The interesting thing about a new frontier model isn’t whether it’s smarter — it’s what it does to the arithmetic of running agents. Artificial Analysis’s breakdown puts Opus 5.5 at the top of its intelligence index (58, ranked first), but the numbers I actually budget around are less flattering: $4 per million input tokens, $20 per million output, and a model flagged as “very verbose” — 260M output tokens across the eval suite against a median of 88M.
That verbosity is the tension. The pitch is efficiency — around 40% fewer tokens and less code for the same result — which genuinely matters when you’re fanning out dozens of tool calls per agent run and token spend, not latency, is the line item that scales with your multi-agent topology. But a model that reasons at length by default can quietly claw those savings back in production, and “cheaper per token” is not the same as “cheaper per finished task.” The HN discussion has the predictable split between people quoting the price drop and people who’ve already watched a bill go up.
The wider read tracks that ambiguity. Digital Trends frames the launch as top-tier performance at a 40% lower price — the clean cost-per-capability story. The code-quality reviewers are more measured: CodeRabbit found it catching a different mix of bugs with real but modest coverage gains, alongside higher token usage in every configuration they tested, and told teams to verify the lower per-token price actually lowers review costs. Sonar saw the same double edge: 27.5% less code and 42% fewer total findings, yet a 12% rise in per-line bug density and 44% more concurrency issues. The reception is landing exactly where the economics do — a clear win on paper that only pays off if you measure cost per task, not per token.