The day single-vendor AI lock-in stopped being defensible

Three numbers tell you what happened today. Gemini 3.1 Pro Preview scored 57.2 on the Artificial Analysis Intelligence Index. GPT-5.4 Pro scored 57.0. The benchmark cost gap was $892 versus $2,950 — and GPT-5.4 burns roughly twice the tokens to get there, so the production gap is closer to 6x than 3x. Open-weights GLM-5 lands at 88% of frontier quality for $547.

That is the entire story of the AI market in March 2026 compressed into a single benchmark line.

The corroborating events landed in the same week. Meta — which spent $14.3B on Scale AI, hired Alexandr Wang as CAIO, and stood up a 100-person lab to ship Avocado — is reportedly in internal discussions to license Gemini because its own model can't beat Gemini 3.0. Microsoft, two years and $40B+ deep into its OpenAI bet, is now bundling Anthropic into Copilot. OpenAI walked away from expanding its Abilene Stargate site from 1.2GW to 2GW after demand-forecasting fights with Oracle, then shipped GPT-5.4 two days after GPT-5.3 with no release notes. Companies don't move like that when they're winning.

The takeaway isn't that OpenAI is in trouble. GPT-5.4 still wins coding (SWE-Bench-Pro, 75% on OSWorld-Verified — above the 72.4% human baseline) and agentic benchmarks. The takeaway is narrower and more useful: raw intelligence has commoditized; cost-per-task has not. If your architecture routes everything to one provider, you are paying a premium tax on every call that didn't need GPT-5.4 to begin with. That's a margin problem masquerading as an engineering preference.

Routing is now the load-bearing decision

Practitioners have already converged on the answer, and it's not subtle. GPT-5.4 XHigh for production code. Opus 4.6 for design and planning. Gemini 3.1 for general reasoning and anything multimodal. GLM-5 self-hosted for high-throughput batch where 88% quality is fine. Droid and Pi CLIs already support mid-conversation model switching. Adobe shipped the enterprise version of this thesis: Firefly now aggregates 25+ third-party models from Google, OpenAI, Runway, and Black Forest Labs, and made unlimited generations the price-point default for paid tiers.

The trap to avoid: don't build a model-agnostic abstraction that hides provider differences. That was the LangChain bet, and Microsoft just spent two years proving it loses. You want orchestration, not abstraction — a router that dispatches by task type with provider-specific adapters underneath, so you keep extended thinking, structured outputs, and prompt caching where they matter. GPT-5.4's $0.25/M cached token price is real money on the table, but only if your prompts are designed to hit the cache.

The quieter story: agents shipped to production without identity

While the headlines fixated on benchmarks, three smaller signals stacked up. Teleport launched an Agentic Identity Framework — cryptographic per-agent identity, scoped permissions, audit trails. Vercel's Skills.sh is becoming the agent app store, with 66 Claude skills and zero systematic vetting for prompt injection. Practitioners are documenting that long agent sessions silently lose state to context compaction and need an external progress.md write-ahead log just to survive a multi-step task.

And Operation Lightning took down SocksEscort — a 17-year-old residential proxy botnet — but the AVRecon malware on 369,000 home routers doesn't self-clean when the C2 dies. A quarter of those routers are in the United States, sitting under your remote workforce's VPN tunnels.

The through-line across both stories — model commoditization and agent infrastructure hardening — is the same. The defensible layer of an AI product is no longer the model. It's the orchestration, the identity boundary, the workflow lock-in, and the data you bring. GPT-5.4 reverse-engineered T3 Code's entire interface before it open-sourced. If your moat is "better prompts and a nicer UI," an LLM can replicate it from a thirty-minute demo.

Meanwhile Block cut 40% of its staff — 4,000 people — not as a cyclical trim but as an explicit thesis that AI changes the headcount-to-output ratio for a fintech. Atlassian cut 10% the same week. Whatever you think of the call, your next headcount request gets read against this backdrop.

What to do this week

One thing, specifically. Pick your single highest-volume AI feature — the one burning the most tokens in production right now — and stand up a parallel evaluation against Gemini 3.1 Pro Preview on your real workload. Not a benchmark. Your actual prompts, your actual quality bar, your actual latency budget. Two weeks, real traffic shadow-mode if you can swing it, then a decision: does the 3x cost premium hold up against your data, or does it not?

If it doesn't, you have an immediate margin recovery and a concrete artifact to show your board. If it does, you have a defensible reason to keep paying GPT-5.4 prices that isn't "we already integrated."

The vendors who lost this week were the ones whose customers couldn't answer that question with their own numbers. Don't be that customer.