Box measured Claude Opus 5.5 using a third of the tokens Opus 5 needed

Box's token ratio speaks to both the billing and the reading, and it is the hardest figure in this launch to check. Verbose output gets billed twice, once as output tokens and once as the minutes a reviewer spends scrolling to the part that answers the question, and only the tokens show up in a cost model.

Two numbers in the coverage both read 40 percent and count different things. ZDNET's framing is that Opus 5.5 delivers Fable 5.1 performance for most work and costs about 40 percent less to run ^[1]. Yashodha Bhavnani, VP of AI Products at Box, said that in the company's evaluations Opus 5.5 "used a third of the tokens Opus 5 did, and its answers were 40% less verbose without losing accuracy" ^[13].

The published price is the part a buyer can hold Anthropic to: tokens on Opus 5.5 cost 20 percent less than on Opus 5 ^[3]. Stack that on Box's ratio and the same task bills at 0.8 times a third of the old token count, about 27 percent of the Opus 5 cost, a reduction near 73 percent ^[17]. The price cut accounts for 20 points of that, and the token ratio for the other 53 ^[18].

The write-up does not split Box's token count into input and output. If most of the saving sits on the output side, a job that reads a large corpus and writes a short answer sees less of it. Bhavnani said the improvement should "matter a lot for teams running agents across their content in areas like financial services and the public sector" ^[14].

The other named customer is talking about how the output reads. John Ruelas, a staff software engineer at Ramp, said "Verbose, hard-to-follow output has been my biggest frustration with frontier models, and Claude Opus 5.5 fixes it" ^[11]. He said the model "writes like a good colleague and follows our writing rules" ^[12]. Anthropic separately says Opus 5.5 generates output more than 30% faster than Opus 5 ^[6].

The efficiency stops at the safeguard boundary. When safeguards fire, requests fall back from Opus 5.5 to Opus 4.8, and ZDNET reports that in practice most cybersecurity tasks get rerouted to Opus 4.8, with biology and frontier LLM development requests sent to Opus 5 ^[7]. So a triage pipeline cannot be priced at one blended token rate. Organisations approved through Anthropic's Cyber Verification Program can start using Opus 5.5 in a few weeks, and ZDNET's writer disclosed that he has been approved into that program himself ^[15][16].

For the $20 a month tier the two changes compound. Anthropic raised the five-hour usage limits by 20 percent ^[4] and says both five-hour and weekly limits go further because Opus 5.5 costs less than Opus 5 ^[5]. ZDNET pairs a 20 percent larger bucket with a 25 percent slower burn and puts effective run capacity about 50 percent higher ^[10]; 1.2 times 1.25 is 1.5 ^[19]. That only holds if the limit is metered in something that tracks the token price.

Running the three jobs that account for most of your token spend on both models, counting input and output tokens separately, costs an afternoon. Two numbers come out of that comparison and only one is safe to put in a budget: 20 percent off per token is published pricing, and a third of the tokens is Box's result on Box's content.