Quick verdict: As of August 2026, DeepSeek V4 flash remains the cheapest serious model in the world at $0.14/M input — 3.6x cheaper than OpenAI’s budget model (gpt-5.6-luna) and 7x cheaper than Claude’s budget tier (Haiku 4.5). On price-per-performance, no US lab currently comes close. The catch: DeepSeek has announced a significant price increase is coming, and Anthropic just cut Sonnet 5’s price to $2/M for launch. The gap is real but the landscape is moving fast.
Price is the most under-reported story in AI right now. Every review compares benchmarks and feature lists, but the actual numbers on your invoice are doing something dramatic — and it matters a lot more than a 2-point MMLU difference. I pulled the official API price sheets from OpenAI, Anthropic, and DeepSeek this week so you don’t have to click through three pricing pages. Here’s the real state of the market in August 2026.
The Price Sheet: Everyone’s Current API Rates
All prices are per million tokens, standard tier, short context, from official pricing pages (August 2026). I’ve listed each vendor’s flagship, mid-tier, and budget models:
| Model | Input | Output | Vendor |
|---|---|---|---|
| DeepSeek V4 flash | $0.14 | $0.28 | DeepSeek |
| DeepSeek V4 pro | $0.435 | $0.87 | DeepSeek |
| gpt-5.6-luna | $0.20 | $1.20 | OpenAI |
| gpt-5.6-terra | $2.00 | $12.00 | OpenAI |
| gpt-5.6-sol | $5.00 | $30.00 | OpenAI |
| Claude Haiku 4.5 | $1.00 | $5.00 | Anthropic |
| Claude Sonnet 5 (intro) | $2.00 | $10.00 | Anthropic |
| Claude Sonnet 5 (standard, from Sep 1) | $3.00 | $15.00 | Anthropic |
A few things jump out immediately. First: DeepSeek V4 flash’s output price ($0.28/M) is cheaper than every other model’s input price except gpt-5.6-luna’s. Second: the price spread across “premium” tiers is enormous — gpt-5.6-sol is 35x more expensive than DeepSeek flash on input, 107x on output. Third: Anthropic is clearly trying something with Sonnet 5’s $2 intro price — a price cut at launch is rare in this market, and it’s a direct response to the pressure from Chinese labs.
What the Gap Actually Means in Practice
These numbers are abstract until you put real workloads on them. Let me make it concrete with three realistic scenarios, all calculated at official rates:
Scenario 1: A content pipeline (100 articles/month)
Say each article costs 25K input tokens (research + drafts) and 8K output tokens. Monthly spend:
- DeepSeek V4 flash: 100 × (25K × $0.14/M + 8K × $0.28/M) = $0.57/month
- Claude Haiku 4.5: 100 × (25K × $1.00/M + 8K × $5.00/M) = $6.50/month
- gpt-5.6-terra: 100 × (25K × $2.00/M + 8K × $12.00/M) = $14.60/month
Same work: $0.57 vs $14.60. Over a year, that’s the difference between $7 and $175 — for the same content volume. At scale this is the difference between a hobby and a business.
Scenario 2: A customer-support agent (1M conversations/month)
Each conversation: 1.5K system prompt + 0.5K user + 0.5K output. With DeepSeek’s context caching (cache hits at $0.0028/M), the system prompt gets cheap fast:
- DeepSeek V4 flash + cache: 1M × (1.5K × $0.0028/M + 0.5K × $0.14/M + 0.5K × $0.28/M) ≈ $214/month
- Claude Haiku 4.5 + cache ($0.10/M): 1M × (1.5K × $0.10/M + 0.5K × $1.00/M + 0.5K × $5.00/M) ≈ $3,150/month
- gpt-5.6-luna + cache ($0.02/M): 1M × (1.5K × $0.02/M + 0.5K × $0.20/M + 0.5K × $1.20/M) ≈ $730/month
This is where the gap becomes existential. A 12x difference in monthly infrastructure cost determines whether your startup burns or survives.
Context Caching: The Hidden Variable
Every vendor has prompt caching now, but the economics differ wildly:
| Vendor | Cache-hit input price | vs. normal input |
|---|---|---|
| DeepSeek V4 | $0.0028/M | 50x cheaper |
| OpenAI (gpt-5.6) | $0.02–0.50/M | ~10x cheaper |
| Claude (Sonnet 5) | $0.20/M | 10x cheaper |
DeepSeek’s cache discount is the most aggressive in the industry — I verified it myself: a 6,500-token system prompt hit the cache for 6,400 tokens on the second call (98.5% hit rate). For agent loops and chat apps with stable system prompts, this is the single biggest lever on your API bill.
But It’s Not Just Price: What Each Vendor Wins On
Price isn’t everything, and pretending otherwise would be dishonest. Here’s my honest take on where each vendor genuinely wins:
- OpenAI (gpt-5.6): ecosystem, tooling, reliability, and the best multimodal support. If you need image/audio/video in one API, it’s still the default. You pay a massive premium for that convenience.
- Anthropic (Claude): long-context quality, coding, and agentic reliability. Claude remains the coding favorite among developers I know — Sonnet 5’s intro pricing is a real bargain for what it offers at $2/M.
- DeepSeek V4: raw price-performance. My benchmark testing showed it holds its own on reasoning and coding at a fraction of the price — and the API is genuinely developer-ready (JSON output, tool calls, Anthropic-compatible endpoint).
The Fine Print Everyone’s Ignoring
Three things to watch before you bet your architecture on today’s numbers:
- DeepSeek’s announced price increase. It’s official, in writing: “a significant increase expected.” Nobody knows the new numbers, but the 50x cache discount and the $0.14/M input rate are both at risk. Lock in savings while you can — but build for portability.
- Sonnet 5’s intro price expires September 1, 2026. $2/M becomes $3/M. Still cheap relative to the market, but budget accordingly.
- Context length pricing. OpenAI charges roughly 2x for long-context windows (gpt-5.6 long context: sol $10/$45 vs $5/$30 short). DeepSeek’s 1M window is flat-rate. If you work with big documents, this flips the comparison even further in DeepSeek’s favor.
My Recommendation
For most builders in August 2026: default to DeepSeek V4 flash, keep a warm standby on OpenAI or Claude. Build an abstraction layer (one base-URL swap), put the heavy lifting on DeepSeek, and use the premium models where they genuinely win — multimodal, complex agentic workflows, or enterprise compliance requirements. If you’re already on Claude, Sonnet 5’s intro pricing is the best reason in a long time to stay. But for new projects with real volume, the DeepSeek price advantage is currently decisive — and the window to exploit it may be closing.
FAQ
Is DeepSeek V4 really the cheapest AI API? As of August 2026, yes for serious models: $0.14/M input and $0.28/M output (flash tier) beats every comparable US model, with Claude Haiku 4.5 at $1/$5 and gpt-5.6-luna at $0.20/$1.20.
How much does the OpenAI API cost in 2026? gpt-5.6-luna starts at $0.20/M input and $1.20/M output; the flagship gpt-5.6-sol is $5/$30. Long-context windows cost roughly double.
Is Claude cheaper than DeepSeek? No — Claude Haiku 4.5 is $1/$5, and Sonnet 5 is $2/$10 (intro). DeepSeek V4 flash at $0.14/$0.28 is 7x cheaper than Claude’s budget tier.
Are AI API prices going down or up? Mixed: Anthropic cut Sonnet 5’s launch price to $2/M, but DeepSeek has officially announced a significant increase is coming. Lock in current pricing and build portability.
All prices pulled from official OpenAI, Anthropic, and DeepSeek pricing pages on August 10, 2026. Prices change often — verify before committing. I paid for every API test referenced here. This post contains affiliate links.