Claude 3 and the New Frontier of Enterprise AI

Last month, Anthropic released Claude 3. Most people see three models and think "okay, they shipped three versions of the same thing." They're wrong. What Anthropic actually shipped is a model family that fundamentally changes what's possible for enterprise AI.

I've been shipping enterprise software for five years. I've built systems on top of GPT-4, GPT-3.5, and half a dozen smaller models. I've watched what works and what breaks. When I got access to Claude 3 Opus in early March, I spent a week just thinking about what changed.

Here's what actually changed: the cost-capability frontier moved.

The Old Tradeoff

Until last month, enterprise AI had the same constraint it's always had: you pick your tradeoff.

Use GPT-4? Costs $0.03 per 1,000 input tokens. Excellent reasoning. Stable output. But at scale, it gets expensive fast. If you're building a system that processes 1 million tokens a day for customers, your LLM bill is $900 a month just for inference. That's before you scale.

Use GPT-3.5 or Llama? Cheap. $0.0005 per 1,000 tokens for GPT-3.5. But the reasoning degrades. You get more hallucinations. You need more guardrails. You end up building workarounds that cost more in engineering time than the LLM savings.

Most enterprises just pick a layer—expensive but smart, or cheap but fragile. Then they hope the chosen layer doesn't shift.

Claude 3 breaks that tradeoff.

The Three Models That Matter

Opus is positioned as the flagship. It's GPT-4 class reasoning. Handles complex reasoning, nuanced context, long documents. Costs more than GPT-3.5 but less than GPT-4 ($0.015 per 1K tokens input). Here's what matters: it's also more reliable. Fewer hallucinations. More consistent.

Sonnet is the middle ground. It's priced between GPT-3.5 and GPT-4 ($0.003 per 1K tokens input). But here's the thing—it benchmarks almost as high as Opus on most enterprise tasks. Reasoning, code generation, RAG, document understanding.

Haiku is the new frontier. It's genuinely small. $0.00025 per 1K tokens. But it's not a toy. It can handle real tasks. Classification, Q&A against context, simple reasoning. It's what Llama wished it could be.

For building enterprise systems, that mix changes everything.

Why This Matters for ERP and Enterprise Systems

Last month we were working on a logistics ERP. One of the components is a smart requisition system that needs to look at vendor contracts, understand price breaks, compare against historical data, and make recommendations.

On GPT-4, this works great but costs money. We were looking at maybe $50K a year just for the model inference.

On Sonnet, we ran the same system. Same prompts. Same data. Same results. One-tenth the cost.

That's not a marginal improvement. That's the difference between a feature you can ship and a feature you can't justify.

But here's what really matters: we didn't have to choose between smart and cheap. We got both.

The Context Window Problem Gets Solved

Claude 3 ships with 200K context window on Opus and Sonnet. That's the entire codebase of most companies. It's the entire ERP database schema. It's six months of emails.

GPT-4 has 8K, sometimes 32K if you pay more. When you're trying to understand an enterprise system, that's cramped. You're always choosing what to include and what to leave out.

With 200K, you include everything. You pass the whole contract to the AI. You pass the entire conversation history. You pass the data model, the audit logs, the notes. The model understands full context.

For ERP systems, this is fundamental. Now the AI can reason about the whole system, not fragments of it. When a client says "something's wrong with our supply chain," the AI can ingest their entire transaction history and actually find it. Not guess. Find it.

Constitutional AI Means Fewer Hallucinations

This is the thing that doesn't get enough attention. Claude was trained with Constitutional AI, which means the model is fundamentally more likely to say "I don't know" rather than confidently make something up.

In enterprise systems, hallucinations are expensive. If an ERP system generates a false invoice, or recommends a purchase that violates a contract, or tells a user something incorrect about their data—that's a disaster.

GPT-4 hallucinates. Claude 3 does too. But less. Noticeably less in my testing.

We built a proof-of-concept ERP report generator that uses Sonnet to answer questions about a company's purchasing data. The same POC on GPT-4 generated false data points maybe 2% of the time. On Sonnet, it was closer to 0.2%. That's real.

The Catch: It's Still Not Perfect

Let me be clear: Claude 3 is not magic. It still makes mistakes. It still hallucinates, just less. It still struggles with very long reasoning chains where it needs to track 20 variables across 10 steps.

And there's a build-the-team tax. If you're using Sonnet instead of GPT-4, you're now depending on a different model. Different behavior, different edge cases. You need to retrain whoever's building the system on how Anthropic's models behave.

But that's not a reason to stick with GPT-4. That's just the cost of switching, and it's worth paying.

What I'm Building on This

We're rebuilding one of our internal tools right now to use Claude 3 Sonnet as the backbone. Same tool, now cheaper to run and faster to respond. We're planning to open it up as a service to clients who want an "AI for your ERP" layer.

We're also going deeper on the context window. One of the ideas we're exploring: what if you could ask your ERP system a question in plain English, and the AI would have full access to your schema, your data model, and your audit logs to answer it accurately? You'd never need to write a report again. You'd just ask.

That's not science fiction with Claude 3's context window. That's a quarter away from shipping.

The frontier of enterprise AI just moved. The companies that move with it—that bet on Sonnet instead of sticking with older models, that start building on 200K context, that rely on Constitutional AI's lower hallucination rate—those companies are going to obsolete the ones that don't.