The current Artificial Analysis Intelligence Index shows a tight cluster at the top, with Claude Opus 5.5 (Max) leading at 58, followed by Claude Sonnet 5.5 (Max), Claude Opus 5.5 (Xhigh), and Muse Wonder Full (Max) all sitting at 56. Out of 127 models evaluated, the top four span exactly two points. For enterprise buyers, this headline composite opens a procurement gap. The 56 in the headlines is the subscription-tier model. The API tier enterprise teams can actually call scores 52, at $4 per million input tokens and $20 per million output. And the composite score buyers are reading is one the index itself says may not govern their workload.
This two-point compression leaves pricing and ecosystem as the primary differentiators, as that analysis mapped. The Muse Wonder launch showed how access tiers complicate procurement: Muse Wonder Full, whose entry timing on the index is unstamped, is gated behind $50-per-month Power or $200-per-month Maximum tiers in Muse Code and is not available on OpenRouter at all. The open-weights landscape sits somewhere else entirely: the highest-ranked open model, MiMo-V2.6-Pro, scores 46, and fifty-nine of the 127 evaluated models are open-weights. The 12-point gap between the closed leaders and the best open model puts the mid-tier squarely in closed-model territory.
The index is a composite run by a single analyzer rather than an independent audit, and its own methodology note says figures represent a model’s first-party API or the median across providers where no first-party API exists. Its relevance note is the one buyers skip: while model intelligence generally translates across use cases, specific evaluations may be more relevant for some use cases. The headline composite, v4.3.2, incorporates 10 evaluations. The same index publishes the workload-specific cuts: AA-Briefcase v1.1 for agentic knowledge work, GDPval-AA v2.1 for agentic real-world work tasks, AutomationBench-AA for agentic SaaS workflows, EnterpriseOps-Gym-AA for agentic business operations. A buyer reading only the headline number is reading past all of them.
What is actually being bought shows up in traffic. On OpenRouter’s listing, the apps sending the most token volume to Muse Wonder are agentic AI tools: Plato Computer Use Agent (11 billion tokens), omp (1.62 billion), Enclave AI Hacking Arena (660 million), Claude Code (615 million), and Kilo Code (471 million). That is bounded production work, computer use and code iteration, already running on the mid-tier. The score certifies little about how that work fails, or about who catches it when it does.
The 150-deployment audit from The Brief Script puts numbers on that gap. Its line: \”technical action often arrives before mature business measurement.\” Of 150 qualifying enterprise deployments, 118 keep explicit human loops for approval, exception handling, governance, or active review, and only 22 disclose any cost or financial impact. When a mid-tier model runs a workload, its margin of error does not disappear; it shifts to the staff managing the output. The review-tax research covered in our Deploying Blind piece shows what that costs when anyone measures it: coding agents raised code volume 30 percent while issues resolved stayed flat, review times jumped 49 percent, revision requests nearly doubled, and 14 percent more workers were pulled into code review.
More than 40 percent of agentic AI projects will be canceled by the end of 2027 on Gartner’s forecast, citing escalating costs, unclear business value, and inadequate risk controls. The belief gap is wider still: 93 percent of leaders believe AI will eventually run core business functions; 7 percent have made it work. The mid-tier trap is what sits between those numbers. A composite score gets treated as a certification, and the cost of managing the model’s output gets treated as somebody else’s problem.
For an enterprise buyer, the useful questions are operational rather than comparative: What does the agent read? Which tools can it choose? What can it write or change? Which actions require approval? What happens when confidence is low or a case falls outside policy? Those five come from the audit’s own checklist, and not one of them is answered by a leaderboard. A score of 52 or 56 describes a model on ten evaluations. The work decides what happens next, and the people in the exception lane carry the margin of error.