Anthropic released Claude Opus 5.5 yesterday, the first model in a new Claude 5.5 family, and the figure it wants you to remember is not a benchmark score. It is 40 percent. That is how much less the model costs to run than Opus 5 on typical workloads at default settings. Capability, per Anthropic, lands at roughly Fable 5.1 level on most work.
Three qualifiers in one claim. Most work. Typical workloads. Default settings. Each one is load-bearing.
Cheaper is the feature
For a couple of years the pitch on every frontier release was a higher ceiling. This one is a floor moving down. Anthropic is telling developers they can get near its own top-tier quality without paying top-tier rates, which only works as a message if the audience has already decided the ceiling is high enough. That seems like a fair read of agentic coding right now. Teams are rarely blocked because the model cannot write the function. They are blocked because running an agent across a large repo for six hours costs real money, and nobody enjoys explaining that invoice.
“At default settings” is where I would push hardest. Agentic models burn tokens unevenly, and running cost depends enormously on how much thinking happens before action. A 40 percent reduction on a default configuration could come from a genuinely more efficient model, from a more conservative default reasoning budget, or from both. Those are not the same thing for anyone who overrides defaults, and most serious users override defaults.
The two writeups I read don’t even describe the same claim. MarkTechPost frames it as lower running cost. TechCrunch says lower prices. Those diverge in practice: per-token pricing is a posted number, while running cost folds in how many tokens the model actually spends to finish a job. A model can be cheaper per million tokens and still more expensive per completed task.
The benchmark loop
Anthropic says Opus 5.5 leads in agentic coding on Anthropic’s own benchmarks. Every lab publishes the evaluations it wins, so this is unremarkable, but it means the release-day numbers carry roughly zero independent weight. Useful signal shows up in about a week, when people have pointed it at ugly legacy code and watched hour four of a long-running task, and when somebody posts cost comparisons from real usage instead of a pricing page.
Something else stands out. Opus 5.5 is measured against Fable 5.1 and Opus 5. The comparison set is entirely internal. That is either confidence or a closed loop, and I lean toward confidence while noting that internal ladders are easy to design.
If the cost claim survives contact with real agent workloads, it moves what teams can afford to automate, which matters more than a few points on a coding eval. TechCrunch got Anthropic calling it “the strongest-performing model we’ve tested to date.” The words doing the work in that sentence are “we’ve tested.”