Sonnet 5.5: Great for agents, not for humans

The positioning. Anthropic's Sonnet 5.5 arrives as the middle sibling in the Claude 5.5 family, positioned as a faster, cheaper complement to Opus 5.5. The speaker makes clear this is not a general-purpose upgrade for human users. Benchmarks show it scoring well on Terminal Bench but using far more tokens per task than competitors. The model costs the same as Sonnet 5 per token, but Anthropic claims 30% lower cost per task due to efficiency gains.

The cost trap. The pricing story is more complicated than it looks. Cache reads, which dominate agentic workloads, were not discounted proportionally on Sonnet 5.5. Where Fable 5.1 dropped cache read costs by 90% and Opus by 60%, Sonnet's cache read pricing remains disproportionately high relative to its input/output token costs. The result is that for real-world agent tasks, Sonnet 5.5 can end up costing more than Opus, not less.

Reasoning effort pitfalls. The speaker strongly warns against using 'max' reasoning effort on any Anthropic model. Max reasoning forces a minimum token floor rather than raising a ceiling, creating 1,500% token increases on simple tasks while often degrading performance through overthinking. Low reasoning is also dangerous, as these models perform poorly without sufficient reasoning budget. The sweet spot is medium or high, but even then, Opus handles reduced reasoning better than Sonnet does.

Where Sonnet actually shines. Despite its flaws as a primary model, Sonnet 5.5 excels as a sub-agent for codebase analysis and research. In a controlled benchmark comparing models on a deep audit task, Sonnet performed slightly better than Opus at roughly half the cost and nearly twice as fast. This makes it an ideal tool for Opus or Fable to call upon for investigative sub-tasks rather than for direct human use. The strength is in agent-to-agent orchestration, not human-to-model interaction.

The design disconnect. Sonnet 5.5 performs poorly on front-end design work. In design comparison tests using a common benchmark, Sonnet's output lagged significantly behind both Opus and especially Fable 5.1. The speaker notes that Sonnet produces less steerable, less polished UI results, making it a poor choice for frontend work despite Anthropic's claims about design capabilities. This reinforces the model's best-fit role as a backend research tool.

The competitive picture. The speaker frames this release as a direct threat to OpenAI, which currently lacks competitive offerings in any of these segments. OpenAI's GPT-6 Soul is described as effectively dead on arrival. Anthropic now claims the smartest model (Opus), the most efficient sub-agent (Sonnet), and a subscription plan that delivers roughly $10,000/month of API-equivalent value for $200. The speaker expects OpenAI to respond aggressively, but currently sees no reason to use any OpenAI model for serious coding work.