DeepSeek
DeepSeek V3.2
6.3
out of 10
DeepSeek's latest model continues to shock with its price-to-performance ratio. V3.2 introduces 'Fine-Grained Sparse Attention' for 50% better compute efficiency. Input costs drop to $0.07/1M tokens with cache hits. The web interface at chat.deepseek.com appears to be free with no hard usage cap.
Context window
128K tokens
API (blended)
$0.48/1M
Consumer access
Free
Multimodal
Text only
Strengths
- +Cheapest frontier-capable API: $0.48/1M blended (drops to ~$0.13 with caching)
- +Open-weight — downloadable and self-hostable
- +Web interface at chat.deepseek.com appears free with no hard cap
- +50% better compute efficiency vs prior generation
Weaknesses
- -Chinese company — data stored under Chinese law; avoid for sensitive work
- -Smallest context window of the group at 128K tokens
- -Full 685B parameter model is difficult to self-host at scale
- -Service reliability has had outage issues during high demand
Best for
budget API usereasoningcodinghigh-volume processingself-hosting
Not ideal for
privacy-sensitive datacreative writinglong documentsenterprise workloads
Pricing details
Subscription plans
Free (chat.deepseek.com)Full web chat access, no announced hard usage cap(Service reliability varies; has experienced outages. Data subject to Chinese law.)
FreeAPI pricing
DeepSeekCache hit input: $0.07/1M (74% discount). Batch API: 50% discount. Cheapest frontier-capable direct API available.
$0.27/$1.1OpenRouterSmall markup. Useful for unified API access across providers.
$0.28/$1.12Together AIfree tier$25 free credits on signup.
$0.3/$1.2Self-hostedOpen-weight — download from HuggingFace. Full 685B param model requires significant multi-GPU infrastructure. Smaller distilled versions available.
Self-hostedPrices verified February 2026. LLM pricing changes frequently — verify at the provider's site before budgeting.
Last updated: February 2026