Full Breakdown
Anthropic Unveils Claude Sonnet 5.5: Faster, Cheaper Model with Enhanced Cyber Safeguards
By Drooid · · How we work
Core Release Details
On September 28, 2026, Anthropic released Claude Sonnet 5.5, the second model in its Claude 5.5 family. It is positioned as a low-cost complement to Claude Opus 5.5 for routine programming, document creation, and other well-scoped tasks. The model is available on AWS, Google Cloud, and Azure under the ID claude-sonnet-5-5.
Background & Context
Earlier in 2026, Anthropic introduced Opus 5.5, a more expensive model for complex work. CEO Dario Amodei has urged AI firms to slow frontier model development, emphasizing alignment and safety. Sonnet 5.5 follows that stance, offering improvements without extending raw capability.
Performance Benchmarks & Pricing
- Speed & Cost: Output is >30 % faster and per-task costs up to 30 % lower than Sonnet 5.
- Pricing: $2 per million input tokens and $10 per million output tokens (plus cache rates of $0.20 and $2.50 per million).
- Coding Benchmark: Terminal-Bench 4.0 score 70.6 %, versus 10.3 % for Sonnet 5 and 66.4 % for Opus 5.5.
- Occupational Test (GDPval-AA): Score 1,844, two points below Opus 5.5 and ~400 points above Sonnet 5.
- Other Evaluations: FrontierCode 1.1 max-effort 46.2 % (Sonnet 5.5) vs 42.4 % (Sonnet 5); CursorBench 4.0 55.5 % vs 34.1 % for Sonnet 5.
Safety Enhancements
Sonnet 5.5 is the first Sonnet model with frontier-style cyber safeguards previously reserved for Opus 5.5:
- Classifiers that block reasoning-extraction attempts.
- A three-stage cyber-risk pipeline (activation probe, lightweight on-model classifier, separate LLM classifier).
- Reported 99.43 % recall on a cyber-harm coverage set and a 21.0 % attack success rate, improved from 57.2 % for Sonnet 5.
- Automated audits across ~1,850 scenarios show lower sandbox-escape attempts and comparable alignment metrics to Sonnet 5.
Official Statements & Responses
- Theo Chu, research product manager at Anthropic, said:
> “Sonnet is really for the cost-conscious customer where they might not need as much intelligence.”
- The firm noted that higher-risk cyber requests will fall back to the older Sonnet 5 model.
Conflicting Reports & Gaps
- Pricing Timeline: Thenextweb reported introductory rates of $2/$10 were slated to rise after August 31; Anthropic’s September 28 announcement lists the same rates, leaving the change unverified.
- Launch Timing: Leaks on September 22 suggested a launch on September 30 or October 1, but Anthropic confirmed only the September 28 release.
- Benchmark Claims: Social-media posts claim Sonnet 5.5 outperforms OpenAI’s GPT-6 Sol on real-world tests; these remain unverified and are not in Anthropic’s published tables.
What’s Next
- Anthropic announced Claude Haiku 5.5, aimed at high-volume, cost-sensitive applications, will join the Claude 5.5 family “in the coming weeks.”
- A Cyber Verification Program will soon grant vetted defenders tiered access to advanced capabilities across Sonnet 5.5, Opus 5.5, and Claude Mythos models.
- For accounts created after August 31, 2026, replaying old thinking blocks after modifying system prompts or tools will trigger errors by default, reinforcing the new safety architecture.