AI Platform Divergence: What It Means for Builders

A tweet crossed my feed a while back that stuck. Paraphrased: OpenAI is specializing in AI agents (Deep Research, Agents SDK). Anthropic in automated program creation and coding assistants. Google in multimodal — image understanding and image generation. Three labs, three product bets.

I wanted to see how much of that was vibes and how much was actually in the releases. So I dug in. This post is the research: what the divergence is, what the numbers and products say, and what it means if you're building on top of these platforms — startup, incumbent, or in between.

What "divergence" actually means

Nobody has an exclusive lock on "agents" or "coding" or "multimodal." All three ship long context, reasoning modes, and some form of tool use. The divergence is where each lab is putting their best work and product — and where they're winning mindshare and usage.

Think of it as specialization, not capability walls. You can still use Claude for research or GPT for code — I've run research pipelines that mix providers, so the lanes are real but not walls. The leading edge and the marketing are aligning around three lanes.

The evidence: three lanes

OpenAI → Agents and research

OpenAI is pushing agentic workflows — multi-step planning, tool orchestration, research that runs on its own.

• Deep Research API and Agents SDK (Python and TypeScript). You build research pipelines that plan, search, synthesize, and stream progress. Not just "answer this" — "run this research job and give me a report."
• The SDK supports web search, MCP, file search, multi-agent patterns. The recommended model for deep research is o4-mini-deep-research-2025-06-26 — faster than the full o3 research model, built for planning and tool use.
• Use cases in the docs: financial research agents, customer service systems, multi-step reasoning. The positioning is clear: agents for work that needs planning, synthesis, and tools. Not for one-shot Q&A. If you're building something that runs research jobs (not single prompts), that's the lane — the same kind of multi-step orchestration I used on top of Exa.

So: agents and research are the product story. That's where the API and the narrative point.

Anthropic → Coding and program creation

Anthropic is doubling down on coding — the AI pair programmer that lives in your editor and terminal.

• Claude Code is the flagship: terminal and IDE integrations (VS Code, JetBrains, Cursor, Windsurf), codebase-aware, natural-language commands for writing, editing, debugging, testing, and git. Built on Claude Opus 4 and Sonnet 4.5.
• Coding benchmarks: Claude Opus 4.5 is advertised as top of the pack on SWE-bench (e.g. 80.9%). The messaging is "best coding model available."
• Features like Computer Use, MCP, and the Effort parameter (high / medium / low — trades response depth for token cost) are aimed at engineers who want control — deterministic behavior, clear tool boundaries. Anthropic's own numbers: 60x faster code review feedback for one AI platform customer, 95% reduction in time to run tests for an enterprise. That's a productivity story, but the identity is "reliable engineer."

So: coding and program creation are the product story. Enterprise trust + developer tooling.

Google → Multimodal + Integrated Reasoning

Google is leaning into multimodal from the ground up while aggressively baking inference-time reasoning into their baseline.

• Gemini 3.1 Pro just dropped as their first-ever ".1" point release. It takes the deep reasoning from their specialized "Deep Think" models and integrates it directly into the workhorse Pro model via adjustable thinking_level parameters (Minimal, Low, Medium, High). It's essentially "Deep Think Mini" on demand.
• Multimodal DNA: Gemini 2.5 and 3 series remain natively multimodal. Long context (up to ~1M tokens), native image understanding, and native image generation — Gemini 2.5 Flash Image and Gemini 3 Pro Image (preview). Generate images with reasoning, interleaved text-and-image output (e.g. blog posts with images), up to 1024px and 4096px depending on model.
• Veo (video) and strong performance on video and audio. The story isn't "we added vision" — it's "the model was designed for text, image, audio, video from day one."
• Image understanding: captioning, classification, visual QA, object detection, segmentation. One model, many modalities. No separate vision stack.

So: multimodal understanding and generation are the product story. "See and create" is the edge.

The strategy layer: not just product, but position

Product specialization sits on top of a go-to-market split. Dani's newsletter tracked 195+ product updates across the three from July–December 2025 and summarized it cleanly:

	OpenAI	Anthropic	Google
Positioning	"Everything platform" — consumer hub, Atlas, broad reach	"Enterprise trust" — safety, compliance, reliability	"Ecosystem integration" — default inside Search, Workspace, Android
Market shift	Enterprise share down (on the order of 50% → ~25%)	Enterprise share up (~12% → ~32%)	Distribution: 2B+ users, Gemini in 7 products same day

So you get two axes: (1) strength — agents vs coding vs multimodal, and (2) strategy — consumer platform vs enterprise trust vs ecosystem default. They reinforce each other. Anthropic's "we're the safe, controllable one" pairs with "we're the coding one." OpenAI's "we're the agentic one" pairs with "we're the big tent."

Context windows, reasoning modes, and agentic behavior have become table stakes. The real differentiation is who makes them economically viable at scale and who owns which use case in the mind of buyers. As Dani's analysis notes, long context burns memory and compute; offering it at scale creates margin pressure unless you own the stack (e.g. Google with TPUs).

Where things stand: convergence + divergence

• Convergence: Everyone has long context, tool use, and now adjustable inference-time reasoning. OpenAI pioneered this with the o-series reasoning_effort, Anthropic adopted it as a hybrid toggle in Claude 3.7 (thinking), and Google just made it a sliding baseline with Gemini 3.1 Pro's thinking_level. Giving the model more compute to "think" before speaking is no longer a separate product lane—it is absolute table stakes. No single lab has a unique capability; the basics are shared.
• Divergence: Best-in-class and investment are splitting. Agents → OpenAI. Coding → Anthropic. Multimodal (vision, image gen, video) → Google. Plus who you trust (enterprise vs consumer) and where you already are (Google stack vs not).
• Enterprise: Anthropic is the one gaining enterprise share in the reported data. OpenAI still has the "smartest model" brand but is more consumer/platform. Google is "everywhere" but often as default, not necessarily as the premium enterprise contract.
• Builders: The practical take from Menlo's 2025 LLM update and similar: choose by use case. Consider multi-model — reasoning from one provider, coding from another, vision from a third. Assume models and leaderboards churn; don't over-optimize for today's "best" model. Open-weight models (Llama, etc.) are advancing and matter for cost-sensitive deployment, but closed-source frontier models still lead in enterprise usage for now.

So the "current state" is: one layer of shared tech, another layer of specialization and strategy. No single winner. Three lanes.

What to do if you're in between

If you're building product:

• Agentic / research workflows → OpenAI's Agents SDK and Deep Research are the obvious first place to look. That's where the APIs and docs are pointed.
• Code gen, codebase tools, deterministic control → Anthropic (Claude Code, API with control knobs). That's where the coding narrative and benchmarks live.
• Vision, image understanding, image or video generation → Google (Gemini image models, Veo). That's where multimodal is the lead story.

You're not locked to one provider. You're choosing by task — and increasingly, combining them. Example: a research-heavy product might use OpenAI for the agent loop and Claude for code gen in the editor; pick by task, not by brand.

If you're a startup: The divergence is optionality. You can position as "best agentic research tool" (OpenAI stack), "best dev/code platform" (Anthropic stack), or "best visual/multimodal product" (Google stack). Or you mix. The risk is betting everything on one vendor's roadmap when each is specializing in a different direction.

If you're an incumbent (consultancy, product company, enterprise): You're in the middle of picking primary partners and filling gaps with the others. The "current state" is: no single winner; three lanes with different strengths; choose by use case and plan for multi-model where it matters.

Bottom line

The tweet was right in spirit. OpenAI is specializing in agents. Anthropic in coding. Google in multimodal and integrated reasoning. The evidence is in the products and the release patterns (like Google's shift to .1 point releases). Underneath that, strategy is diverging too: platform vs enterprise trust vs ecosystem.

For builders: treat the divergence as specialization. Pick the right tool for the job. Plan for multiple tools. Don't assume one lab will own every use case — and don't assume the current leaderboard will look the same in six months. The useful move is to understand the three lanes and build with that in mind, not to wait for a single winner.