In September 2024, OpenAI released o1-preview, a model that fundamentally changed what we expected from language models. Rather than generating answers in a single forward pass, o1 "thinks" -- it generates an internal chain of reasoning before producing a final answer. The result was a dramatic leap in performance on hard problems: competition-level math, PhD-level science questions, and complex coding challenges that standard models struggled with. Since then, DeepSeek-R1 and Alibaba's QwQ have joined the reasoning model category, and the implications for the AI industry are significant.
What Makes Reasoning Models Different
Standard language models (GPT-4o, Claude 3.5 Sonnet, Gemini 2.0) generate tokens left-to-right in a single pass. They are remarkably capable, but their inference process is fundamentally "fast" -- they commit to each token without extensive deliberation. This is efficient and produces good results on most tasks, but it limits performance on problems that require deep, multi-step logical reasoning.
Reasoning models add an explicit "thinking" phase. When o1 receives a problem, it first generates a chain-of-thought (CoT) -- a long internal monologue where it breaks down the problem, considers approaches, checks its work, and refines its answer. This thinking process can involve hundreds or thousands of tokens of internal reasoning before the model produces its final response.
This is analogous to the difference between answering a math problem instantly in your head versus writing out your work on paper. The written-out approach is slower but catches more errors and handles more complexity.
The technical mechanism varies by model: - o1 uses reinforcement learning to train the model to produce effective chains of reasoning. The reasoning tokens are generated but hidden from the user in the API response -- you see only the final answer (though you are billed for the reasoning tokens). - DeepSeek-R1 was trained using a combination of reinforcement learning and distillation from a larger model. Its reasoning chain is visible in the output, giving users transparency into the model's thinking process. - QwQ (Alibaba's Qwen with Questions) takes a similar approach to DeepSeek-R1, producing visible chain-of-thought reasoning.
Performance Breakthroughs
The performance gains on hard problems are not incremental -- they are transformative.
Mathematics On the MATH benchmark (competition-level math), o1 scores approximately 94.8%, compared to GPT-4o's 76.6%. That is a jump from "decent undergraduate" to "competitive mathematician" territory. DeepSeek-R1 achieves 97.3% on MATH, actually surpassing o1 on this benchmark.
On the American Invitational Mathematics Examination (AIME), o1 scored in the 83rd percentile of human contestants. For context, GPT-4o scored in the 13th percentile on the same exam. This is not a marginal improvement -- it represents a qualitative shift in mathematical capability.
Science On GPQA Diamond (PhD-level science), o1 scores 78.0% compared to GPT-4o's 53.6% and Claude 3.5 Sonnet's 65.0%. This is particularly notable because GPQA Diamond questions are designed to be difficult even for human domain experts.
Coding On SWE-bench Verified, o1 achieves approximately 48.9%, competitive with Claude 3.5 Sonnet's leading score. On Codeforces-style competitive programming problems, o1 scores at the 89th percentile of human contestants -- a remarkable achievement.
Explore how reasoning models compare to standard models on our leaderboard page.
The Cost and Latency Tradeoff
Reasoning models are expensive. The internal chain-of-thought generates hundreds to thousands of "thinking tokens" that you are billed for, even though they do not appear in the final output. A single o1 request can consume 10-50x more tokens than a GPT-4o request for the same query.
Current o1 pricing: - Input: $15.00 per million tokens - Output (including thinking tokens): $60.00 per million tokens
Compare this to GPT-4o at $2.50/$10.00. A query that costs $0.01 on GPT-4o might cost $0.30-1.00 on o1, depending on how much reasoning is required. o1-mini offers a more affordable option at $3.00/$12.00, but with reduced capability on the hardest problems.
DeepSeek-R1 is dramatically cheaper: - Input: $0.55 per million tokens - Output: $2.19 per million tokens
This is roughly 10-30x cheaper than o1 for comparable reasoning quality on many tasks. DeepSeek-R1 has made the reasoning paradigm accessible to cost-conscious teams. See our pricing page for current pricing across all reasoning models.
Latency is also higher. While GPT-4o returns first tokens in 300-600ms, o1 can take 10-60 seconds for complex problems because it needs to complete its entire reasoning chain before producing the final answer. For interactive applications, this latency can be prohibitive.
DeepSeek-R1: The Open Source Reasoning Breakthrough
DeepSeek-R1, released in January 2025, was a watershed moment for the reasoning model paradigm. It demonstrated that:
1. Reasoning capabilities can be achieved without the massive compute budgets of OpenAI 2. Open-weight reasoning models can match or exceed o1 on key benchmarks 3. Distilled versions (DeepSeek-R1-Distill-Qwen-32B, DeepSeek-R1-Distill-Llama-70B) can bring reasoning capabilities to smaller, self-hostable models
The availability of DeepSeek-R1's weights has accelerated the entire field. Researchers and companies can study the reasoning patterns, fine-tune on domain-specific reasoning tasks, and deploy reasoning models without dependency on OpenAI's API.
DeepSeek-R1's visible chain-of-thought also enables a level of transparency that o1 lacks. You can see exactly how the model approaches a problem, where it considers alternatives, and how it arrives at its answer. For applications where auditability matters (medical, legal, financial), this transparency is valuable.
QwQ: Alibaba's Reasoning Entry
Alibaba's QwQ (Qwen with Questions) takes a different approach. Built on top of the Qwen 2.5 architecture, QwQ emphasizes question-asking as part of its reasoning process. When faced with an ambiguous problem, QwQ explicitly identifies what it does not know and reasons through the uncertainty.
QwQ-32B-Preview demonstrated competitive performance on math and coding benchmarks while being small enough to run on consumer hardware (with quantization). This makes reasoning models accessible to individual developers and small teams, not just enterprises with API budgets.
When Reasoning Models Make Sense
Reasoning models are not a universal upgrade over standard models. They are a specialized tool that excels in specific scenarios:
Good Use Cases - Competition-level mathematics and scientific reasoning - Complex multi-step logical deductions - Problems that require exploring and evaluating multiple approaches - Code generation for complex algorithmic challenges - Analysis tasks where accuracy is more important than speed
Poor Use Cases - Simple classification or extraction tasks (massive overkill) - Conversational chatbots (too slow, too expensive) - High-volume content generation (cost prohibitive) - Tasks requiring low latency (reasoning adds 10-60 seconds) - Routine coding tasks that standard models handle well
The mistake many teams make is using reasoning models for every task. GPT-4o or Claude 3.5 Sonnet is the right choice for 90%+ of production workloads. Reasoning models are for the 5-10% of tasks where standard models genuinely fall short.
The Reasoning Model Landscape
Here is how the current reasoning models compare:
o1 (OpenAI) - Strengths: Best overall reasoning quality, strong across math/science/coding - Weaknesses: Expensive ($15/$60 per M tokens), hidden reasoning chain, high latency - Best for: Teams that need the highest possible accuracy on hard problems and can absorb the cost
o1-mini (OpenAI) - Strengths: Cheaper ($3/$12 per M tokens), good coding performance, lower latency than full o1 - Weaknesses: Reduced performance on the hardest science/math problems compared to full o1 - Best for: Coding tasks where reasoning helps but full o1 is overkill
DeepSeek-R1 (DeepSeek) - Strengths: Near-o1 performance at 10-30x lower cost, visible chain-of-thought, open weights available - Weaknesses: Less mature API ecosystem, occasional reasoning loops on edge cases - Best for: Cost-conscious teams that need reasoning capabilities, research, self-hosting
QwQ-32B (Alibaba) - Strengths: Small enough for self-hosting, strong math/coding for its size, open weights - Weaknesses: Smaller model means lower ceiling on the hardest problems, less diverse training - Best for: Self-hosting reasoning capabilities, resource-constrained environments
Claude 3.5 Sonnet with Extended Thinking (Anthropic) - Strengths: Integrated into the familiar Claude API, strong coding reasoning, visible thinking - Weaknesses: Higher latency when extended thinking is enabled, premium pricing - Best for: Teams already in the Anthropic ecosystem who need reasoning on demand
Compare all reasoning models on our comparison page to see detailed benchmark breakdowns.
What This Means for the Industry
The rise of reasoning models signals several important trends:
Compute-Time Scaling Traditional AI scaling laws focus on training compute -- bigger models trained on more data. Reasoning models introduce inference-time scaling: you can improve outputs by spending more compute at inference time. This changes the economics of AI from a fixed-cost model (training) to a variable-cost model (per-request reasoning depth).
Specialization Over Generalization Rather than one model that handles everything, we are moving toward a portfolio approach: fast, cheap models for routine tasks and slower, expensive reasoning models for hard problems. Smart routing between these tiers will become a critical capability.
Open Source Competitiveness DeepSeek-R1 proved that reasoning capabilities are not exclusive to well-funded labs. The open source community can replicate and extend these techniques, which keeps the field competitive and accessible.
The Verification Problem Reasoning models produce more accurate answers on hard problems, but they are also more expensive and slower. Verifying that a problem truly needs reasoning (rather than just using a standard model) becomes an important engineering challenge.
Practical Advice
If you are evaluating reasoning models for your workload:
1. Start with standard models (GPT-4o, Claude 3.5 Sonnet). Only add reasoning models if you identify specific failure modes. 2. Use reasoning models surgically -- for the specific request types where standard models fail. 3. Consider DeepSeek-R1 before o1 unless you need the absolute best quality. The 10-30x cost difference is significant. 4. Build your routing layer early. The ability to send different requests to different models is the foundation of cost-effective AI architecture. 5. Monitor reasoning token usage. A single o1 request can consume thousands of thinking tokens; set usage alerts.
The reasoning model paradigm is real and here to stay. But like every AI capability, it is a tool to be used strategically, not a default for every request. Explore the full reasoning model landscape in our model directory to find the right fit for your needs.