Menu

Comparisons · both · deepseek vs claude

DeepSeek versus Claude: price on the snapshot, quality on the public boards

Published 16 Oct 2026. Updated 16 Oct 2026. Researched 16 Oct 2026.

People search for this when a coding bill arrives. The honest split is: the token price gap is large and sourced, and the “it codes as well as Claude” claim is not something this site can support from the files we have.

Price

DeepSeek’s pricing page, not a reseller:

V4.1 Flash peak V4.1 Flash off-peak V4 Pro peak V4 Pro off-peak
Input, cache miss $0.30 $0.15 $1.32 $0.66
Output $1.20 $0.60 $3.96 $1.98

Per million tokens. Off-peak is half, on the hours in the DeepSeek hub. The OpenRouter chart is a different number for V4 Pro.

DeepSeek on the OpenRouter list, not the direct off-peak ratedeepseek-v4.1-flash$0.198/Mdeepseek-v4-pro$1.19/Mclaude-opus-5.5$8.00/Mgpt-6.1-sol$4.00/M
DeepSeek on the OpenRouter list, not the direct off-peak rate. OpenRouter list price, blended 75% input / 25% output. Snapshot 2026-10-16 06:10 UTC. Comparators resolve to Anthropic: Claude Opus 5.5 and OpenAI: GPT-6.1 Sol.
ModelBlendedvs Opusvs Sol
deepseek-v4.1-flash$0.19898% cheaper95% cheaper
deepseek-v4-pro$1.1985% cheaper70% cheaper
Illustrative Coding-agent session costDeepSeek-V4.1-Flash$0.178DeepSeek-V4-Pro-0813$0.752Anthropic: Claude Opus 5.5$2.94OpenAI: GPT-6.1 Sol$1.47
Coding-agent session: 3 million input tokens, 90% served from cache, 60 thousand output tokens. Cache-write fees are not included. Chinese rows use the lab published rate (DeepSeek uses the peak rate). Comparators are OpenRouter list prices for Anthropic: Claude Opus 5.5 ($8.00/M blended) and OpenAI: GPT-6.1 Sol. Snapshot 2026-10-16. Cache-write fees are excluded. A missing cache price is billed as a full input token.
ModelSessionvs Opusvs SolCache price
DeepSeek-V4.1-Flash$0.17894% cheaper88% cheaperpublished
DeepSeek-V4-Pro-0813$0.75274% cheaper49% cheaperpublished
Anthropic: Claude Opus 5.5$2.94——published
OpenAI: GPT-6.1 Sol$1.47——published

The percentages in this table compare session cost. The price table's percentages compare blended token price. They answer different questions.

The task chart uses DeepSeek’s peak rate for those two rows and the OpenRouter list price for the Opus and Sol comparators. Off-peak would cut the DeepSeek bars in half. Cache writes are not in the formula.

Preference votes, not coding tests

On the LMArena style-control snapshot published 15 Oct 2026, the coding category places the current Opus row and the current Sol row above deepseek-v4.1-flash-max. The gap in rating points is modest next to the price gap. Votes are blind preferences on prompts people chose to send, not a fix-the-issue benchmark. Intervals overlap more than a rank number suggests. Read the chart’s detail column.

LMArena coding category, style controlClaude Opus 5.5 (High)1547.52gpt-6.1-sol-max1539.79Kimi K3 (Max)1537.83mimo-v2.6-pro1537.58glm-5.3-flash1524.53deepseek-v4.1-flash-max1523.37qwen3.8-max1523.36glm-5.3-max1518.51kimi-k2.61516.03mimo-v2.6-flash1512.93deepseek-v4-pro1500.58qwen3.8-27b1499.13minimax-m31496.69deepseek-v4-flash1484.16hunyuan-hy3-preview1452.74
Style-controlled Bradley-Terry ratings from the LMArena leaderboard dataset. Not Elo. A higher rating means the model was preferred more often in blind votes. Intervals are the dataset's rating_lower and rating_upper. Ratings are rounded to 1 decimal for display in the chart; full values are stored. Category: coding. Published 2026-10-15, fetched 2026-10-16. Source: https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset (text_style_control, latest).
EntryratingDetail
Claude Opus 5.5 (High)1547.52rank 6, 1,894 votes, 1533.61–1561.44
gpt-6.1-sol-max1539.79rank 10, 1,257 votes, 1522.63–1556.95
Kimi K3 (Max)1537.83rank 12, 8,306 votes, 1530.58–1545.09
mimo-v2.6-pro1537.58rank 13, 1,105 votes, 1519.93–1555.23
glm-5.3-flash1524.53rank 27, 8,457 votes, 1517.07–1531.99
deepseek-v4.1-flash-max1523.37rank 29, 3,527 votes, 1513.1–1533.65
qwen3.8-max1523.36rank 30, 7,677 votes, 1515.79–1530.93
glm-5.3-max1518.51rank 38, 5,400 votes, 1509.88–1527.13
kimi-k2.61516.03rank 44, 11,091 votes, 1509.15–1522.92
mimo-v2.6-flash1512.93rank 51, 2,625 votes, 1500.97–1524.89
deepseek-v4-pro1500.58rank 78, 17,136 votes, 1494.66–1506.51
qwen3.8-27b1499.13rank 81, 6,417 votes, 1491.03–1507.23
minimax-m31496.69rank 87, 18,242 votes, 1490.99–1502.39
deepseek-v4-flash1484.16rank 105, 15,288 votes, 1477.95–1490.37
hunyuan-hy3-preview1452.74rank 153, 2,055 votes, 1439.33–1466.16

SWE-bench does not yet answer the V4 question

SWE-bench Verified rows we keptClaude 4.5 Opus (high)76.8MiniMax M2.5 (high)75.8Claude 4.6 Opus75.6GLM 5 (high)72.8Lingxi v1.5 x Kimi K271.2Kimi K2.5 (high)70.8DeepSeek V3.2 (high)70OpenHands + Qwen3-Coder-480B-A35B-69.6GLM-4.668.2OpenHands + Kimi K265.4
Resolved percentage on SWE-bench Verified, as published. Harnesses differ, so this is not a pure model-vs-model ranking. The current Opus and Sol aliases do not appear as rows on this board. Newer Chinese flagships (K3, GLM-5.3, DeepSeek V4) are also absent. Fetched 2026-10-16. Source: https://raw.githubusercontent.com/SWE-bench/swe-bench.github.io/master/data/leaderboards.json.
Entry% resolvedDetail
Claude 4.5 Opus (high)76.82026-02-17
MiniMax M2.5 (high)75.82026-02-17
Claude 4.6 Opus75.62026-02-17
GLM 5 (high)72.82026-02-17
Lingxi v1.5 x Kimi K271.22025-10-14
Kimi K2.5 (high)70.82026-02-17
DeepSeek V3.2 (high)702026-02-17
OpenHands + Qwen3-Coder-480B-A35B-Instruct69.62025-08-05
GLM-4.668.22025-09-30
OpenHands + Kimi K265.42025-07-16

The newest DeepSeek entry in that extract is V3.2 at 70.0 resolved, dated 17 Feb 2026, against Claude rows from the same mini-agent submission set in the mid-70s. Harnesses differ across the rest of the board. There is no V4 row and no current-Opus row. A claim that “V4 beats Claude at SWE-bench” is not available from this file.

What you give up

You give up the option to pretend the data path is the same. DeepSeek is a PRC company, and several governments have restricted official use. See the safety guide. You also give up a single contract: the app, the API, and OpenRouter are three vendors.

For a lot of drafting and boilerplate, the flash price is the story. For a long autonomous run on a private repo, the missing V4 SWE-bench row and the data-path question matter more than another percent off the token price.

Related: Kimi versus Claude, is the app free?.

Questions

Is DeepSeek cheaper than Claude?

On DeepSeek's published peak rate, V4.1 Flash is much cheaper per token than the latest Opus row on OpenRouter. That is a token price. A long agent trace with retries can spend the savings.

Does DeepSeek beat Claude on SWE-bench?

The Verified leaderboard file we fetched does not contain a DeepSeek V4 row. The newest DeepSeek row in our extract is V3.2. Do not cite a V4 SWE-bench score from this site.

Sources

  1. DeepSeek API pricing, accessed 2026-10-16
  2. OpenRouter models API, accessed 2026-10-16
  3. SWE-bench leaderboards.json, accessed 2026-10-16
  4. LMArena leaderboard dataset, accessed 2026-10-16