Can Sonnet 4.6 Match Opus Performance at One‑Fifth the Cost?

Introduction

On Feb 17 2026 Anthropic launched Claude Sonnet 4.6 only 12 days after Opus 4.6. Sonnet is positioned as a “mid‑scale flagship” model and is priced at $3 per million input tokens and $15 per million output tokens, roughly one‑fifth of Opus 4.6.

Positioning and Release Background

Anthropic’s model family is divided into Opus (flagship), Sonnet (core), and Haiku (lightweight). Sonnet 4.6 matches or exceeds Opus 4.5 on several tasks while costing far less, a claim supported by Replit CEO Michele Catasta’s comment on its cost‑performance.

Overall Capability Improvements

Benchmark comparisons show the following gains over Sonnet 4.5:

OSWorld (computer use) : 61.4 % → 72.5 % (+11.1 pp)

SWE‑bench Verified (programming) : – → 79.6 % (frontier level)

Enterprise document re‑reasoning Q&A : 62 % → 77 % (+15 pp)

Mathematics accuracy : 62 % → 89 % (+27 pp)

PDF/Word extraction : – → >80 %

Public service tasks : – → 88 %

Retail operations : – → 94 %

The most notable jumps are OSWorld rising from 14.9 % to 72.5 % over sixteen months and mathematics accuracy climbing from 62 % to 89 %.

Computer‑Use Capability Leap

Computer Use now allows the model to manipulate desktop software, browsers, and forms without any API, enabling automation of legacy internal systems such as insurance portals, ERP, and hospital scheduling tools. Pace’s CEO Jamie Cuffe reported a 94 % score on a complex insurance computer‑use benchmark, noting autonomous reasoning and self‑correction on failures.

Long‑Context Window and Agent Planning

Sonnet 4.6 expands the context window to 1 million tokens (beta). This directly influences agent workflow design. In the Vending‑Bench Arena simulation of a year‑long business operation, Sonnet 4.6 achieved profits nearly three times higher than its predecessor, reflecting stronger multi‑step planning and resource allocation.

API Changes

Four notable API updates:

effort parameter – previously only in Opus, now available for Sonnet; default “medium” balances speed, cost, and performance.

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=4096,
    thinking={"type":"enabled","budget_tokens":8000},
    # effort: low / medium / high
    extra_headers={"anthropic-beta":"interleaved-thinking-2025-05-14"},
    messages=[{"role":"user","content":"请对以下代码进行安全审计..."}]
)

output_format → output_config.format – old syntax still works but is deprecated.

# Old (still usable)
response = client.messages.create(
    output_format={"type":"json_schema","schema":schema},
    ...
)
# New (recommended)
response = client.messages.create(
    output_config={"format":{"type":"json_schema","schema":schema}},
    ...
)

Free code execution with web search/fetch – when web_search or web_fetch tools are used, code execution incurs no extra token cost, and the engine filters irrelevant results before they enter the context.

Dynamic filtering public beta – models can generate and run their own filtering logic, improving relevance for long‑context agents.

Security Enhancements

Prompt‑injection resistance has been significantly improved, matching flagship Opus 4.6. Anthropic’s evaluation shows Sonnet 4.6’s “harmful” and “high‑risk misalignment” scores are on par or better than recent Claude models, reinforcing safety for browser‑automation agents.

Enterprise Adoption Assessment

Box CTO Ben Kus observed a 15‑point lift in enterprise document Q&A over Sonnet 4.5. Mercury Banking’s Ryan Wiggins described the model as “faster, cheaper, and gets it right the first time.” A comparative chart (not reproduced) places Sonnet 4.6 ahead of Opus 4.6, Gemini 3 Pro, and GPT‑5.2 on the GDPval office‑task benchmark.

Conclusion

Sonnet 4.6 demonstrates that the performance ceiling of mid‑scale models is being systematically raised. Five‑fold OSWorld growth, mathematics accuracy jumping to 89 %, and flagship‑level injection resistance together make it a far more cost‑effective choice for enterprise AI deployments. With fine‑grained effort control, a 1 M token window, and free code‑execution‑with‑search, its practical applicability surpasses any prior Claude release.

API model identifier: claude-sonnet-4-6, now available on Claude.ai, Claude Cowork, Claude Code, and major cloud platforms.