Frontier AI just consolidated to three labs, and your stack is priced for the old world

Two numbers from independent benchmarks this week: Gemini 3.1 Pro Preview scored 57.2 on the Artificial Analysis Intelligence Index. GPT-5.4 Pro scored 57.0. The cost delta on the same benchmark suite was $892 versus $2,950 — a 3.3x gap for effectively identical general intelligence, and closer to 6-7x once you account for GPT-5.4 burning roughly twice the tokens per equivalent output. Open-weights GLM-5 came in at 88% of frontier quality for 18.5% of the price.

That's the cover story, and it's not really about benchmarks. It's about what happened around it in the same 72 hours.

Meta — after spending $14.3B on Scale AI, poaching Alexandr Wang as Chief AI Officer, and standing up a 100-person TBD Lab — is reportedly in internal discussions about licensing Gemini to power its own products because Avocado couldn't clear Gemini 3.0. Microsoft, on its third AI pivot in eighteen months, is bundling Anthropic into Copilot rather than continuing to build the wrapper layer itself. OpenAI walked away from expanding its Abilene Stargate site from 1.2GW to 2GW, citing demand-forecasting disputes with Oracle. The largest AI compute consumer on the planet can't confidently project its own compute curve.

Yes, but — GPT-5.4 still leads on the two things a lot of production workloads actually need: coding (SWE-Bench-Pro, Terminal-Bench-Hard) and agentic tool use (75% on OSWorld-Verified, above the 72.4% human baseline). If your product is a coding agent or an autonomous computer-use system, the premium is defensible for now. For everything else, you're paying 3x for a tie.

The single-vendor bet is the fragile bet

The repricing here isn't just at the API line item. Adobe now routes 25+ third-party models through Firefly and just set unlimited generations as the paid-tier standard. Practitioners running production agents already switch mid-conversation — GPT-5.4 XHigh for code, Opus 4.6 for design and planning, Gemini for general reasoning, GLM-5 for high-throughput batch. The pattern that wins is orchestration with provider-specific adapters, not a universal abstraction that flattens each model's advantages into a lowest-common-denominator interface. Microsoft learned that at $40B scale so you don't have to.

If more than 70% of your inference spend sits with one provider, that's not a technical decision anymore. It's a fiduciary one, and Gemini plus GLM-5 just handed you the leverage to renegotiate.

The infrastructure story cuts both ways

The demand crack at Abilene sits inside a stranger picture. $19B in VC megafunds closed in a single week — Founders Fund $6B, General Catalyst $10B, Spark $3B — while OpenAI faces a skeptical IPO window and Nvidia-backed Nscale is racing to acquire permitted, power-secured physical sites before it goes public. 46 off-grid gas-fired plants now account for 30% of planned US data center capacity, 90% announced in 2025 alone. Electricity prices rose at 2x inflation last year. Iran's declared closure of the Strait of Hormuz puts roughly $300B in Gulf AI capex in play.

Capital is abundant. Conviction is fracturing. The scarce asset stopped being GPUs or money — it's permitted sites with secured power. Any strategic plan built on "compute keeps getting cheaper and more abundant" needs to be stress-tested against a scenario where hyperscaler expansion decelerates 20-30% and inference costs rise 20-40% over eighteen months. Model switching is how you preserve optionality when that scenario shows up.

The blast radius nobody's staffing for

While the model layer consolidates, the agent layer is shipping to production with the security posture of npm circa 2013. Vercel's Skills.sh registry is filling up fast — Claude alone lists 66 skills and 9 workflows — with no signature verification, no sandboxing, no permission scoping. AGENTS.md and CLAUDE.md files auto-load into agent context at every session start and execute with the agent's full permissions. Teleport just launched an Agentic Identity Framework because "ungoverned production agents" is now a monetizable category.

GPT-5.4 achieves 75% on OSWorld-Verified computer-use at $2.50 per million input tokens. Superhuman offensive capability is a credit card away. Meanwhile Operation Lightning took down SocksEscort, but the AVRecon malware on 369,000 residential routers — over a quarter of them in the US — persists until each device is manually rebooted or patched. Those routers sit between your remote workforce and your VPN concentrator. And CISA is running on shutdown fumes into the fifth week while Block and Atlassian cut a combined 5,600 people, many of whom held the credentials and institutional knowledge that touched your data.

What to do this week

One thing, concrete: stand up a task-aware model routing layer and run a 30-day parallel evaluation across GPT-5.4 Pro, Gemini 3.1 Pro Preview, and one open-weights option (GLM-5 or equivalent) on your actual production workloads, not benchmarks. Map your top ten AI features to task categories — general reasoning, code, design, high-volume batch — and assign a primary and fallback model for each. Instrument cost-per-successful-task, not cost-per-token. Bring the delta to your board before the quarter closes.

The 3.3x cost gap on general reasoning is the highest-leverage optimization sitting in your infrastructure right now. Every week you don't route around it is margin you're not getting back.