Claude Opus 5.5 went live inside Perplexity Computer on Tuesday as the platform’s new Standard effort tier, with Perplexity’s own account posting benchmark numbers within minutes of the switch.
Key Takeaways
Claude Opus 5.5 became Perplexity Computer’s new Standard effort tier on Tuesday
The model scored 0.610 on Perplexity’s WANDR benchmark at a cost of $4.13 per task
Arav Srinivas said Opus 5.5 would become the default option for all Perplexity Computer users
Perplexity Computer is a browser-based AI agent product that performs multi-step tasks including research and purchases
The model scored 0.610 on Perplexity’s internal WANDR benchmark, a wide-and-deep research evaluation, at a cost of $4.13 per task, posting that the result “slightly” beat rival model Fable 5.1. Perplexity Computer is the company’s browser-based AI agent product that carries out multi-step tasks like research and purchases on a user’s behalf, rather than just answering chat queries.
Perplexity CEO Arav Srinivas followed with his own post confirming the rollout to all Perplexity Computer users, framing the win around cost efficiency rather than raw score.
He said Opus 5.5 “compares favorably” to Fable 5.1 “at a fraction of the cost” and would become the default option going forward.
A third account, AI-tracking outlet TestingCatalog, independently flagged the release as a new state-of-the-art result in agentic coding and computer-use tasks, describing it as an “exponential slowdown” moment, a reference to the narrowing gains between successive frontier releases.
Why Cost Per Task Is Becoming The New Scoreboard
Model releases used to compete purely on benchmark score. Claude Opus 5.5’s launch marks a shift toward cost-normalized comparisons, where a model’s dollar-per-task figure matters as much as raw accuracy for companies running agents at scale.
Perplexity’s WANDR benchmark measures how well a model handles open-ended research tasks requiring multiple search and reasoning steps, not single-shot question answering.
A $4.13 per-task cost only matters at volume, and that is exactly the scale enterprise AI agent deployments are starting to hit.
From Chat Assistant To Paid Agent Platform
Anthropic built its reputation on the Claude chatbot before pushing into enterprise agent workflows through 2026, including security integrations with firms like Palo Alto Networks‘s Unit 42 division, which began pairing Claude models with continuous threat-defense tooling.
The Opus 5.5 release extends that push into a third-party consumer product rather than Anthropic’s own interface, a distribution move that mirrors how OpenAI has placed GPT-6 Astra inside partner tools such as Higgsfield AI’s video generator.
Also Read: Anthropic’s Claude for Financial Advisors Gets First Real Test With Schwab, BlackRock Data
What Happens When Frontier Gains Start Shrinking
TestingCatalog’s framing points to a live debate among AI labs over whether successive model generations deliver smaller jumps even as costs fall faster. If Opus 5.5’s edge over Fable 5.1 proves marginal on score but decisive on price, buyers may start choosing models on cost curves rather than leaderboard rank.
That would change how labs market releases, and how fast rivals respond with their own price cuts.
Read Next: Corporate AI Push Gets Official Anthropic Deal at Societe Generale