Anthropic released Claude 4 Opus last month, and enterprise technology leaders are asking whether it changes their AI platform decisions. I’ve been running evaluations with several organisations. Here’s what we’re finding.

What Anthropic Claims

The headline improvements:

Better reasoning. Significantly improved performance on complex, multi-step problems. Particularly strong on logic, mathematics, and code.

Longer context. 200,000 token context window, with claims of better utilisation of that context than previous models.

Improved instruction following. More reliable adherence to detailed prompts and constraints.

Enhanced safety. Updated constitutional AI approach with better refusal calibration (less over-refusing safe requests).

Multimodal improvements. Better image understanding and reasoning about visual content.

These claims are meaningful if true. Let’s check against real-world testing.

Our Testing Approach

We tested Claude 4 Opus against GPT-4 Turbo and Claude 3.5 Opus on tasks relevant to enterprise use:

• Document analysis and summarisation
• Contract clause extraction and comparison
• Code review and debugging
• Multi-document synthesis
• Data analysis and interpretation
• Customer communication drafting

Testing used real enterprise documents (anonymised as necessary) rather than synthetic benchmarks.

What We Found

Document Analysis: Strong Performance

Results: Claude 4 Opus performed noticeably better than previous Claude versions on long document tasks. Summarisation quality was high. Important details were captured with fewer omissions.

Comparison with GPT-4: Performance was comparable. Neither had a clear advantage on standard document tasks. Claude’s longer context window was advantageous for very long documents.

Verdict: Genuine improvement. Enterprises with document-heavy use cases should evaluate.

Complex Reasoning: Meaningful Improvement

Results: On multi-step reasoning tasks – contract analysis requiring cross-referencing, code debugging requiring tracing – Claude 4 Opus showed clear improvement over Claude 3.5.

Comparison with GPT-4: Competitive. Different models had different strengths. Claude 4 Opus was particularly strong on code-related reasoning.

Verdict: The reasoning improvements are real and practically useful.

Instruction Following: Better but Imperfect

Results: Claude 4 Opus was more reliable at following complex, multi-part instructions. Fewer cases of ignoring constraints or missing requirements.

Comparison with GPT-4: Similar performance. Both models are now quite good at instruction following. The gap with previous generations has closed.

Verdict: Improvement evident, though not dramatic compared to other current models.

Context Utilisation: Where It Matters

Results: For tasks involving 100k+ tokens of context, Claude 4 Opus maintained coherent understanding better than alternatives. Information from early in documents was still accessible when needed.

Comparison with GPT-4: Claude’s longer context window is a genuine differentiator for appropriate use cases. GPT-4’s context is sufficient for most enterprise tasks but limiting for some.

Verdict: If your use cases involve very long documents, Claude’s context advantage is real.

Safety and Refusals: Calibration Improved

Results: Fewer inappropriate refusals on legitimate enterprise tasks. Previous Claude versions sometimes over-refused on business content that happened to mention sensitive topics.

Comparison with GPT-4: Both models are now well-calibrated for enterprise use. Previous Claude versions were notably more conservative.

Verdict: Important improvement for enterprise adoption. Reduces friction in legitimate use cases.

What This Means Practically

For Organisations Evaluating AI Platforms

Claude 4 Opus is now a credible option for enterprise AI. The performance gap with GPT-4 that existed in earlier generations has largely closed.

Platform choice should be driven by:

• Existing cloud relationships (Azure for GPT-4, AWS/GCP for Claude)
• Specific use case requirements
• Enterprise features and compliance requirements
• Pricing for your expected usage

Model capability is no longer a deciding factor – it’s table stakes.

For Organisations Already Using Claude

If you’re on Claude 3.5, evaluating Claude 4 Opus makes sense. The improvements are meaningful for many use cases. Migration should be straightforward.

Focus evaluation on your actual use cases. Generic benchmarks matter less than performance on your specific tasks.

For Organisations on GPT-4

No compelling reason to switch based on Claude 4 Opus alone. The capabilities are competitive, not clearly superior.

Worth evaluating if:

• Your use cases involve very long documents (context advantage)
• You’re unhappy with current performance on specific tasks
• You want competitive pressure on pricing

Not worth evaluating just because there’s a new model.

Enterprise Considerations

Beyond raw capability:

Availability: Claude 4 Opus is available through Amazon Bedrock and directly from Anthropic. Azure availability is via partnership arrangements.

Pricing: Competitive with GPT-4. Token-based pricing means actual costs depend on usage patterns.

Enterprise features: Anthropic’s enterprise offering has matured but still trails Azure OpenAI on some enterprise governance features.

Australian data handling: Clarity on Australian data processing is important for local enterprises. Confirm specifics before committing.

Our Recommendation

For enterprises:

1. Don’t switch platforms based on one model. Platform capabilities, ecosystem, and your existing investments matter more than marginal model differences.

2. Evaluate if considering Claude. If you’re AWS-native or evaluating alternatives, Claude 4 Opus makes Anthropic’s offering competitive.

3. Test on your use cases. Our findings may not match yours. Benchmark what matters to you.

4. Monitor the space. Model capabilities are converging. The differentiators are increasingly elsewhere.

Final Thought

Claude 4 Opus represents continued maturation of enterprise AI options. It’s a capable model that closes previous gaps with GPT-4.

The exciting conclusion: enterprises now have multiple credible options. Competition benefits customers through better products and pricing.

The pragmatic conclusion: model capability is becoming commoditised. Where you add value is in how you apply AI to your business, not which model you choose.

Choose wisely, but don’t overthink it. The model matters less than what you do with it.