DeepSeek R1 Review: Is the Open-Weight Reasoner Still Worth It?

If you landed here in 2026 asking whether DeepSeek R1 is still the right reasoning model for your workflow, the short answer is “mainly if you run it yourself”: DeepSeek no longer hosts it. deepseek-reasoner served R1 from January 2025 until DeepSeek moved it to V3.1 on 21 August 2025, and the legacy name was retired on 24 July 2026 at 15:59 UTC. We do not run our own evaluations; this DeepSeek R1 review rests on published evidence and pricing as of 22 October 2026.

Our verdict on DeepSeek R1

In our assessment, R1 matters because DeepSeek shipped a reasoning model it billed as “on par with OpenAI-o1” as MIT-licensed open weights (R1 release note, 20 January 2025), and its training method later passed peer review in Nature (17 September 2025). As of 22 October 2026 the weights remain on Hugging Face under the MIT License (R1 model card), and on the evidence of DeepSeek’s January 2025 R1 report we consider R1 a credible option for self-hosted math and logic work. It is no longer on DeepSeek’s API; new code uses deepseek-flash (DeepSeek-V4.1-Flash-1017) or deepseek-v4-pro (DeepSeek-V4-Pro-0813).

Who should use DeepSeek R1, and who shouldn’t

A good fit if you

• Need an open-weight reasoning model for on-premises or regulated workloads; code and weights are MIT-licensed and commercial use is allowed.
• Run math- or logic-heavy pipelines where a long, inspectable chain of thought matters more than latency.
• Want a smaller reasoner: R1 shipped with six distilled models, and the license permits distilling your own.
• Do academic research and need a training pipeline documented in an arXiv report and a peer-reviewed Nature paper.

Not a good fit if you

• Need a hosted API: DeepSeek no longer serves R1, and deepseek-flash costs less than R1’s 2025 launch card at peak and off-peak.
• Want repository-scale coding: DeepSeek reports a higher SWE-bench Verified score for V3.1 (see Coding below).
• Still call deepseek-reasoner: that name was retired on 24 July 2026.
• Need short answers: R1’s reasoning traces are long (see Math and formal reasoning).

How we assessed it

This review is based on DeepSeek’s official documentation, technical reports, official pricing and named public benchmarks. We do not run our own evaluations; scores are vendor-reported, API cost figures are arithmetic on DeepSeek’s published prices, and training costs are DeepSeek’s own published estimates. Sources:

• DeepSeek-R1 technical report, arXiv 2501.12948 (v1, 22 January 2025); peer-reviewed in Nature 645, 633–638 (17 September 2025), with Supplementary Information.
• DeepSeek release notes: R1 (20 January 2025), R1-0528 (28 May 2025), V3.1 (21 August 2025), V4 Preview (24 April 2026), V4-Pro GA (13 August 2026), V4.1-Flash (10 September 2026), V4.1-Flash-1017 (17 October 2026).
• DeepSeek API Change Log, entries of 20 January, 28 May, 21 August, 22 and 29 September and 1 December 2025, and 24 April, 13 August, 10 September and 17 October 2026.
• DeepSeek-R1 and DeepSeek-R1-0528 model cards on Hugging Face (fetched 22 October 2026).
• Models & Pricing page; Thinking Mode, JSON Output and Anthropic API guides (fetched 22 October 2026).
• DeepSeek-V3 technical report, arXiv 2412.19437 (v1, 27 December 2024).

Verdicts are the editors’ assessment of this evidence, not test results.

What the published evidence shows, by task type

Math and formal reasoning

In our assessment of DeepSeek’s published figures, math is R1’s strongest area. DeepSeek’s R1 technical report (arXiv 2501.12948, January 2025) gave R1 79.8 Pass@1 on AIME 2024 (o1-1217: 79.2) and 97.3 on MATH-500 (o1-1217: 96.4), plus 90.8 on MMLU, 84.0 on MMLU-Pro and 71.5 on GPQA Diamond.

For R1-0528, the Change Log entry of 28 May 2025 reported AIME 2025 rising from 70.0 to 87.5 and GPQA from 71.5 to 81.0; the R1-0528 model card (May 2025) lists AIME 2024 at 79.8 → 91.4. The cost was length: on the AIME test set the card reports about 23K tokens per question for R1-0528 against about 12K for R1, and the Change Log warned that complex reasoning tasks “may consume more tokens”.

Coding

DeepSeek’s R1 report gave R1 a Codeforces rating of 2,029 (96.3 percentile; o1-1217: 2,061) and 49.2 on SWE-bench Verified (o1-1217: 48.9). For R1-0528, the Change Log (28 May 2025) reported LiveCodeBench 63.5 → 73.3 and Aider 57.0 → 71.6; the model card lists SWE Verified at 57.6 (Agentless framework). DeepSeek reports a higher SWE-bench Verified score for V3.1 (66.0; Change Log, 21 August 2025). For deepseek-flash (V4.1-Flash-1017) it reports 77.0 on DeepSWE v1.1 and 68.2 on NL2Repo-Bench (Change Log, 17 October 2026); R1 has no score on those, so they are not a direct comparison. All are vendor-reported; versions and harnesses differ.

Long-context document work

DeepSeek’s R1 model card lists a 128K context length. R1 scored 82.5 on FRAMES against 73.3 for DeepSeek-V3 (R1 report); the Nature paper’s Supplementary Information calls FRAMES “a long-context-dependent QA task”. The original R1 card advised against using a system prompt; the R1-0528 card says “System prompt is supported now.” DeepSeek’s current API models offer 1M context and up to 384K output tokens (Models & Pricing page, 22 October 2026).

Structured output (JSON mode, tools)

R1-0528 added JSON output and function calling (release note and Change Log, 28 May 2025). Since DeepSeek’s API no longer serves R1, the JSON advice below applies to deepseek-flash and deepseek-v4-pro, which both list JSON Output and Tool Calls. DeepSeek’s JSON Output guide (fetched 22 October 2026) says to set response_format to {"type": "json_object"}, include the word “json” and an example in the prompt, and set max_tokens so the JSON is not truncated; the API may occasionally return empty content, so validate it. More in the DeepSeek models hub.

Benchmark table (from the R1 technical report)

From the DeepSeek-R1 technical report (arXiv 2501.12948, v1, 22 January 2025); the same values appear on the R1 model card and in Supplementary Table 5 of the Nature version (17 September 2025). Columns use the versions DeepSeek cited; don’t cross-compare with newer generations without a fresh source. A dash means no value was reported.

More on R1 against o1 at the time: DeepSeek R1 vs OpenAI o1.

How to access R1 in 2026

DeepSeek itself now offers R1 only as open weights.

1. Web chat and mobile app (no longer R1). R1 launched behind the “DeepThink” button on chat.deepseek.com (20 January 2025); on 17 October 2026 DeepSeek’s official web and app models were updated to DeepSeek-V4.1-Flash-1017 (Change Log). See our DeepSeek chat guide.
2. DeepSeek API (no longer R1). deepseek-reasoner served R1 from 20 January 2025 (R1-0528 from 28 May 2025), moved to V3.1 on 21 August 2025 (then V3.1-Terminus on 22 September, V3.2-Exp on 29 September and V3.2 on 1 December 2025), and from 24 April 2026 routed to deepseek-v4-flash in thinking mode, with deepseek-chat on its non-thinking mode. Both names were retired on 24 July 2026 at 15:59 UTC. DeepSeek V4-Flash was itself retired on 10 September 2026; its name is still accepted but served by DeepSeek-V4.1-Flash-1017 at the Flash price.
3. Self-hosting open weights (still open). Code and weights are MIT-licensed. DeepSeek points to the DeepSeek-V3 repository for the full model (671B parameters, 37B activated) and gives vLLM and SGLang commands for the distilled checkpoints. See install DeepSeek locally or running DeepSeek on Ollama.

Minimal API call (OpenAI SDK)

Chat requests hit POST https://api.deepseek.com/chat/completions (OpenAI format); DeepSeek also offers an Anthropic-format endpoint at https://api.deepseek.com/anthropic. A current call uses deepseek-flash:

from openai import OpenAI
client = OpenAI(base_url="https://api.deepseek.com", api_key="...")

# deepseek-flash (DeepSeek-V4.1-Flash-1017). Thinking is the default;
# these fields make it explicit, as in the official docs.
resp = client.chat.completions.create(
    model="deepseek-flash",
    messages=[{"role": "user", "content": "Prove that sqrt(2) is irrational."}],
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
)
print(resp.choices[0].message.reasoning_content)  # chain of thought
print(resp.choices[0].message.content)            # final answer

The API keeps no conversation state: as the multi-turn sample in DeepSeek’s Thinking Mode guide shows, the client resends the message list each turn. reasoning_effort accepts low, high (default) or max, and thinking takes {"type": "enabled"} or {"type": "disabled"}. In thinking mode temperature, presence_penalty and frequency_penalty have no effect. With tools, earlier reasoning_content must be passed back. More in our DeepSeek API documentation.

Pricing and value for money

The first row below is R1’s launch rate card (deepseek-reasoner, 20 January 2025, R1 release note); it is history. Current prices as of 22 October 2026 (Models & Pricing page), USD per 1M tokens:

deepseek-flash is cheaper than R1’s launch card on every line, at peak and off-peak. At peak, deepseek-v4-pro is cheaper than R1’s launch card only on cache-hit input (its cache-miss input is 2.4× and its output about 1.81× R1’s). Off-peak it is cheaper on cache-hit input and on output ($1.98 vs $2.19), but its cache-miss input is still 1.2× R1’s. Self-hosting moves the cost to hardware you rent or own; the DeepSeek pricing calculator estimates hosted API cost on deepseek-flash or deepseek-v4-pro at peak and off-peak rates.

Worked example (illustrative): deepseek-flash, peak and off-peak

Hypothetical, at 22 October 2026 rates: 1,000,000 calls, each with a 2,000-token cached system prompt, a 200-token uncached user message and 300 output tokens. Assumptions: every call hits the cached prefix (the first call’s miss is ignored as negligible), and the 300 output tokens include any reasoning_content, since thinking is the default. Reasoning-heavy prompts can produce far more output (R1-0528’s model card cites about 23K tokens per AIME question), and output cost scales linearly with length.

• Cached input: 2,000 × 1,000,000 = 2,000,000,000 tokens. Peak: 2,000,000,000 ÷ 1,000,000 × $0.006 = $12.00; off-peak: 2,000,000,000 ÷ 1,000,000 × $0.003 = $6.00
• Uncached input: 200 × 1,000,000 = 200,000,000 tokens. Peak: 200,000,000 ÷ 1,000,000 × $0.30 = $60.00; off-peak: 200,000,000 ÷ 1,000,000 × $0.15 = $30.00
• Output: 300 × 1,000,000 = 300,000,000 tokens. Peak: 300,000,000 ÷ 1,000,000 × $1.20 = $360.00; off-peak: 300,000,000 ÷ 1,000,000 × $0.60 = $180.00
• Total: $432.00 at peak; $216.00 off-peak (exact, no rounding)

For context, the same volumes on other rate cards (tokens ÷ 1,000,000 gives 2,000 cached, 200 uncached and 300 output, each multiplied by the rate):

• deepseek-v4-pro, peak: 2,000 × $0.044 = $88.00; 200 × $1.32 = $264.00; 300 × $3.96 = $1,188.00; total $1,540.00
• deepseek-v4-pro, off-peak: 2,000 × $0.022 = $44.00; 200 × $0.66 = $132.00; 300 × $1.98 = $594.00; total $770.00
• R1 launch card (historical): 2,000 × $0.14 = $280.00; 200 × $0.55 = $110.00; 300 × $2.19 = $657.00; total $1,047.00

DeepSeek’s earlier off-peak discount ran until 5 September 2025; off-peak pricing returned at half the peak rate from 16:00 UTC on 16 August 2026 (Change Log, 13 August 2026). Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; all other hours are off-peak.

Training cost context

DeepSeek’s peer-reviewed Nature paper (17 September 2025; Supplementary Table 4) estimated $294,000 for R1-Zero ($202,000), SFT data creation ($10,000) and R1 ($82,000) combined: 147,000 H800 GPU-hours at an assumed rental price of $2 per GPU-hour, with R1-Zero and R1 trained on 64×8 H800 GPUs. That is an estimate, not reported spending, and it excludes the base model: the DeepSeek-V3 technical report (December 2024) estimated $5.576M for V3’s official training (2.788M H800 GPU-hours at the same $2), excluding prior research and ablation experiments. Together that is $5.576M + $0.294M = $5.870M, still a rental-price estimate.

Competitor context

DeepSeek’s headline comparison was with OpenAI-o1-1217: in its January 2025 report, R1 led on AIME 2024, MATH-500 and SWE-bench Verified (49.2 vs 48.9) and trailed on GPQA Diamond, MMLU and Codeforces. That is a dated, vendor-reported snapshot, not a verdict on OpenAI’s current models; see our DeepSeek vs OpenAI o1 page, and the DeepSeek alternatives for reasoning roundup for other open-weight and proprietary reasoners.

DeepSeek’s own API now serves deepseek-flash and deepseek-v4-pro (see our DeepSeek V4 guide). DeepSeek has kept releasing open weights (V4 Preview, 24 April 2026; V4.1-Flash, 10 September 2026; V4.1-Flash-1017, 17 October 2026), so R1 is not the only DeepSeek reasoning checkpoint you can pin. Our assessment: R1 is the better choice mainly when you need that specific, peer-reviewed 2025 reasoner, to audit or reproduce published results.

Strengths and weaknesses, distilled

Strengths

• At release (January 2025), DeepSeek reported 97.3 on MATH-500 and 79.8 on AIME 2024.
• MIT-licensed code and weights; commercial use and distillation permitted.
• Documented training pipeline (arXiv, January 2025; Nature, September 2025) and six distilled checkpoints released with R1.
• Documented run paths for full R1 and the distilled models (R1 model card).

Weaknesses

• Long reasoning traces (about 23K tokens per AIME question for R1-0528); DeepSeek says V3.1-Think answers in less time than R1-0528 (Change Log, 21 August 2025; vendor claim).
• Not a coding specialist: SWE-bench Verified 49.2 (R1) and 57.6 (R1-0528), against V3.1’s 66.0 (DeepSeek-reported; harnesses may differ).
• 128K context (R1 model card), against 1M on DeepSeek’s current API models.
• No DeepSeek-hosted option since August 2025; the legacy names were retired on 24 July 2026.

For strengths across the family, see our DeepSeek capabilities guide and the DeepSeek reviews hub.

Final verdict

Our assessment, based on DeepSeek’s January 2025 R1 report, the R1 and R1-0528 model cards and official pricing as of 22 October 2026: R1 is still worth running in 2026 if you want MIT-licensed open weights for on-premises reasoning, or are studying how a peer-reviewed reasoning pipeline was built. For hosted use, point new code at deepseek-flash, cheaper than R1’s launch card at peak and off-peak, or deepseek-v4-pro, whose peak cache-miss and output rates are above that card. In our view, R1’s practical role in a 2026 stack is as an archived, auditable baseline.

DeepSeek AI Guide is an independent resource and is not affiliated with DeepSeek or its parent company. Model IDs, pricing and API behaviour change; check the official DeepSeek documentation and pricing page before committing to a production decision.