Claude Sonnet 5
v5Anthropic
Anthropic's Sonnet-tier flagship (released 2026-06-30), positioned as the cheaper way to run agents: 72.7% SWE-bench Verified (vs Sonnet 4.6's 62.3%), 80.4% Terminal-Bench 2.1, 81.2% OSWorld-Verified — approaching Opus 4.8 at a fraction of the cost. Same $3/$15 standard price as Sonnet 4.6 with intro $2/$10 through 2026-08-31, 1M context, effort parameter incl. xhigh, and deliberately low cyber capability with default-on safeguards.
Trust Vector Analysis
Dimension Breakdown
🚀Performance & Reliability+
Biggest Sonnet-generation jump in the registry: +10.4 points SWE-bench Verified and +13.4 Terminal-Bench 2.1 over Sonnet 4.6, closing most of the gap to Opus 4.8 (and edging it on GDPval-AA v2). Scores held slightly conservative pending independent benchmark replication — the model is nine days old.
Provider-published coding benchmarks cross-checked against independent launch coverage; SWE-bench Verified, SWE-bench Pro, and Terminal-Bench 2.1
Provider-published reasoning benchmarks; independent replication still limited nine days post-launch
Knowledge-work and computer-use benchmark review from provider announcement and third-party comparisons
Review of documented behavior controls; repeated-run consistency data still limited post-launch. Note the updated tokenizer maps identical input to roughly 1.0-1.35x more tokens than Sonnet 4.6
Median latency for API requests with standard prompt sizes; limited launch-window sample
95th percentile response time across diverse workloads; limited launch-window sample
Specification review from launch documentation and coverage. Max output tokens not yet confirmed in public docs
Platform uptime from official status page; model-specific track record is only ~9 days old, so score held slightly below the platform baseline
🛡️Security+
Strong safety posture with a notable design choice: offensive-cyber capability is intentionally kept much lower than the Opus line, with cyber safeguards on by default. Anthropic reports fewer undesirable agentic behaviors than Sonnet 4.6.
Review of provider safety evaluations against OWASP LLM01 patterns; third-party red-team data still limited at launch
Testing against adversarial prompt datasets and review of provider safety documentation
Analysis of privacy policies and data handling practices
Review of provider safety evaluations across harmful content and offensive-cyber categories
Review of API security features and best practices
🔒Privacy & Compliance+
Inherits Anthropic's full enterprise privacy posture from day one: ephemeral data handling, SOC 2 Type II, GDPR, HIPAA eligible.
Review of enterprise documentation and privacy policies
Analysis of privacy policy and data usage terms
Review of terms of service and data retention policies
Review of data protection capabilities and customer responsibilities
Verification of compliance certifications and audit reports
Review of data handling practices
👁️Trust & Transparency+
Transparent launch documentation, including an unusually candid cyber-capability section and a public benchmark-methodology correction. Knowledge cutoff not yet published.
Evaluation of reasoning transparency and explanation capabilities
Review of provider evaluations; independent factual-QA measurement still limited at launch
Evaluation on bias benchmarks and diverse demographic testing
Qualitative assessment of confidence expression in outputs
Review of documentation completeness and clarity
Review of public disclosures about training data
Analysis of built-in safety mechanisms
⚙️Operational Excellence+
Drop-in adoption for existing Claude users, with one real migration caveat: the updated tokenizer maps the same input to roughly 1.0-1.35x more tokens than Sonnet 4.6, so per-request cost can rise even at identical per-token prices. Vertex AI availability still pending as of 2026-07-09.
Review of API design, consistency, and feature completeness
Review of SDK quality, documentation, and maintenance
Review of versioning policy and historical practices
Review of available monitoring tools and metrics
Assessment of documentation, community, and support responsiveness
Analysis of availability surfaces and third-party integrations; score held below Sonnet 4.6 until Vertex AI availability lands and framework defaults migrate
Review of licensing terms and restrictions
- +72.7% SWE-bench Verified and 80.4% Terminal-Bench 2.1 — a 10+ point generation-over-generation jump at the same standard price
- +Near-Opus performance (57.4% HLE with tools vs Opus 4.8's 57.9%; edges Opus 4.8 on GDPval-AA v2) at $3/$15
- +Introductory $2/$10 pricing through 2026-08-31 makes agent migration trials cheap
- +1M token context window with effort parameter (low/medium/high/xhigh)
- +Self-verification behavior and lower hallucination/sycophancy rates than Sonnet 4.6
- +Deliberately low offensive-cyber capability with default-on safeguards — a safety plus for most deployments
- +HIPAA eligible with ephemeral data handling from day one
- !Released 2026-06-30 — independent benchmark replication and production track record still limited
- !Updated tokenizer maps identical input to roughly 1.0-1.35x more tokens than Sonnet 4.6, partially offsetting per-token price parity
- !Google Vertex AI availability still pending at evaluation time (AWS and Microsoft Foundry live)
- !Intentionally weak at sanctioned cybersecurity work — use Opus tier for authorized cyber tasks
- !Knowledge cutoff and max output tokens not yet published
- !Opus 4.8 still leads on the hardest agentic coding (69.2% vs 63.2% SWE-bench Pro)
Use Case Ratings
code generation
72.7% SWE-bench Verified and 80.4% Terminal-Bench 2.1 — near-Opus agentic coding at Sonnet pricing. The new best default for production coding agents.
customer support
The Sonnet support sweet spot, now smarter at the same price; use effort 'low' for high-volume tiers. Intro pricing makes migration trials cheap through August.
content creation
Strong long-form and marketing content with fast turnaround; Opus/Fable tiers still lead on the most nuanced pieces.
data analysis
1M context for large datasets with GDPval-AA v2 knowledge-work performance that edges Opus 4.8, at workhorse pricing.
research assistant
1M context, strict BrowseComp improvement over Sonnet 4.6, and strong synthesis. Opus/Fable preferred for the hardest research.
legal compliance
Anthropic cites legal research and insurance workflows as target use cases; HIPAA eligible with 1M context for contract repositories.
healthcare
HIPAA eligible with strong privacy controls; real-world clinical validation is still early nine days post-launch.
financial analysis
Near-Opus reasoning (57.4% HLE with tools) at predictable cost; escalate the hardest modeling to Opus 4.8 or Fable 5.
education
Fast, patient explanations at a price point that scales; default model on Free/Pro plans broadens reach.
creative writing
Capable creative writing with good narrative flow; Opus tier produces more distinctive prose.