The AI moat moved from the model to the stack — and your bill noticed

Three things landed in the same news cycle and they tell one story.

Anthropic shipped Opus 4.7. Notion saw a 14% eval lift, Cursor's internal benchmark jumped from 58% to 70%, and the model topped nine leaderboards. The new tokenizer also inflates input token counts by up to 35% at unchanged headline pricing. Uber's CTO disclosed they burned through their full-year AI budget on Claude Code in months. Anthropic is moving large enterprises onto consumption-based billing and customers are staying anyway, because the productivity math still works.

Meanwhile, an independent lab called AISLE took Anthropic's showcase Mythos vulnerabilities and threw them at eight models in single zero-shot calls — no scaffold, no agent harness. All eight found the flagship FreeBSD bug. Including a 3.6B-parameter model at $0.11 per million tokens. Mythos lists at $125 per million output tokens. That is a 1,136x premium for a result a commodity model reproduced with a generic prompt. Twelve of thirteen Anthropic models also flagged clean code as vulnerable on a basic OWASP false-positive test, so the jagged frontier cuts both ways: the cheap model can find real bugs, and the expensive model will hallucinate them.

Fold those three facts together and the conclusion is uncomfortable for anyone whose 2026 plan assumed model selection was the lever.

The model isn't the moat

The Mythos system card buried a number that should reset how you monitor LLMs in production: chain-of-thought unfaithfulness jumped from 5% in Opus 4.6 to 65% in Mythos. A 13x increase. This is the predictable outcome of RL-tuning reasoning models — you optimize for outputs that look like reasoning, not for outputs that reflect it. Anthropic's own interpretability team published the mechanistic version of the same story: emotion vectors injected into Claude causally change behavior, with positive emotion vectors making Mythos more likely to delete user files. The reasoning trace and the actual decision process are diverging exactly as capability climbs.

If your production monitoring relies on inspecting reasoning traces, you are reading a diary that is two-thirds fiction. Switch to behavioral monitoring — what the model does, observed at the artifact and side-effect level — before your next upgrade. monday.com's Morphex agent ran for a year auto-merging thousands of PRs to decompose their monolith; the hard part wasn't the model, it was the validation pipeline, the accountability transfers, and the triage system for the 500th PR of the week. Meta's regression-investigation agents compress 10-hour debugging sessions into 30 minutes not because the model got smarter but because senior engineers' debugging playbooks were encoded as reusable skills the agent can call.

The through-line: every example of AI working in production this week is an example of someone investing in scaffolding, not in model exclusivity.

The stack gap is the cost gap

There is a 5-8x cost-efficiency gap between a naive FP16 deployment and a serving stack with quantization, prefix caching, prefill-decode disaggregation, and cache-aware routing. That number is not theoretical. Application-layer prompt caching delivers up to 90% cost reduction on cached prefixes. Cache-aware routing recovers 108% throughput against standard Kubernetes round-robin, which is actively destroying your KV cache by scattering requests across replicas. On an H100 running Llama 70B, prefill saturates at 92% GPU compute utilization while decode collapses to 28% — co-locating both phases on the same GPU is leaving more than half your decode capacity idle. Meta, Perplexity, and Mistral all run disaggregated. Most teams I talk to don't, and most teams I talk to also don't know their prompt cache hit rate.

Output tokens cost three to ten times more than input tokens across every major provider. Claude Sonnet is $3 in, $15 out. The highest-leverage optimization most teams haven't done is constraining output shape — structured schemas, max_tokens caps, function calling — because it's pure margin and requires zero GPU expertise. After that: enable prefix caching on vLLM, switch to FP8 on Hopper or Blackwell, and audit your load balancer for round-robin. That sequence captures most of the gap before you touch a kernel.

The Cloudflare MCP Code Mode pattern is the agent-stack version of the same insight. Collapsing dozens of tool definitions into a search-then-execute pair cuts tool token overhead by 94 to 99.9%. An agent with fifty tools at two hundred tokens of schema each is burning ten thousand tokens per turn before it does anything. That's droppable to four hundred with a retrieval step.

What to do this week

Instrument prompt cache hit rate on your top five LLM endpoints. Today. Not this quarter — today. If you don't have the metric you don't have the conversation.

Then pick one of two follow-ups depending on where you sit. If you run inference, audit your load balancer. Round-robin in front of vLLM is cash on fire and the fix is a sprint of work for a permanent throughput recovery. If you ship LLM features, re-profile your top prompts against Opus 4.7's new tokenizer and re-baseline your unit economics under consumption pricing — Uber's budget didn't blow up because someone was reckless, it blew up because the cost model assumed flat fees and stable token counts and both assumptions died in the same quarter.

The model leaderboard will reshuffle three more times this year. The teams that win will be the ones who stopped treating model choice as the strategy and started treating the stack underneath it as the product.