Claude Sonnet 5

Claude Sonnet 5 has been the new default for free and pro users since late June 2026. The successor to Sonnet 4.6 brings Adaptive Thinking as default behavior, speed close to Opus, and processes text and images with a 1,000,000-token context window. Safety is significantly improved, with better resistance to prompt injection and fewer hallucinations.

Anthropic Version 5 Commercial use permitted Dense 1000 K Context 01/2026 $2 / $10 per 1M

Proprietary
Frontier
Anthropic
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 details

Asset performance, reliability, and runtime profile

Name	P1	P2	Combined
EU License Research	100	60	76
HTTP Fetch & Extract	80	100	90
Tool Failure Handling (404)	80	80	80
Web Search & Tool Selection	100	80	91
URL Construction & Fetch	80	60	70
Multilingual Search & Synthesis	100	40	70

Name

Combined

EU License Research

100

HTTP Fetch & Extract

100

Tool Failure Handling (404)

Web Search & Tool Selection

100

URL Construction & Fetch

Multilingual Search & Synthesis

100

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

Reliability

Tool Call Valid: No

Retry: Not required

Hallucination: Not 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: 2.48; First Request

MCP: 0.89; Protocol Latency

Synthesis: 10.12; Response Generation

Total: 80.95; Sum of All Phases

Token: 16031; 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.

Deployment Verdict

Conditional deploy: tool execution is strong, but calls were not consistently valid and protocol-compliant, and synthesis quality is too uneven for production decision pipelines.

Tool Execution Profile

Claude Sonnet 5 demonstrates genuine tool intelligence rather than rigid pattern-matching. On the Web Search & Tool Selection test — which checks whether the model selects web_search over fetch without being prompted — it reliably identifies the correct access path. This speaks to workable orchestration in open MCP pipelines. It also executes the required steps consistently on EU License Research and Multilingual Search & Synthesis.

Execution weakens where the model must derive a target URL on its own. On the URL Construction test, it constructs the target URL adequately, but not with the precision required for deterministic pipelines. This is not a planning failure — it is a precision problem in the final step. More critically, the tool call overall was flagged as invalid. For production, this means: often correct tool selection on the technical merits, but MCP integration only with strict schema validation, guardrails, and potentially a call sanitizer ahead of actual tool dispatch.

Synthesis Fidelity

How well does it compress tool results? Adequately — but not at the level one should blindly approve for reliable executive summaries or compliance extracts. The strong 100-point score on HTTP Fetch & Extract shows it can cleanly pull structured web content. By contrast, EU License Research and URL Construction & Fetch fall off noticeably in compression, and on Multilingual Search & Synthesis the German-language consolidation is the weakest point. The pattern is clear: solid extraction, inconsistent synthesis.

Does it stay within tool results or fall back on training? On the honeypot EU License Research test — which checks whether current license restrictions are actually retrieved from web sources — the model stays within the queried results. No hallucination was detected. That is a trust signal. However, the P2 score of 60 indicates that accurate retrieval does not automatically translate into a precise, reliable final summary.

Error Resilience

The model is production-ready when it comes to tool failures. On the 404 test — which checks for transparent error communication versus fabricated fallback content — it does not hallucinate page content and communicates the failure in a traceable way. This is exactly the behavior a tool pipeline requires: a visible failure rather than an invented success.

Operational Profile

Call 1: 2.48s. Call 2: 10.12s. MCP latency: 0.89s. Total: 80.95s.
Cost/run: local.
For the performance shown: on the slower end for end-to-end runs. Cost not assessable here.

Conclusion & Recommendation

Suitable for agentic research, routing, and tool-selection pipelines where a downstream validator catches structural errors and a human or rule set reviews the final synthesis. Not suitable as an uncontrolled terminal instance for compliance syntheses, multilingual decision memos, or workflows where a formally invalid tool call already constitutes an incident. Teams deploying Claude Sonnet 5 should give it the tools — but not the final word without a safety net.

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.