Anthropic on June 30, 2026 released Claude Sonnet 5 — a new mid-tier model that becomes the default for Free and Pro plans. The model closes the performance gap between the Sonnet and Opus class, reigniting price competition in the agentic AI segment.
Key takeaways
- Launch pricing: $2 per million input tokens and $10 per million output tokens — valid through August 31, 2026
- After August 31 pricing moves to $3/$15 per million input/output tokens
- Sonnet 5 is the default model across all Anthropic plans (Free, Pro, Max, Team, Enterprise)
- On OSWorld-Verified benchmark, the model outperforms Sonnet 4.6 and approaches Opus 4.8
- Available in Claude Code and via API as claude-sonnet-5
How Sonnet 5 performs on agentic benchmarks
Anthropic published results on two key agentic benchmarks: BrowseComp (multi-step web research) and OSWorld-Verified (computer use). In both, Sonnet 5 clearly surpasses Sonnet 4.6 and approaches Claude Opus 4.8 at a significantly lower per-token cost.
The key difference: for tasks where top-tier accuracy is not required, Sonnet 5 reduces operational costs for agentic workflows without requiring the more expensive Opus 4.8 ($5/$25 per MTok). Anthropic illustrates this through cost-performance curves at varying effort levels — both Sonnet 5 and Opus 4.8 now cover a single continuous capability range, at different price points.
Safety and behavior improvements
Pre-deployment safety evaluations show Sonnet 5 produces fewer undesirable behaviors than Sonnet 4.6 — lower rates of hallucination, sycophancy, and cooperation with misuse. Anthropic notes, however, that the model shows a slightly higher misalignment rate than Opus 4.8 and Claude Mythos Preview.
On cybersecurity evaluations, Anthropic deliberately did not train Sonnet 5 for offensive tasks. The model failed to construct a working exploit in any test. The model ships with real-time cybersecurity safeguards enabled by default. Developers can access via Claude Code and the API.
Availability and pricing
Sonnet 5 is available today across all plans: Free, Pro, Max, Team, and Enterprise — as the default model replacing Sonnet 4.6. The model uses an updated tokenizer that processes the same text in 1.0–1.35× more tokens. Developers can access the model through Amazon Bedrock, Vertex AI, and Microsoft Foundry.
Why it matters
Sonnet 5 changes the economics of agentic AI systems. Until now the choice was: cheaper Sonnet with limited agentic capabilities, or more expensive Opus with full capabilities. Sonnet 5 eliminates that dichotomy — a single cost-performance range now spans both models, letting developers match effort level (and cost) to the specific task.
For companies building multi-step agentic pipelines (coding, computer use, complex research), Sonnet 5 cuts operational costs without sacrificing Opus-class quality. At scale, a difference of a tenth of a cent per token translates to tens of thousands of dollars per month in intensive deployments.
What's next?
- Price increases to $3/$15 per MTok after August 31, 2026 — already announced
- Anthropic's Claude Science beta for scientists is live as a separate product on the same API
- Updated rate limits for higher effort levels are already active on the platform
Sources
- Anthropic — Introducing Claude Sonnet 5
- TechCrunch — Anthropic launches Claude Sonnet 5 as a cheaper way to run agents
- VentureBeat — Anthropic launches Claude Sonnet 5 at a steep discount
