhorse that responds in seconds. They occupy different positions in the emerging split between fast conversational AI and deep reasoning engines.

Where Sonnet 4.6 dominates the comparison is practical software engineering and computer use. Its 72.5% on OSWorld-Verified is nearly double GPT-5.2's 38.2%, and Deep Think doesn't even compete on that benchmark. SWE-bench Verified at 79.6% reflects real-world coding ability that Deep Think's Codeforces ranking (competitive algorithm puzzles) doesn't capture. For the 95% of developers who need a model to refactor code, debug production issues, and navigate browser-based workflows, Sonnet 4.6 is the better tool.

Availability also differs. Deep Think is restricted to Google AI Ultra subscribers and a limited API early access program. Sonnet 4.6 is available to everyone on claude.ai Free and Pro tiers, the full API, and Claude Code. For teams evaluating which model to build on, accessibility and pricing matter as much as peak benchmark scores.

Two beta features shipped alongside the model upgrade. Adaptive thinking lets Sonnet 4.6 dynamically adjust how much reasoning effort it applies to a problem - spending more compute on hard questions and less on simple ones. This is similar in concept to what powers Deep Think's extended reasoning, but calibrated for real-time use rather than marathon problem-solving sessions.

Context compaction automatically summarizes older conversation context, effectively enabling unlimited conversation length within the 1M token window. Instead of hitting a wall when context fills up, the model compresses earlier turns while preserving key information. For long coding sessions or multi-hour research workflows, this means fewer restarts and less context management overhead.

Anthropic ran extensive safety evaluations. The results: Sonnet 4.6 is as safe as, or safer than, other recent Claude models. Their safety researchers described it as having "a broadly warm, honest, prosocial, and at times funny character, very strong safety behaviors, and no signs of major concerns around high-stakes forms of misalignment."

The real improvement is in prompt injection resistance. Computer use is powerful but risky - malicious actors can hide instructions on websites to hijack the model. Sonnet 4.6 is a major improvement over Sonnet 4.5 in resisting these attacks, performing similarly to Opus 4.6. If you're building computer use agents, this directly reduces your security surface.

Sonnet 4.6 is available now across all channels:

Same pricing, substantially more capable. Anthropic is following the pattern they set with Opus 4.6 - delivering more intelligence at the same cost per token. For production workloads, this is a straightforward model swap with the ID claude-sonnet-4-6.

Anthropic has maintained a rapid release pace. Here's where Sonnet 4.6 sits in the current model family:

An updated Haiku model is expected to follow in the coming weeks, which would complete the 4.6 generation across all three tiers.

If you're building on the Claude API, the practical impact is clear. The 1M context window means simpler architectures - fewer RAG pipelines needed when you can just feed the full document set. The coding improvements mean fewer iterations per task. The prompt injection resistance means safer computer use agents.

For teams that were considering upgrading to Opus for specific workloads, Sonnet 4.6 narrows the gap enough that many use cases no longer justify the 5x price premium. Document comprehension matching Opus on OfficeQA is a concrete example - there's no reason to pay Opus pricing for PDF extraction if Sonnet delivers the same accuracy.

The model ID for the API is claude-sonnet-4-6. If you're on claude.ai Free or Pro, you're already using it.

Claude Sonnet 4.6 delivers near-Opus intelligence at Sonnet pricing: 79.6% SWE-bench, 72.5% OSWorld, 60.4% ARC-AGI-2, and 89% on math. The 1M token context window, adaptive thinking, context compaction, and Opus-matching document comprehension make this the most capable mid-tier model available.

Gemini 3 Deep Think wins on pure reasoning benchmarks (84.6% ARC-AGI-2, 48.4% Humanity's Last Exam) but operates as a specialized slow reasoner with limited availability. For the practical work most developers actually do - coding, computer use, document processing, and general-purpose AI tasks - Sonnet 4.6 at $3/$15 per million tokens is the new default. Opus 4.6 still leads on the hardest scientific reasoning, but the gap has never been thinner.