"OpenAI o1 ranks in the 89th percentile on competitive programming questions and places among the top 500 students nationally on a USA Math Olympiad qualifier." — OpenAI Blog, September 12, 2024
OpenAI released o1 in preview this month, and it's a fundamentally different kind of language model.
Not different in degree — faster, more tokens, better benchmarks — but different in kind. o1 (previously codenamed "Strawberry") uses chain-of-thought reasoning at inference time. Before it answers, it thinks. Sometimes for a few seconds; sometimes for much longer on hard problems. And that thinking process is substantive: the model explores the problem, considers approaches, spots errors in its own reasoning, and revises.
The results on complex reasoning tasks are dramatic. PhD-level performance on physics, chemistry, and biology benchmarks. Near-perfect scores on math olympiad problems. Competitive coding performance that would place it among the top human coders.
For most of what businesses use AI for today, this doesn't change much. But for a specific class of problems, it changes everything.
What's Different About Reasoning Models
Standard language models like GPT-4 or Claude work in a single forward pass. They see your prompt, they generate a response token by token, and they're done. Their "reasoning" happens implicitly through patterns in their training, not through an explicit, iterative thinking process.
o1 adds an extended thinking phase before output. The model reasons through the problem — the visible "thinking" traces show it examining cases, considering alternatives, catching mistakes — and then produces a final answer based on that reasoning process.
The key differences:
- Slower: o1 takes longer to respond than GPT-4 because it's actually thinking. Responses can take 10-30+ seconds for complex problems.
- More expensive: Significantly higher cost per query at current pricing.
- Dramatically more reliable on hard reasoning tasks: On problems that require multi-step logic, mathematical reasoning, or complex deduction, o1 is dramatically more accurate.
- Worse on some simple tasks: The extended reasoning is overkill for simple tasks and can actually introduce more errors by overthinking.
The Business Use Cases That This Actually Changes
Complex Financial and Legal Analysis
The difference between o1 and GPT-4 is most pronounced on problems that require careful multi-step reasoning where errors compound. Financial modelling validation, contract risk assessment, regulatory compliance analysis — tasks where a chain of reasoning errors leads to a wrong conclusion — are genuinely different propositions with o1.
Before o1, these use cases required extensive prompt engineering, chain-of-thought scaffolding, and even then often produced unreliable results on harder cases. o1 handles them more naturally because the reasoning is built into how the model works.
Technical Document Analysis and Engineering Review
Analysing engineering specifications, identifying inconsistencies across large technical documents, reviewing code for security vulnerabilities — tasks that require following complex logic chains and checking work — benefit significantly from reasoning-first models.
A manufacturing client we're working with has documentation across multiple interconnected standards documents, and reasoning about how specifications in one document affect requirements in another is genuinely hard multi-step reasoning. o1 handles this meaningfully better than what we had before.
Research Synthesis and Competitive Analysis
When you need the model not just to summarise but to reason — "given these five market reports, what's the most likely competitive response to a price change?" — reasoning model quality shows up clearly in output quality.
Code Generation for Complex Systems
o1's coding performance is particularly notable. For complex algorithmic problems or systems design involving multiple interacting constraints, o1's approach produces more correct solutions and catches more edge cases. For routine code generation, GPT-4 or Claude 3.5 remain better choices on cost/quality grounds.
Where o1 Doesn't Help (Or Actively Hurts)
High-volume, routine tasks. For RAG-based question answering, routine summarisation, classification, or extraction — tasks that don't require complex multi-step reasoning — o1's cost and latency make it the wrong choice. You're paying a large premium for a capability you're not using. Conversational interfaces. The latency is incompatible with snappy conversational experiences. Users waiting 15+ seconds for a response in a chat interface have a bad experience. Creative generation. o1 isn't optimised for fluent creative writing. Claude 3.5 Sonnet or GPT-4o is still better for content generation tasks.The Practical Guidance
Think of o1 as a specialist, not a generalist. It's the model you reach for when the task genuinely requires careful multi-step reasoning and accuracy is more important than cost or speed.
The pattern I'd recommend for most production systems: use GPT-4o or Claude 3.5 Sonnet as your default. When a specific task within your system requires hard reasoning — analysis, validation, complex planning — route that specific task to o1. A hybrid architecture like this gets you the cost efficiency of standard models for 90% of queries and the reasoning capability of o1 where it matters.
We're starting to design systems this way. It's a bit more complex architecturally but the cost economics are much more favourable than routing everything through o1.
The Bigger Picture
Reasoning models represent a meaningful expansion of what AI can do for knowledge-intensive work. The gap between "AI that produces plausible text" and "AI that can actually think through a complex problem" is narrowing in meaningful ways.
This changes the ROI calculation for knowledge-worker automation. Tasks that were previously "AI can assist but humans need to verify the reasoning" start to shift toward "AI can do the reasoning, humans verify the conclusions." That's a meaningful change in how much of the workflow the AI can own.
Watch the pricing trajectory. As reasoning model costs decline — and they will, on the same trajectory as every previous generation of models — the business applications that become viable will expand significantly.
Want to implement AI in your business? Book a 30-minute architecture call at blok-ai.com.