Depth vs. Speed: What This Week's AI Drops Tell You About the Next Two Years

Two major AI releases dropped this week. Both from companies operating at a scale most engineers will never touch. Both trying to solve what they believe is the main problem with AI right now. They landed on completely different answers.

Google shipped Gemini 3 Deep Think — an upgraded reasoning mode that scored 84.6% on ARC-AGI-2, hit gold-medal level on the International Math Olympiad 2025, and helped a Rutgers mathematician catch a logic flaw in a peer-reviewed paper that human reviewers had missed. Google's bet: the bottleneck is intelligence. Models need to think harder.

OpenAI shipped GPT-5.3-Codex-Spark — a smaller Codex model running on Cerebras Wafer Scale Engine 3 hardware, hitting 1,000+ tokens per second, with 80% reduced per-client roundtrip overhead and 50% faster time-to-first-token. OpenAI's bet: the bottleneck is latency. Models are already smart enough. They just need to stop making you wait.

Both can't be right. Or maybe the question is wrong.

The depth bet

Deep Think's numbers are real. ARC-AGI-2 was specifically designed to resist pattern-matching from training data — it tests genuine reasoning, not memorization. 84.6% on that benchmark means something. The 48.4% on Humanity's Last Exam (expert-level questions across dozens of domains, no tools) is another step up.

But the benchmark that stuck with me wasn't either of those. It was the Wang Lab at Duke using Deep Think to design fabrication recipes for growing semiconductor thin films over 100 µm — a precision target that previous methods couldn't reliably hit. That's not a "AI wrote my boilerplate" story. That's an AI solving a materials science problem.

And the Rutgers mathematician finding the logic flaw in a peer-reviewed paper? That's operating in territory that was off-limits two years ago. Human experts missed it. The model didn't. Speed had nothing to do with it.

The speed bet

Codex-Spark's headline number is 1,000+ tokens per second. That gets attention. But the number I care about is the 80% reduction in per-client/server roundtrip overhead.

That's infrastructure, not model capability. They rewrote their inference stack, moved to persistent WebSocket connections, reworked session initialization so the first token arrives faster. These are the kinds of wins that change how people use a product, not just how fast it is. When latency drops far enough, you stop batching your thoughts into discrete queries and start thinking alongside the model in real time. The cognitive overhead of "compose query, wait, read, compose again" disappears.

The Cerebras partnership is architecturally interesting too. GPUs are cheaper per token at scale — they stay. Cerebras wafer-scale chips are faster at inference but not cost-competitive for bulk token generation. So you run a latency-first serving tier on specialized hardware for interactive work, and standard GPU infrastructure for everything else. Watch this pattern. It'll spread.

Why both bets can coexist

The use cases genuinely don't overlap much.

Gemini 3 Deep Think is for researchers and scientists — people who hand a model a hard, messy problem and wait for a thorough answer. Sometimes they wait overnight. Latency is irrelevant when you're working on semiconductor fabrication parameters. Accuracy and depth of reasoning are everything.

Codex-Spark is for developers mid-session — asking a question, seeing an answer, reacting, asking again, four times in two minutes. For that workflow, 500ms vs 50ms time-to-first-token is the difference between the tool feeling like a collaborator and a vending machine.

These are different products. They're not really competing.

What the divergence signals

A few things I'm taking from this week.

The general-purpose model era is fragmenting. A single model that's fine for everything is increasingly losing to specialized models that are excellent for something specific. The routing layer — the thing that decides which model handles which request — becomes the real engineering problem. We're not there yet. But the pressure is building.

The infrastructure layer is catching up. OpenAI spent meaningful engineering time this week not training a better model, but cutting WebSocket overhead. That happens when models are good enough that the bottleneck shifts downstream. This will keep happening. More of the interesting work in AI over the next two years will look like systems engineering, not ML research.

And the depth-vs-speed split suggests that different stages of a workflow will need different approaches. Codex-Spark handles the interactive back-and-forth. A reasoning model handles the hard problem you kick off at end of day and review in the morning. The architecture that combines both — routing within a single session based on query type — is what I'd be building if I were working on an AI coding product right now.

The bottleneck in AI right now isn't models. It's the plumbing around them.

---

About the Author: Muhammad Khan is a Principal Full Stack Engineer with 9+ years of experience building scalable systems for millions of users.