DeepSeek R1 in 2026: What the Open Reasoning Model Still Teaches

DeepSeek R1 is the model people still cite when they talk about open-weight reasoning, and it is now one of the older models you can reasonably still call through an API. Released on 20 January 2025, it showed that a reasoning model with visible chain of thought could be published under the MIT licence and land within a few points of OpenAI o1 on maths and code.

Almost two years later the question is different. Newer DeepSeek models do more for less, so R1 is no longer the obvious choice for anything. What it still offers is a well-documented reference point, a family of distilled small models, and some prompting rules that catch people out if they treat it like a chat model.

What R1 is

R1 is a mixture-of-experts model with 671B total parameters and 37B active per token, with a 128K context window according to DeepSeek's model card. The repository and weights are MIT-licensed, and DeepSeek explicitly permits commercial use and distillation into other models.

Hosted context windows vary by provider and have often been shorter than the model's maximum. OpenRouter's listing for the original R1 shows 64,000 tokens of context, for example, so check what your provider actually serves rather than assuming the full 128K.

At release DeepSeek compared it with OpenAI o1-1217. R1 scored 79.8 on AIME 2024 against 79.2, 97.3 on MATH-500 against 96.4, and 65.9 on LiveCodeBench against 63.4. It trailed on GPQA-Diamond, 71.5 against 75.7, and slightly on MMLU and Codeforces. Close to parity on maths and code, a step behind on graduate-level science questions.

How it was trained, and why that mattered

DeepSeek published two models together. R1-Zero was trained with large-scale reinforcement learning directly on the base model, with no supervised fine-tuning first. It learned to reason, but suffered from endless repetition, poor readability and mixing languages mid-answer.

R1 added a small amount of curated "cold-start" data before the reinforcement learning stage, which fixed most of those problems while keeping the reasoning ability. The wider significance was that reasoning behaviour could emerge from reinforcement learning on verifiable rewards rather than needing vast amounts of hand-written reasoning traces. Reasoning models explained covers the general mechanism, and test-time compute explained covers why spending tokens on thinking buys accuracy on some tasks.

The 0528 update changed real behaviour

On 28 May 2025 DeepSeek shipped R1-0528, a post-training update on the same architecture. It is not a cosmetic refresh. From DeepSeek's model card, original R1 against 0528:

• AIME 2025: 70.0 to 87.5
• GPQA-Diamond: 71.5 to 81.0
• LiveCodeBench: 63.5 to 73.3
• SWE-bench Verified: 49.2 to 57.6
• Aider-Polyglot: 53.3 to 71.6
• Humanity's Last Exam: 8.5 to 17.7
• SimpleQA: 30.1 to 27.8, the one regression

The cost of that improvement is length. DeepSeek reports average reasoning on AIME questions rising from about 12K tokens to about 23K. The model thinks roughly twice as long to get there, and you pay for every one of those tokens. The token cost of reasoning models works through what that does to a bill.

0528 also fixed practical irritations. It supports a system prompt, which the original did not, has better function calling, a lower reported hallucination rate, and no longer needs the thinking tag forced at the start of the output. If you are choosing between the two R1 endpoints, 0528 is the better model for nearly everything. The dated identifier is also the safer thing to put in code, because it names one specific snapshot.

Prompting the original R1 correctly

The original R1 has rules that differ from ordinary chat models, and ignoring them produces output that looks like a weak model rather than a misconfigured one. From DeepSeek's usage recommendations:

• Temperature between 0.5 and 0.7, with 0.6 recommended. Higher values cause repetition or incoherent output.
• No system prompt. Put every instruction in the user message.
• For maths, ask it to reason step by step and put the final answer in a boxed expression, so the answer is easy to extract.
• Force the output to begin with the <think> tag, because the model sometimes skips reasoning otherwise.

Agent frameworks that inject a long system prompt and run at temperature 1.0 break at least two of those. If someone tells you R1 performed badly in their harness, ask about the settings first.

The distilled models are the lasting contribution

Alongside R1, DeepSeek released six dense models distilled from its reasoning traces: Qwen2.5-based models at 1.5B, 7B, 14B and 32B, and Llama-based models at 8B and 70B. With 0528 came another, R1-0528-Qwen3-8B, which DeepSeek says matches Qwen3-235B-thinking on AIME 2024.

These small models are where R1's influence is most practical today, because most of them fit on a single consumer or workstation GPU, the larger ones once quantised. They also made a point that has held up: a large reasoning model's traces are an effective way to teach a small model to reason. Model distillation explained covers the technique and its limits.

Should you still run R1 in 2026

For most new work, no. On OpenRouter at the time of writing, the original R1 is listed at around $0.70 per million input tokens and $2.50 output, and 0528 at about $0.50 and $2.15. DeepSeek's V4-generation models are listed for a fraction of that, with 1M-token windows, native tool use inside reasoning and far stronger agentic scores. The V4.1 Flash guide is the better starting point for a current DeepSeek deployment.

R1 still makes sense in three situations: reproducing research that used it, running one of the distilled models locally, or keeping an existing pipeline stable while you evaluate a replacement. In the last case, treat it as a migration project with a deadline. Hosted availability of an older model is at the provider's discretion, and the cheapest time to move is before an endpoint disappears.