DeepSeek vs Claude: The 2026 Price and Capability Gap
The short answer
DeepSeek V4-Flash costs $0.14 per million input tokens. Claude Sonnet 5 costs $2.00. On output the gap widens to $0.28 against $10.00 — roughly 97% cheaper.
But the context-window argument that every older comparison makes is dead. Both are at 1M tokens now. If you read somewhere that DeepSeek is limited to 128K and Claude to 200K, that page is describing 2025.
Last verified 25 July 2026 against DeepSeek's API documentation and Anthropic's model and pricing documentation. We re-check this page weekly.
Current models
DeepSeek
| Model | Total params | Active | Context | Max output |
|---|---|---|---|---|
| deepseek-v4-flash | 284B | ~13B | 1M | 384K |
| deepseek-v4-pro | 1.6T | ~49B | 1M | 384K |
Text-only, MoE, MIT licence, open weights. Both offer thinking and non-thinking modes. The legacy deepseek-chat and deepseek-reasoner aliases were retired on 24 July 2026 — see the migration guide. More on the architecture in DeepSeek V4 explained.
Anthropic
| Model | Context | Max output | Thinking |
|---|---|---|---|
| Claude Fable 5 | 1M | 128K | Adaptive, always on |
| Claude Opus 5 | 1M | 128K | Adaptive |
| Claude Sonnet 5 | 1M | 128K | Adaptive |
| Claude Haiku 4.5 | 200K | 64K | Extended |
All current Claude models take text and image input and produce text output. That is the first real capability difference: Claude sees images, DeepSeek does not.
API pricing
Per 1M tokens:
| Model | Input | Cache read | Output |
|---|---|---|---|
| DeepSeek V4-Flash | $0.14 | $0.0028 | $0.28 |
| DeepSeek V4-Pro | $0.435 | $0.003625 | $0.87 |
| Claude Haiku 4.5 | $1.00 | $0.10 | $5.00 |
| Claude Sonnet 5 | $2.00 | $0.20 | $10.00 |
| Claude Opus 5 | $5.00 | $0.50 | $25.00 |
| Claude Fable 5 | $10.00 | $1.00 | $50.00 |
One thing to diarise: Claude Sonnet 5's price rises on 1 September 2026. The $2.00/$10.00 rate is introductory pricing running through 31 August; from 1 September the standard rate of $3.00 input / $15.00 output applies. Anthropic states this on its own pricing page. If you are modelling costs for Q4, use the September numbers.
The multiples
| Comparison | Input | Output |
|---|---|---|
| Sonnet 5 vs V4-Flash | 14.3× more | 35.7× more |
| Opus 5 vs V4-Pro | 11.5× more | 28.7× more |
| Fable 5 vs V4-Pro | 23.0× more | 57.5× more |
Put as a percentage, since that is how most comparisons phrase it: against Sonnet 5, DeepSeek V4-Flash is about 93% cheaper on input and 97% cheaper on output. Comparisons claiming "80–90% cheaper" are understating it. Our DeepSeek pricing page has the full rate card and a cost calculator.
Consumer plans
| Claude | Price |
|---|---|
| Free | $0 |
| Pro | $20/mo, or $17/mo billed annually ($200 up front) |
| Max | from $100/mo — 5× or 20× Pro usage, higher output limits |
| Team | $20/seat/mo annually, $25 monthly; premium seats $100/$125 |
| Enterprise | $20/seat plus usage |
Pro and above include Claude Code and Claude Cowork. DeepSeek's consumer chat is free with an account and has no paid tier.
Where each one wins
DeepSeek when cost per token dominates, when your workload is text at volume, when you want open weights you can run yourself, or when you need 1M context at a flat rate.
Claude when you need image input, when you want adaptive thinking that decides its own reasoning depth, when you are doing agentic or coding work with Claude Code, or when your organisation needs the enterprise controls Anthropic ships.
The models are closer than the price gap suggests on plain text generation, and further apart than it suggests on anything agentic or visual.
What we do not claim
We do not publish a star rating, a benchmark table, or a language count for either model.
On benchmarks: the scores circulating for these models mostly cannot be traced to the vendor or to a named leaderboard. Artificial Analysis publishes dated, documented comparisons — use those. Anthropic publishes evaluations in its own release posts.
On languages: you will see "Claude supports 70+ languages" and "DeepSeek supports 50+" on comparison pages across the web. Neither company publishes a total. Anthropic's multilingual documentation benchmarks 15 languages and says Claude is capable in many more; DeepSeek publishes no count at all. Both numbers are invented.
On speed: we have not published response-time figures because we have no methodology we would defend.