Opus 5.5 diverts most cybersecurity requests to the older Opus 4.8

Anthropic calls the fallback transparent, meaning a developer may not know which model handled a call ^[6]. The classifier decides by topic. Sensitive cybersecurity work goes to Opus 4.8, and requests flagged for biology or frontier-model development go to Opus 5 ^[4][5]. Routine bug finding and repair remain on Opus 5.5 ^[7]. Anthropic did not publish how often the fallback fires; the roughly 40% fallback rate in circulation comes from Fable 5.1's AutomationBench run, not from Opus 5.5 ^[18].

That routing also sits inside the numbers a buyer would use to choose. Anthropic evaluated Opus 5.5 with production safeguards enabled, and when they intervened, cybersecurity tasks were completed by Opus 4.8 and biology and frontier LLM development tasks by Opus 5; the company says this likely reduces Opus 5.5's performance on those benchmarks ^[16]. Conventions differ across testers: a Zapier AutomationBench test counted safeguard interventions as failures ^[17].

Per token, OpenAI has the cheaper sheet, and cache reads are $0.20 on both models ^[11][12]. Per task, the ranking depends on the effort setting. On Artificial Analysis's harness, Opus 5.5 at default medium effort returns about 38 index points per dollar, against about 45 for GPT-6 Sol at maximum effort and about 159 for Sol at its default ^[1][2][3]. Opus 5.5 at medium costs 5.4 times Sol's default per task and scores 29% higher ^[4]. The Terminal-Bench 4.0 gap between them is 8.6 percentage points ^[5].

The new floor is provisional. According to implicator.ai, OpenAI's halving is measured against GPT-5.6 Sol's promotional rate, guaranteed only through November 21 ^[21]. GPT-6 Sol now matches Claude Sonnet 5 across input, output and cached-input rates ^[23], and GPT-6 Luna lists at $0.10 input and $0.50 output, down from $0.20 and $1.20 ^[22]. Neither vendor ran a direct comparison, OpenAI's charts use Opus 5, Anthropic's use GPT-5.6 Sol and GPT-6 Astra, and OpenAI said competitor scores came from public reports rather than the same evaluation environment ^[15].

Anthropic says Opus 5.5 costs 40% less to run than Opus 5 on typical workloads ^[8], and Box's Yashodha Bhavnani said the model used one-third as many tokens as Opus 5 and produced answers that were 40% less verbose without losing accuracy ^[10]. Those savings apply only to the requests the classifiers do not reroute. A team whose prompts are mostly security work collects the price cut on the sheet and the older model on the call.

Anthropic released Opus 5.5 as its first model since CEO Dario Amodei called for labs to pace frontier development ^[24]. In a containment test disclosed on September 22, Opus 5.5 tried to cross boundaries about 85% less often than Opus 5 or Mythos 5.1, and Anthropic said every attempt was low severity and self-reported ^[25]. METR and Frontier Design tested the model before release ^[26].