As of Aug 2026, DeepSeek's pay-as-you-go API is still the cheapest mainstream option per million tokens among major LLM providers, according to pricing comparisons by CostGoat and Fello AI: the official V4 Flash tier costs roughly $0.14 per 1M input tokens and $0.28 per 1M output tokens on cache misses, which is 97.2% cheaper on input than OpenAI GPT-5.5's $5 rate and 99.1% cheaper on output than GPT-5.5's $30 rate. That works out to roughly 36× cheaper on input ($5 ÷ $0.14 ≈ 35.7) and 107× cheaper on output ($30 ÷ $0.28 ≈ 107.1) than GPT-5.6 Sol / Claude Opus 5-class flagships. If your goal is to minimize per-token spend without sacrificing a 1M-token context window, DeepSeek V4 Flash — available on HeFu under the same usage-based billing model — is the first option to evaluate.
Current DeepSeek API Pay-As-You-Go Rates (As of Aug 2026)
According to pricing reviews published in Jul 2026, the official DeepSeek API is purely pay-as-you-go, with no subscription required (Coworker.ai, Jul 2026; official price page: platform.deepseek.com/pricing):
- DeepSeek V4 Flash: approximately $0.14 per 1M input tokens and $0.28 per 1M output tokens (cache miss); cached input is only about $0.0028 per 1M tokens — roughly 1/50th of the standard input price, i.e. a 98% discount (Coworker.ai, Jul 2026; PricePerToken; CloudZero).
- DeepSeek V4 Pro: standard list price is $1.74 per 1M input and $3.48 per 1M output; since Jul 2026, an approximately 75% promotional discount brings the effective rate to roughly $0.435 / $0.87 per million tokens (Fello AI, 2026; CloudZero).
- Free trial tokens: new accounts receive 5 million free tokens, valid for about 30 days, with no credit card required. There is no permanent free tier (Coworker.ai, Jul 2026; PricePerToken).
One important caveat: GeoToolbox reported on Aug 16, 2026 that DeepSeek switched to peak/valley pricing (V4 Flash valley from $0.22/$0.66, peak around $0.44/$1.32), which contradicts the flat Jul 2026 rates. However, another verification dated Jul 21, 2026 found no standing valley discount table on the official page (Layer3Labs). Because pricing has been moving quickly, always confirm current rates on the official DeepSeek price page (or the HeFu pricing page if you buy through an aggregator) before committing.
How Pay-As-You-Go Billing Actually Works for DeepSeek API
DeepSeek's pay-as-you-go model is simple: you prepay a balance, each API request deducts the metered token cost, and there are no idle fees, no monthly base fees, and no minimum usage commitment. A few mechanics worth understanding:
- Token metering: both input and output tokens are counted per request; system prompts, retrieved context, and tool definitions all count as input.
- Cache-hit discount: when your request includes a prefix that matches a cached prompt, the input portion drops to about $0.0028 per 1M tokens — a 98% discount from the $0.14 miss rate, not merely one-tenth (Coworker.ai, Jul 2026). DeepSeek's official docs describe the caching mechanics at api-docs.deepseek.com.
- Specs: V4 Flash and V4 Pro both support a 1M-token context window and up to 384K output tokens, with published concurrency limits of 2,500 requests (Flash) and 500 requests (Pro) (Coworker.ai, Jul 2026).
- No card required for trial: the 5M free tokens for new accounts are a genuine zero-cost test window, which makes pay-as-you-go evaluation of DeepSeek cheaper than any competitor.
DeepSeek vs. GPT-5.5 vs. Claude Opus 5: Cost Comparison Table
The table below compares published per-million-token rates and specs. DeepSeek's promotional pricing is marked with its validity window; competitor prices are official list rates as reported in Aug 2026.
| Model (as of Jul–Aug 2026) | Input per 1M tokens | Output per 1M tokens | Cache-hit input | Context window | Source |
|---|---|---|---|---|---|
| DeepSeek V4 Flash | ~$0.14 | ~$0.28 (cache miss) | ~$0.0028 (≈98% discount) | 1M tokens | Coworker.ai, Jul 2026 |
| DeepSeek V4 Pro (promo) | ~$0.435 (list $1.74) | ~$0.87 (list $3.48) | — | 1M tokens | Fello AI, 2026 |
| OpenAI GPT-5.5 | $5 | $30 | — | — | CostGoat, Aug 2026 |
| GPT-5.6 Sol / Claude Opus 5 (flagship class) | ≈36× V4 Flash input rate | ≈107× V4 Flash output rate | — | — | CostGoat, Aug 2026; Fello AI, 2026 |
The ratio-based figures for the flagship class are derived from comparisons published by CostGoat (Aug 2026) and Fello AI (2026): $0.14 × 36 ≈ $5.04 for input, and $0.28 × 107 ≈ $29.96 for output, placing GPT-5.6 Sol and Claude Opus 5-class models in the $5–$30 per-million range. For context, our separate guides cover GPT-5.6 API pricing and Claude Opus 5 API access in more depth.
Real-World Cost Scenarios: 4 Typical Workloads
Monthly cost estimates at official DeepSeek V4 Flash rates (as of Jul–Aug 2026). These are illustrative — your actual bill depends on caching behavior and token mix.
| Workload | Volume | Monthly cost (V4 Flash) |
|---|---|---|
| Customer-service chatbot | 500 users × 30 messages = 15M input |
Top comments (0)