DeepSeek V4 Pro 0423
DeepSeek released 2026-04-24
DeepSeek V4 Flash 0731 is both better and cheaper.
It scores +6.5 higher and costs 79% less ($0.105/M against $0.496/M) on this workload — and it does everything this model does.
GPT-5.6 Luna is cheaper still (9% less) but drops 384K → 128K max output.
Our take
editorial — not a measurementThe most unusual pricing model in the dataset: DeepSeek bills by wall-clock time. Peak rate is $1.32/$3.96, and everything outside 01:00-04:00 and 06:00-10:00 UTC Monday-Friday is exactly half that. A batch job that runs at 14:00 UTC costs half what the identical job costs at 08:00, which no other vendor here does and which most cost models cannot express. Combined with a cache-hit rate of $0.044 — about 3% of the cache-miss price, the most aggressive caching discount in the dataset — a well-scheduled, well-cached DeepSeek pipeline is extraordinarily cheap. Open weights on Hugging Face, though the exact licence was not verified.
Strengths
- 50% off outside peak hours — schedulable savings no competitor offers
- Cache hits at ~3% of cache-miss price
- 1M context window, open weights available
- Very cheap at the frontier-adjacent tier
Weaknesses
- Price varies by time of day — genuinely hard to budget
- Licence not independently verified despite open weights
- OpenRouter lists a blended $1.122/$3.366 that matches neither published rate
- Text-only
Reach for it when
- Schedulable batch pipelines that can run off-peak
- Cache-heavy repeated-context workloads
- Self-hosting where licence permits
Avoid it if
- You need predictable per-request cost
- Your traffic is latency-bound to business hours in the peak window
Sources: api-docs.deepseek.comopenrouter.ai
Every price dimension
| Input | $0.397/M |
|---|---|
| Output | $0.794/M |
| Cached input | $0.033/M |
Price varies by time of day. Listed rate is the PEAK rate. Off-peak is 50% lower ($0.66 input / $0.022 cache hit / $1.98 output). DeepSeek defines peak hours as 01:00-04:00 and 06:00-10:00 UTC, Monday-Friday. OpenRouter lists a blended $1.122/$3.366.
Independent scores
| Intelligence | 45.3 |
|---|---|
| Coding | 59.4 |
| Agentic | 37.8 |
| LMArena Elo | 1450.9default |
Sources are listed separately rather than averaged. Across the 62 models both have scored they correlate at r = 0.806 — close agreement overall, but four models rank very differently between them, and a blended score would hide exactly those.
Best rankings by task
- 3d #16of 136 1301
- godotgamedev #27of 43 1059
- gamedev #28of 144 1269
- asciiart #30of 83 1178
- codecategories #31of 145 1267
Capability
| Context window | 1.0M |
|---|---|
| Max output | 384K |
| Input modes | text |
| Tool use | yes |
| Reasoning | optional |
| Open weights | yes |
Provenance
| Price source | openrouter.ai |
|---|---|
| Fetched | 2026-08-24 |
| Quality data | verified |
| Cross-checked | vendor page |
DeepSeek V4 Pro is a large-scale Mixture-of-Experts model from DeepSeek with 1.6T total parameters and 49B activated parameters, supporting a 1M-token context window. It is designed for advanced reasoning, coding,...