Anthropic on Monday released Claude Sonnet 5.5, an updated mid-tier model designed for everyday professional work including coding, bug fixes, and generation of documents, slides, and spreadsheets. The company states the model runs more than 30 percent faster than its predecessor, Sonnet 5, and uses tokens at a lower rate for many tasks, resulting in up to 30 percent lower cost per completed task even though list prices remain unchanged.

Pricing stays at $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20 per million tokens. That rate matches Sonnet 5 and stands at half the cost of Anthropic’s flagship Opus 5.5, which lists at $4 and $20 respectively. The new model is positioned for well-scoped tasks where sustained complex judgment is not required, while Opus 5.5 continues to be recommended for work demanding deeper reasoning.

Coding Benchmarks and Independent Results

On Terminal-Bench 4.0, a measure of whether an AI agent can complete complex professional tasks through autonomous command-line work, Anthropic reported Sonnet 5.5 scoring 70.6 percent. That figure exceeds the 66.4 percent recorded for Opus 5.5 and the 10.3 percent posted by Sonnet 5. Independent testing by Artificial Analysis produced similar relative rankings, with Sonnet 5.5 at 63.6 percent against 59.6 percent for Opus 5.5 and 59.1 percent for OpenAI’s GPT-6 Astra.

On the GDPval-AA evaluation, which assesses real-world professional performance across 44 occupations using an Elo ranking system, Sonnet 5.5 scored 1844 compared with 1846 for Opus 5.5, effectively a statistical tie. GPT-6 Sol registered 1487 on the same measure. Anthropic notes that Sonnet 5.5 at high effort settings matches certain GPT-6 Sol results on FrontierCode at roughly one-fifth the cost per task.

The model also demonstrates cyber-related capabilities comparable to those of Opus 5 and is the first Sonnet version subject to the same safeguard framework applied to higher-tier Claude models. Its ability to coordinate multiple agents without immediately exceeding cost thresholds is cited as one factor behind the coding gains.

Token Efficiency and Cost Trade-Offs

While list prices are unchanged, Anthropic states that lower default effort settings deliver meaningful savings. At medium effort, the setting used by default in its applications, the company claims Sonnet 5.5 exceeds Sonnet 5’s best coding scores at less than one-tenth the prior cost. Independent measurements, however, show that at maximum effort the model generates approximately 193,000 tokens per test task—the highest volume recorded in those evaluations—and roughly 60 percent more tokens than Opus 5.5. That usage translated to an estimated $7.60 per task in the independent tests, about 50 percent higher than Sonnet 5 under similar conditions.

Artificial Analysis identified high-effort mode as offering the strongest value among the settings it examined. For routine enterprise workloads the practical outcome is near-flagship coding performance at a fraction of the higher-tier price, provided effort dials remain moderate.

Broader Industry Shift Toward Cost Per Task

The release occurs as evaluation criteria across the sector move beyond pure benchmark scores toward price-performance and cost per completed task. Enterprise customers, which account for the majority of Anthropic’s revenue, increasingly select the least expensive model that reliably finishes required work. OpenAI recently aligned GPT-6 Sol pricing with the same $2 and $10 rates, while its mid-tier GPT-5.6 Terra lists at $2 and $12.

Longer-term data show rapid declines in the cost of achieving a given performance level, with one analysis estimating roughly 47 percent quarterly reductions since 2023. At the same time, industry forecasts indicate that total inference spending per agentic workflow could rise more than fivefold through 2028 as cheaper tokens enable more complex multi-step processes.

A smaller model, Claude Haiku 5.5, remains scheduled for release in the coming weeks and is expected to target high-volume, cost-sensitive applications. Anthropic continues to position Opus 5.5 as the stronger choice for tasks requiring extended judgment, while presenting Sonnet 5.5 as the practical option for the bulk of daily professional workloads.