DeepSeek V4 API, Pricing, Benchmarks & Guides
Follow DeepSeek V4 news, compare it with ChatGPT, Claude, Gemini, and Grok, read Reasonix and local deployment guides, and jump to in-stock one-off Coding Plans when you need configured API access.
- Models
- deepseek-v4-pro · deepseek-flash
- Context
- 1M tokens
- API
- OpenAI-compatible
- Off-peak from
- $0.003 / 1M cache-hit input
DeepSeek Chat, Kimi Coding Plan, DeepInfer and Developer Guides
Evidence-first benchmark update
DeepSeek Flash-1017 and V4 Pro: benchmarks and model comparison
Compare current DeepSeek models with selected frontier models using independent intelligence scores and DeepSeek's separately reported agent benchmarks.
Current DeepSeek API models
- Flash-1017
- deepseek-flash · Flash-1017 · Text and image input
- V4 Pro-0813
- deepseek-v4-pro · Pro-0813 · Text input
DeepSeek V4.1 Flash-1017: October 17 release benchmarks
DeepSeek reports gains across reasoning, coding, agent tasks, and vision. This release table compares Flash-1017 with the September Flash release and four other models; the scores are publisher-reported, not an independent evaluation. The API model ID remains deepseek-flash and pricing is unchanged.
Terminal-Bench 4.0
October 17 release comparison using the Claude Code harness. Flash-1017 scores 42.3, versus 31.2 for the September Flash release.
| Model | Reported score |
|---|---|
| DeepSeek-V4.1-Flash-1017 | 42.3 |
| DeepSeek-V4.1-Flash (2026-09-10) | 31.2 |
| Kimi K3 | 12.6 |
| GLM-5.3 | 37.9 |
| GPT-5.6 Sol | 39.9 |
| Claude Opus 5.0 | 51.8 |
Six models across 11 benchmarks
Compare scores within each benchmark. HLE bars show the full test without tools; text-only results (†) appear separately in the table. ZeroBench-main uses tools and Pass@5. Missing results are left blank.
Reported score · 0–100
View benchmark data as a table
| Benchmark | DeepSeek-V4.1-Flash-1017 | DeepSeek-V4.1-Flash (2026-09-10) | Kimi K3 | GLM-5.3 | GPT-5.6 Sol | Claude Opus 5.0 |
|---|---|---|---|---|---|---|
| HLE (w/o tools) | 42.6 (44.0†) | 36.8 (39.1†) | 43.5 | Not reported (42.0†) | 44.5 | 56.3 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 12.6 | 37.9 | 39.9 | 51.8 |
| DeepSWE v1.1 | 77.0 | 74.2 | 67.5 | 66.9 | 73.0 | 74.0 |
| ProgramBench | 29.1 | 20.3 | 17.5 | 19.0 | 23.0 | 37.0 |
| NL2Repo-Bench | 68.2 | 65.4 | 58.0 | 58.0 | 56.8 | 75.3 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 46.7 | 48.8 | 45.8 | 50.3 |
| Agents’ Last Exam | 35.4 | 31.8 | 27.6 | 28.5 | 26.7 | 28.6 |
| Apex Bench v1.1 | 72.1 | 64.6 | 47.8 | 56.6 | 54.3 | 65.8 |
| Chartography (w/tools) | 83.0 | 78.9 | 68.1 | Not reported | 79.9 | 84.0 |
| BabyVision (w/tools) | 92.5 | 89.6 | 85.7 | Not reported | 88.9 | 94.1 |
| ZeroBench-main (w/tools, Pass@5) | 51.0 | 49.0 | 41.0 | Not reported | 53.0 | 52.0 |
† HLE text-only subset, not the full test. GLM-5.3 has only a text-only HLE result in this source. Terminal-Bench 4.0 uses Claude Code; DeepSWE v1.1 uses Mini-SWE-V2. Chartography, BabyVision, and ZeroBench-main use tools; ZeroBench-main is Pass@5. Automation-Bench v1.0.6 and Apex Bench v1.1 retain their published versions. No new Pro-0813 or Terminal-Bench 2.1 result is supplied in this release table.
Source: DeepSeek Flash-1017 model card · October 17, 2026 · Official model card and evaluation settings · Official changelog
Independent comparison: one shared methodology
Artificial Analysis Intelligence Index v4.3.2 · verified October 16, 2026. Thirteen selected model variants with their evaluated effort and fallback settings. Index scores are not percentages. The Flash score is for the September release, not Flash-1017.
Intelligence Index
DeepSeek V4.1 Flash (2026-09-10)
Max effort
Source39DeepSeek V4 Pro-0813
Max effort
Source36Claude Opus 5.5
Max effort · with fallback
Source58Claude Sonnet 5.5
Max effort · with fallback
Source56Claude Fable 5.1
Max effort · with fallback
Source53GPT-6 Astra
Max effort
Source53GPT-6.1 Sol
Max effort
Source52Gemini 4 Argon
High effort
Source53Grok 4.7
Xhigh effort
Source46Qwen3.8 Max (0902)
Effort not specified in leaderboard
Source45GLM-5.3
Max effort
Source45Kimi K3
Max effort
Source44MiniMax-M3
Effort not specified in leaderboard
Source29
Read scores with their methodology and model variant
These v4.3.2 scores cannot be compared with the earlier v4.1 scores. Effort levels across providers do not imply equal compute budgets. Claude results labeled with fallback include that fallback.
AA methodologyDeepSeek V4.1 Flash and V4 Pro API token pricing
Official rates per 1M tokens, verified 2026-10-11. Peak hours are 09:00–12:00 and 14:00–18:00 Beijing time (01:00–04:00 and 06:00–10:00 UTC), Monday–Friday, excluding Chinese public holidays. All other hours, weekends, and Chinese public holidays are off-peak. These rates are separate from Coding Plans. Flash pricing is for DeepSeek-V4.1-Flash; the dated benchmark results above remain historical measurements.
| Model | Cache hit input | Cache miss input | Output | Pricing window |
|---|---|---|---|---|
| DeepSeek V4.1 Flash | $0.003 / ¥0.02 | $0.15 / ¥1.00 | $0.6 / ¥4.00 | Off-peak |
| DeepSeek V4.1 Flash | $0.006 / ¥0.04 | $0.3 / ¥2.00 | $1.2 / ¥8.00 | Peak |
| DeepSeek V4 Pro | $0.022 / ¥0.15 | $0.66 / ¥4.50 | $1.98 / ¥13.50 | Off-peak |
| DeepSeek V4 Pro | $0.044 / ¥0.30 | $1.32 / ¥9.00 | $3.96 / ¥27.00 | Peak |
Official release benchmarks: October 17, 2026 · AA snapshot: October 16, 2026
The API serves Flash-1017 via deepseek-flash; Pro-0813 remains available. September Flash scores are historical measurements. Benchmark coverage remains separate from Coding Plan inventory and availability.
Four Coding Plans · One-time or monthly
DeepSeek Coding Plan leads the current lineup
Get 50M tokens every 5 hours and 400M tokens every 7 days across DeepSeek V4.1 Flash and the existing V4 routes. New configurations use the native multimodal Model ID deepseek-flash. Also in stock while inventory lasts: Agent Plan, Coding Plan Plus, and Coding Plan Max — each with one-time purchase or 50%-off monthly renewal.