Claude Sonnet 4.6

Claude Sonnet 4.6

Quick answer: Claude Sonnet 4.6 is Anthropic's cost-efficient frontier model, released 17 February 2026. It pushes near Opus-level intelligence at Sonnet pricing ($3/$15 per million tokens), adds a 1M-token context window (beta), and excels at coding, computer use, and agent planning. Anthropic reports 80.2% on SWE-bench Verified. It is a closed, API-only model (claude-sonnet-4-6).

At a Glance

Where Claude Sonnet 4.6 leads

• Near-flagship intelligence at a third of Opus's price — $3 / $15 per 1M tokens.
• Strong agentic coding (80.2% on SWE-bench Verified) and computer use (94% on a partner insurance benchmark; large OSWorld-Verified gains).
• A 1M-token context window (beta) that it reasons across, not just stores.

Where it lags

• Opus 4.6 remains stronger for the deepest reasoning — large refactors and multi-agent coordination.
• Closed and API-only with no open weights to self-host, and the 1M window is still in beta.

Best for: teams that want most of the frontier's capability at mid-tier cost for broad, high-volume agent work.

What Claude Sonnet 4.6 Is

Claude Sonnet 4.6 is the mid-tier of Anthropic's Claude family, but the gap to the flagship has narrowed sharply. Anthropic frames it as approaching Opus-level intelligence at a price point — $3/$15 per million input/output tokens — that makes it practical for far more tasks. In Anthropic's early testing, developers preferred Sonnet 4.6 to the previous Sonnet 4.5 roughly 70% of the time, and even preferred it to the November 2025 flagship Opus 4.5 about 59% of the time, citing less overengineering, fewer false claims of success, and stronger instruction following.

Like Opus 4.6, it is a hybrid reasoning model supporting both adaptive and extended thinking, with a 1M-token context window in beta — enough to hold an entire codebase, a long contract, or dozens of papers in one request. Anthropic stresses that Sonnet 4.6 reasons effectively across that context, not just stores it, which translates into better long-horizon planning. The release also leans hard into computer use, with major gains on the OSWorld-Verified benchmark and partner reports of human-level performance on multi-step web and spreadsheet tasks.

For Benchgen, Sonnet 4.6 is an API-benchmarked model with no open weights. Its appeal is the cost-to-capability ratio: if a much cheaper model completes domain agent tasks nearly as reliably as a flagship, that is exactly the kind of finding a verified, repeated-trial evaluation surfaces and an aggregator's single-number leaderboard hides. The question Benchgen answers is how reliably that value holds across repeated attempts on a specific task.

Specifications

Pricing

Pricing is unchanged from Sonnet 4.5. Source: Anthropic — Introducing Claude Sonnet 4.6.

Context Window

Claude Sonnet 4.6 has a 1M-token context window (in beta) — roughly 1,500 pages of text in a single request. That is enough to hold an entire codebase, a long contract, or dozens of papers at once, and Anthropic stresses that Sonnet 4.6 reasons effectively across that context rather than merely storing it, which translates into better long-horizon planning.

Public Benchmark Scores

Scores above are reported by Anthropic (some via named partners) and shown for context. They are not Benchgen measurements. Figures depend on harness, effort level, and tool settings — see the source for methodology.

Claude Sonnet 4.6 vs Alternatives

The most telling comparison is internal: Sonnet 4.6 trails Anthropic's own Opus 4.6 by roughly a point on SWE-bench Verified while costing 40% less, and Anthropic concedes Opus stays stronger for the deepest reasoning. Against GPT-5 it offers a far larger context window at a higher token price. For agent builders the decision is usually budget-driven — how much capability per dollar a workflow actually needs. (Rival scores attributed to Anthropic, OpenAI, and Google; not Benchgen measurements.)

How Claude Sonnet 4.6 Performs on Real Agent Tasks

Sonnet 4.6's public story is unusually agent- and work-focused for a mid-tier model: strong computer-use results on OSWorld-Verified, an inventive long-horizon strategy on the Vending-Bench Arena business-simulation eval, and partner benchmarks in insurance, financial services, and enterprise document Q&A. Anthropic explicitly markets it for branched, multi-step workflows like contract routing and CRM coordination — the connective tissue of real agentic systems.

What none of those public numbers establish is reproducible reliability on a specific domain task under a single rubric. A leaderboard number measures general capability; what agent builders actually need is reliable, repeated completion of the task in front of them — including how often the model fails, and why. Because Sonnet 4.6's headline selling point is value — near-flagship behaviour at a third of the cost — that reliability-under-repetition view is exactly what Benchgen is built to surface.

Use Claude Sonnet 4.6 via API

from anthropic import Anthropic

client = Anthropic(api_key="YOUR_API_KEY")

message = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Summarize this contract clause..."}],
)
print(message.content[0].text)

Related Models & Benchmarks

• Compare: Claude Opus 4.6 · GPT-5 · Gemini 3 Pro · Grok 4
• Benchmarks: SWE-bench Verified

Frequently Asked Questions

Specs and scores sourced from Anthropic's Claude Sonnet 4.6 announcement. Third-party benchmark scores attributed inline to Anthropic. Benchgen scores, where shown, are produced on Benchgen infrastructure. Last updated 2026-06-17.