Skip to content
CrucibleMark

Claude Sonnet 4.6

Where the Opus class is too expensive, Claude Sonnet 4.6 steps in: coding, computer use, and agentic workflows at near-Opus level, at the lower Sonnet price. The model operates with adaptive thinking in three effort levels, processes text, images, and PDF documents, and offers a context window of one million tokens, generally available since March 2026.

Anthropic Version 4.6 Commercial use permitted Dense 1000 K Context 08/2025 $3 / $15 per 1M

  • Proprietary
  • Frontier
  • API
  • Text
  • Vision
  • Long Context
  • Interactive

Sovereign Risk: MEDIUM Anthropic is a US-based provider; relevant risks relate to cloud processing under US law, as no open weights are available.

Tool-use profile: 6 assets in detail

Comparison of asset performance (P1/P2/Combined) against the fleet average

Asset performance (radar)

Score Breakdown vs. Fleet Average

Tool-use details

Asset performance, reliability, and runtime profile

Name P1 P2 Combined
EU License Research 100 15 49
HTTP Fetch & Extract 80 100 90
Tool Failure Handling (404) 40 40 37
Web Search & Tool Selection 100 60 82
URL Construction & Fetch 80 80 80
Multilingual Search & Synthesis 100 15 57.5

CrucibleMark evaluates tool use across 6 independent tests. Click a test name for details.

Reliability

  • Tool Call Valid: No
  • Retry: Not required
  • Hallucination: Detected

Reliability measures how consistently a model actually executes tool calls: Tool Call Valid schema and format accepted, Retry Required successful only after retry, Hallucination Flag fabricated tools or parameters detected. All three green means production-ready.

Runtime profile

Call 1
61.88
First Request
MCP
0.81
Protocol Latency
Synthesis
21.12
Response Generation
Total
502.85
Sum of All Phases
Token
33485
Input + Output
Cost
$0
Cost per Run

The runtime profile shows the latency and cost metrics for the model run: Call 1 First Request, MCP Protocol Latency, Synthesis Response Generation, Total sum of all phases, supplemented by Token input and output and Cost cost per run.

Tool-use review

· Long Context

Deployment Verdict

Conditional deploy, because tool execution is robust, but the detected hallucination in the honeypot breaks trust in any fact-critical tool pipeline.

Tool Execution Profile

Claude Sonnet 4.6 behaves with discipline at the MCP level. Tool calls were valid, no retry was required, and the P1 score of 83.33 indicates a robust operational baseline. What matters here is not just format compliance, but tool selection: in the Web Search & Tool Selection test — which checks whether the model chooses web_search over fetch without an explicit hint — the model made the correct decision consistently. This argues against rigid pattern-following and in favor of genuine orchestration logic.

In the URL Construction & Fetch test, which measures autonomous derivation of a target URL followed by a fetch, the model remains usable but not deterministic enough for fragile pipelines. P1=80 is solid, but not a signal of high precision when constructing endpoints independently. For agentic workflows with clearly defined tool boundaries, this is deployable. For pipelines where the model must autonomously form URLs or query paths, guardrails are required.

Synthesis Fidelity

How well does it condense tool results? Inconsistently. Claude Sonnet 4.6 can synthesize extracted content very effectively when the input is clearly structured — visible in HTTP Fetch & Extract with P2=100. As soon as the task becomes more research- and interpretation-driven, synthesis quality drops sharply. EU License Research and Multilingual Search & Synthesis both land at P2=15. This is not a general summarization problem; it is a fidelity problem under uncertainty.

Does it stay within the tool result, or does it fall back on training? No, not reliably. In the honeypot EU License Research — which checks whether current license restrictions are answered from web sources rather than from training — a hallucination was detected. Content Verification State B1 at P2=15 is not merely a quality deficiency; it is a security risk. When a model outputs fabricated or unconfirmed facts as the result of a tool-based research task, it undermines the control function of the entire infrastructure.

Error Resilience

In the 404 test — which checks for transparent error communication versus fabricated fallback content — the model responded in a production-appropriate manner. P2=80 and no hallucination despite the error show that it surfaces failures rather than inventing page content. This is acceptable for real-world tool chains.

Operational Profile

38.46s first call, 16.51s second call, 339.33s total. Slow relative to the synthesis quality achieved. MCP latency of 1.58s is non-critical. 0.296922 USD per run: moderately priced, but not a bargain given the trust gap.

Conclusion & Recommendation

Suitable for agentic pipelines with clear tool guidance, structured extraction, and tolerable response times. Not suitable for compliance, policy, licensing, legal, or other high-trust research chains where the model must remain strictly bound to tool findings. If you deploy it, do so only with downstream verification, enforced source constraints, and a hard separation between extraction and final output.

This analysis was generated automatically based on benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and full methodology are documented in the GitHub project.