There's no point in having sonnet workers anymore

Detailed Analysis

A Reddit post titled "There's no point in having sonnet workers anymore" captures a notable shift in how power users are approaching Claude model orchestration following the release of Opus 5. The author describes an "orchestration skill" workflow—apparently built around a tool or framework called "fable"—that historically spun up cheaper Sonnet-tier models as parallel "worker" processes beneath a more capable orchestrating model. The core claim is that Opus 5 has broken this cost-performance calculus entirely: it now reportedly outperforms Sonnet while also being more token-efficient, undercutting the traditional rationale for using Sonnet as a budget workhorse in multi-agent pipelines. The poster further notes that even high-effort configurations of Opus 5 ("xhigh" and "max") are becoming less relevant, with the new preferred pattern being low-effort Opus 5 workers coordinated by a high-effort Opus 5 or fable-based orchestrator.

This matters because it reflects a broader recalibration happening among developers who build agentic systems on top of Claude. For much of the past two years, the standard architecture for cost-sensitive AI applications involved tiering: reserve the most expensive, most capable model (Opus) for planning, reasoning, or supervision, while delegating repetitive or lower-stakes subtasks to cheaper, faster models like Sonnet or Haiku. This tiering strategy was central to keeping API costs manageable at scale, especially for orchestration frameworks that spawn many parallel "worker" agents to complete a larger task. If Opus 5 genuinely delivers better output at lower effective token cost than Sonnet, it inverts the assumption that capability and price move in lockstep—suggesting Anthropic has made significant efficiency gains in the flagship model rather than only in the smaller ones.

The technical detail about effort levels ("low effort" workers versus "high effort" or "xhigh"/"max" orchestrators) points to Anthropic's increasingly granular control over inference-time compute allocation, a feature set that lets developers dial reasoning depth up or down within the same model family rather than switching between distinct model sizes. This is consistent with an industry-wide trend, seen also at OpenAI and Google, toward exposing "thinking effort" or reasoning-token budgets as a tunable parameter rather than relying solely on model selection to manage cost and latency. If a single model can be throttled to behave like a cheap worker or a premium reasoner depending on the effort setting, it reduces the practical need for maintaining multiple distinct model tiers in a product lineup—a dynamic the poster is implicitly flagging when questioning "what exactly is the point of sonnet now."

The post also underscores how quickly community-built orchestration patterns must adapt to new model releases. Frameworks like the one referenced ("fable") that hard-code assumptions about which model tier is cheapest or fastest can become suboptimal or obsolete within days of a new flagship release, forcing developers to re-benchmark their pipelines. This is emblematic of a larger pattern in the generative AI ecosystem: as frontier labs like Anthropic ship models with improved efficiency alongside improved capability, the economic logic underpinning multi-agent and mixture-of-model architectures is in constant flux, and yesterday's cost-optimization strategy can quickly become today's unnecessary complexity.