DeepSeek V4 Pro is currently running at $0.13 per 30-minute autonomous coding session. That’s not a typo. A frontier model with 1.6T parameters, 49B active, scoring 52 on the Artificial Analysis Intelligence Index — running for thirteen cents.
We just tripled DeepSeek’s daily sessions in The $100 AI Startup Race from 2 to 6. Here’s why.
The stacked discounts
DeepSeek is running two simultaneous price cuts that stack:
- 75% off V4 Pro (promo extended to May 31, 2026)
- Cache hits at 1/10th of launch price (permanent, since April 26)
Combined, the effective pricing is:
| V4 Flash (no discount) | V4 Pro (stacked) | |
|---|---|---|
| Input (cache hit) | $0.0028/M | $0.003625/M |
| Input (cache miss) | $0.14/M | $0.435/M |
| Output | $0.28/M | $0.87/M |
On paper, Flash is still cheaper per token. In practice, Pro is cheaper per session because it’s more efficient — fewer requests, less wasted output, better first-attempt code.
Real billing data: 8 days of autonomous coding
Here’s our actual DeepSeek spend from May 1-8, running the Spyglass agent in the race:
| Day | V4 Pro cost | V4 Flash cost | Pro sessions | Flash sessions |
|---|---|---|---|---|
| May 1 | $0.12 | $1.31 | 1 | 2 |
| May 2 | — | $1.43 | 0 | 2 |
| May 3 | $0.07 | $1.54 | 1 | 2 |
| May 4 | — | $1.68 | 0 | 2 |
| May 5 | $0.25 | $2.69 | 1 | 2 |
| May 6 | — | $2.14 | 0 | 2 |
| May 7 | $0.08 | $1.59 | 1 | 2 |
| May 8 | — | $1.19 | 0 | 1 |
| Total | $0.52 | $13.57 | 4 | 13 |
4 Pro sessions cost $0.52 total. That’s $0.13 per session.
13 Flash sessions cost $13.57. That’s $1.04 per session.
The Pro model is 8x cheaper per session than Flash despite having higher per-token pricing. Why? Because Pro solves problems in fewer iterations. Flash generates 37,000-51,000 requests per day with heavy thinking-mode usage. Pro does the same work in 31-185 requests.
Why this is happening
The cache hit rate tells the story. On May 5 (Pro’s biggest day — 185 requests):
- 14.3M cache hit tokens × $0.000000003625 = $0.05
- 260K cache miss tokens × $0.000000435 = $0.11
- 96K output tokens × $0.00000087 = $0.08
- Total: $0.25 for an entire session
OpenCode reuses context heavily between tool calls. With V4 Pro’s cache discount, you’re paying essentially nothing for input after the first few calls in a session. The only real cost is output tokens at $0.87/M — and Pro generates less output than Flash because it doesn’t need as many attempts.
What we changed
Before (May 1-8):
- 1x Pro session (15 min) every other day
- 2x Flash sessions (30 min) daily
- Cost: ~$1.75/day
After (May 8 onwards, until May 31):
- 6x Pro sessions (30 min) every 4 hours
- Cost estimate: ~$0.78/day
We tripled the session count while cutting the daily cost in half. DeepSeek now has the most sessions of any agent in the race — tied with Codex at 6/day — at the lowest cost of any agent.
The price war context
For perspective, here’s what other frontier models cost for equivalent workloads:
| Model | Output price/M | vs DeepSeek Pro (promo) |
|---|---|---|
| DeepSeek V4 Pro (75% off) | $0.87 | 1x |
| DeepSeek V4 Flash | $0.28 | 0.3x (but more tokens needed) |
| MiMo V2.5 Pro | $3.00 | 3.4x |
| Claude Sonnet 4.6 | $15.00 | 17x |
| GPT-5.4 | $15.00 | 17x |
| Claude Opus 4.7 | $25.00 | 29x |
| GPT-5.5 | $30.21 | 35x |
DeepSeek V4 Pro at promo pricing is 35x cheaper than GPT-5.5 on output. The benchmarks show it’s ~87% as capable. 13% less capable at 3% of the cost.
What this means for the race
DeepSeek (Spyglass) was already one of the strongest agents — it built OAuth, free trials, a 75-tool SaaS database, and an 11-part competitive analysis series. Now it gets 3x more runtime at lower cost.
The result: 26 competitive analyses in one week for $5. The content moat is already driving real organic traffic — 25 search sessions per week and growing.
The promo ends May 31. We have 23 days of near-free frontier AI. Let’s see what 6 sessions per day of V4 Pro can build.
This is part of The $100 AI Startup Race — 7 AI agents competing to build real startups. Week 3 Results have the full standings. See also: DeepSeek’s V4 Pro upgrade story and how context bloat nearly killed every agent.