As of July 2026, two stories landed within 48 hours of each other that matter more together than either does alone. Moonshot AI shipped a 2.8-trillion-parameter model that needs dedicated GPU infrastructure just to serve a single request. And a live benchmark run showed Anthropic's newest coding model, Opus 5, passing only 24% of a real, multi-step coding test — the best score of the three Claude models tested, and still a failing grade by any normal standard. Different stories, same lesson for anyone deciding whether to build an AI agent in-house or hire it out: the models keep getting bigger and smarter, and that is not the same thing as safe to trust lights-off.
The infrastructure story: a 2.8-trillion-parameter model needs its own power plant
On July 28, 2026, Moonshot AI released Kimi K3, a 2.8-trillion-parameter model — the first open-weight model to reach the 3-trillion-parameter class. It went live the same day on Telnyx Inference at $2.70 per million input tokens and $13.50 per million output tokens, with a 1-million-token context window. Telnyx was specific about why this took a dedicated release: the model runs on GPU infrastructure the company "owns and operates," because a model this size cannot be served off shared or commodity hardware.
That is the part every "just build it yourself" pitch leaves out. The weights might be open and free to download, but running a model at this scale for real customers means provisioning dedicated GPU capacity and eating a hosting bill that scales before a single lead ever talks to your chatbot. Model size keeps climbing. Infrastructure cost climbs with it.
The reliability story: even the best model still fails 76% of the time
The second story is more useful for a different reason: it shows what happens after you've solved the infrastructure problem and actually put an agent to work. A developer at Humanlayer ran three Claude models — Opus 4.8, Sonnet 5, and Opus 5 — through SlopCodeBench, a benchmark that doesn't hand a model the whole problem up front. Instead it drips out new requirements in checkpoints, the way real client work actually unfolds, and a model fails a checkpoint the moment any earlier defect breaks a later requirement.
Opus 5 won the comparison and still only cleared 4 of 17 checkpoints — a 24% strict pass rate. Opus 4.8 and Sonnet 5 each cleared just 1 of 17. None of the three models finished a single one of the three test problems, including the one graded "easy," without leaving at least one defect behind.
The quality numbers behind that score are the part worth sitting with. Opus 5 wrote 29,065 lines of code across the test set against roughly 9,000 for each of the other models — five times as many functions. And 93% of the lines Opus 5 wrote tripped at least one of the benchmark's code-quality rules, versus 89% for Sonnet 5 and 98% for Opus 4.8. When the same researcher pointed a smaller rule set at a real, human-reviewed TypeScript codebase for comparison, Opus 5's unattended output came back with 11.6 times more quality-rule triggers per thousand lines of code than the reviewed codebase.
Read plainly: more code, more functions, more test volume, and a benchmark score that still means three out of every four checkpoints failed. The model got smarter. The failure rate barely moved.
What this means for an agent you're running unattended
Put the two stories side by side and the pattern is obvious. Frontier models are getting large enough that even the API providers have to build dedicated hosting just to answer questions with them, and the newest, most capable coding model on the market still fails the majority of a realistic, multi-step build when nobody is watching. If your booking agent, your outreach sequencer, or your customer-facing chatbot was assembled by pointing a coding agent at a spec and walking away, the SlopCodeBench numbers say that setup accumulates defects the same way — quietly, checkpoint by checkpoint, until something a customer touches breaks. We wrote about this exact failure mode in our piece on AI agents failing silently: the problem is never that agents make mistakes. It's that most builds have no way to catch the mistake before a customer does.
None of this means don't use agents. It means don't run them without a verification layer — tracing, checkpoints, a human review step — the same way SlopCodeBench catches Opus 5's defects instead of letting them slide. That layer is expensive to build and easy to skip when you're moving fast, which is exactly why it gets skipped.
What this means if you're weighing AI marketing or an agent build
If you're deciding whether to build your own agent stack this year, budget for both halves of this problem: the infrastructure to run a frontier-scale model, and the verification layer to catch what it gets wrong before a customer does. If you'd rather skip both bills, find out first whether your business even shows up when someone asks an AI for the best option near them — get the free free AI Visibility Report and see where you stand.