Sonnet 5.5 hits third place by writing 410M tokens

Claude Sonnet 5.5 at max effort scores 56 on the Artificial Analysis Intelligence Index. That puts it third of 216 models in its comparison class, where the median score is 26. To get there it generated 410M output tokens across the index’s ten evaluations, against a class median of 88M. Both figures come from Artificial Analysis’s page for the model, which Anthropic released on September 28.

The configuration matters. The page tests “Adaptive Reasoning, Max Effort, Default Fallback”, the top of the effort scale, and notes that a non-reasoning variant may also exist. The 56 and the 410M both describe the model when it is allowed to think at its maximum setting.

Index v4.3.2 leans heavily on agentic work. Four of its ten evaluations have the model act over several steps: AA-Briefcase v1.1 for knowledge work, GDPval-AA v2.1 for real-world work tasks, AutomationBench-AA for SaaS workflows, and Terminal-Bench 4.0 for coding and terminal use. The other six are SciCode, Humanity’s Last Exam, GDP.pdf, CritPt, AA-Omniscience and the long-context test AA-LCR v1.1. Proprietary models are ranked only against others in the same price band, which for Sonnet 5.5 means models with a blended price above $1 per 1M tokens at a 3:1 input-to-output ratio.

The per-token price is ordinary for that band. Anthropic charges $2.00 per 1M input tokens and $10.00 per 1M output, and both equal the class medians. What drives the bill is volume, because a reasoning model’s thinking is billed at the output rate. Artificial Analysis averages the cost at $7.60 per index task. That figure weights each evaluation and covers input, cache writes, cache hits, reasoning and answer tokens.

Anthropic’s Sonnet page pitches the model the other way round. It says Sonnet 5.5 “needs far fewer tokens than Sonnet 5, making it practical to run repeatedly” on long-horizon agent tasks, and that it costs up to 30% less than Sonnet 5 to run for typical workloads billed by token. Yet at max effort it still wrote more than its bigger sibling: in September, Opus 5.5 in the same max-effort configuration used 260M output tokens to score 58. A cheaper Sonnet does not mean a shorter one.

Speed is the missing column. Artificial Analysis lists output tokens per second as N/A, so nobody has independently measured how long 410M tokens take to generate. Anthropic says the model runs 30% faster than Sonnet 5.

Anthropic positions Sonnet 5.5 as “a faster, lower-cost complement to Opus 5.5” for well-scoped everyday tasks in coding, agents and knowledge work. The specs fit that role: a 1M-token context window, text and image input, text output. Artificial Analysis counts three API providers. Anthropic lists the Claude Platform plus Amazon Web Services, Google Cloud and Microsoft Foundry.

If you run agents on Sonnet 5.5, the per-token price stays fixed and the billing options are where you have room:

• Prompt caching cuts the cost of cached input by up to 90%. At Artificial Analysis’s 7:2:1 mix of cache hits, fresh input and output, the blended rate drops to $1.54 per 1M tokens.
• Batch processing takes 50% off, for work that doesn’t need an immediate answer.
• US-only inference costs 1.1x the standard price, which matters if data residency requires it.

Effort is a separate setting. Anthropic says “API users have fine-grained control over the model’s thinking effort”, and the 410M-token run sits at the top of that scale. Set effort per request on claude-sonnet-5-5 rather than defaulting to max.