**Anthropic** unveiled its next-generation mid-tier AI model for Monday, ‘Claude Sonnet 5.5.’ The company said the model is more than 30% faster at processing than its predecessor and reduces the cost per task by up to 30%.
Key Points
Sonnet 5.5 scored 70.6% on Terminal-Bench 4.0, a coding benchmark. This is a big jump from Sonnet 5’s 10.3%.
Anthropic first applied its top-tier cybersecurity safeguards to the Sonnet line, having previously used them only on its highest-end models.
However, a single independent benchmark found that, under the maximum performance mode, Sonnet 5 costs more per task than Sonnet 5.
Claude Sonnet 5.5 benchmark
Sonnet 5.5 is the second model in the Claude 5.5 family, revealed just six days after the release of the flagship model ‘Claude Opus 5.5’ (September 22). In the release announcement, Anthropic introduced Sonnet 5.5 as a “faster and cheaper” model designed to complement Opus 5.5. It is described as suitable for day-to-day work, clearly defined tasks, bug fixes, and refining work on documents, slides, and spreadsheets.
Pricing was set at $2 per 1 million input tokens and $10 per 1 million output tokens.
On Terminal-Bench 4.0, an agent-based coding test, Sonnet 5.5 scored 70.6%. This is higher than not only Sonnet 5’s 10.3% but also the more expensive Opus 5.5 (66.4%). Sonnet 5.5 also became the first Sonnet model to clear Pokémon Red (Pokémon Red) by looking only at screenshots. It is a benchmark designed to test long-task performance and image understanding.
The new model is available on all platforms supported by Anthropic, including Amazon Web Services (Amazon Web Services), Google Cloud, and Microsoft Azure. The company also said it plans to launch a smaller-scale ‘Claude Haiku 5.5’ within the next few weeks. Haiku 5.5 is a model aimed at high-volume calls and cost-sensitive demand, and once it is released, the three-model lineup of the 5.5 generation will be complete.
Related article: OpenAI adjusts the pace of Frontier AI development after the September 20 ‘sandbox escape’ incident
Safety features and cost structure of Sonnet 5.5
In its release announcement, Anthropic cited an assessment by Epic Games’ (Epic Games) Chief Operating Officer (COO) **Daniel Vogel**. Vogel said that Sonnet 5.5 handled gameplay system code spanning tens of thousands of lines in initial tests, and added that it “cleared the quality standards expected from a top-tier model with ease.”
Based on evaluations that Sonnet 5.5’s cybersecurity performance reached a level similar to Opus 5, Anthropic has for the first time moved security measures previously applied only to the highest-tier models into the Sonnet line. While it does not affect routine bug fixes, requests with higher cybersecurity risk are designed to be noticeably ‘downgraded’ to Sonnet 5. In addition, a new classifier was introduced to block attempts to extract reasoning processes in reverse.
However, not all tests necessarily support Anthropic’s claim of ‘cost savings.’ Research firm Artificial Analysis estimated the weighted average cost of the benchmark tasks performed by Sonnet 5.5 in the maximum performance mode at $7.60 per task. That is actually higher than Sonnet 5’s $5.09. Still, under the same conditions, Sonnet 5.5’s index score increased from 38 to 56. It is possible to interpret this as indicating that more weight was placed on absolute performance gains rather than cost efficiency.
Before this launch, Anthropic unveiled its flagship model Opus 5.5 on September 22 and cut the price for 1 million input tokens by 20% to $4, while adjusting the price for 1 million output tokens to $20 (a 20% reduction). At the time, the company said that, for typical workloads, it reduced costs by about 40% compared to Opus 5 and increased the output generation speed by more than 30%. On the same day, **OpenAI** also released GPT-6 Sol and GPT-6 Luna, offering prices at roughly half the level of previous-generation models.
Next article: BlackRock forecasts, “Computing power ultimately heads toward the gift market”