Quick answer. Claude Sonnet 4.6 has the highest Truth Score in the dataset at 96, with privacy 94, safety 93 and a weighted score of 87. The API rate is $3 in and $15 out per million tokens. It is not self-hostable. Choose it for coding, writing and analysis when truth and privacy outrank raw speed or the lowest token price.
Claude Sonnet 4.6 earns this profile because its Truth Score is 96, the highest figure in the published set, alongside safety 93 and privacy 94. Rank is 2 and the weighted score is 87. The decision is whether that reliability is worth the workhorse price for the job, not whether Anthropic is a familiar brand. Task performance is 89. Context is a 200,000-token window, scored 88. Speed is 82. Cost efficiency is 76, with API rates of $3 per million input tokens and $15 per million output tokens. The model is not self-hostable and the jurisdiction flag is the United States. GDPR, HIPAA and SOC 2 are all true. Perspective Score is minus 9 and stays outside the weighted score. Best-fit jobs in the dataset are coding, writing, analysis, high-volume API work and agentic workflows. If the work needs the strongest published truth and privacy marks, start here and confirm the match. If the job is mostly cheap high-speed classification, the scorecard points elsewhere.
Truth Score is why this profile exists
Safety is 93 and the Truth Score is 96. The dataset traces safety to Anthropic's Constitutional AI approach and published harm evals, and task performance to SWE-bench Verified, MMLU and Anthropic evals. Those are source notes in the JSON, not a claim that Best AI Match re-ran the benches. Read the method on methodology and the meaning of the mark on Truth Score.
| Reliability and privacy lens | Published figure |
|---|---|
| Truth Score | 96 |
| Safety | 93 |
| Privacy | 94 |
| Weighted score (rank 2 of 21) | 87 |
| Task performance | 89 |
| Context window (context score 88) | 200,000 |
| Speed | 82 |
| Cost efficiency | 76 |
| API price, input / output per million | $3 / $15 |
| Perspective Score (excluded) | −9 |
Figures from /data/ai-comparison-2026.json. Last verified 2026-06-26. Scoring version v1.0. Affiliate links are off.
Privacy flags on the published record
Privacy is 94. GDPR, HIPAA and SOC 2 are true. Jurisdiction is the United States. The model is not self-hostable, so privacy here is a provider-policy score, not an on-premises deployment. For writing and analysis jobs, pair this profile with best AI for writing. For code, use best AI for coding before you treat the weighted score as a staffing plan.
The workhorse price
Cost efficiency is 76 at $3 and $15 per million tokens. That output rate matches GPT-5.4. The input rate is higher. If the job is high-volume classification where speed and price dominate, the scorecard's own best-fit list does not pretend this is the budget model. See Gemini vs Claude for that trade.
Perspective is disclosed
Perspective Score is −9. It does not move the weighted score. Best-fit jobs in the record are coding, writing, analysis, high-volume API work and agentic workflows. The official site is anthropic.com/claude, linked as the provider page while affiliates stay off.
Still matching a job? Run the match engine, or open GPT-5.4 when integration breadth is the constraint.
FAQ
What is Claude Sonnet 4.6's Truth Score?
The published Truth Score is 96, the highest in the Best AI Match dataset, last verified 2026-06-26.
What does the Claude Sonnet 4.6 API cost?
The dataset lists $3 per million input tokens and $15 per million output tokens.
Does a high Truth Score mean citations can be trusted?
No. Truth Score is a reliability mark. An unsourced citation is a separate problem, covered on the citations page.