DeepSeek V4-Flash
Language Retired Full comparison ↗ deepseek/deepseek-v4-flash · by DeepSeek
· mixture-of-experts
Pricing — 1 offering(s)
Input tokens
Retired — no longer sold · last price $0.44 / 1M tokens (input), until 2026-09-09
- $0.14 / 1M tokens (input) Historical 2026-08-04 → 2026-08-15
- $0.44 / 1M tokens (input) Historical 2026-08-16 → 2026-09-09
Output tokens
Retired — no longer sold · last price $1.32 / 1M tokens (output), until 2026-09-09
- $0.28 / 1M tokens (output)
Historical
2026-08-04 → 2026-08-15
Same sourcing basis as the input tier.
- $1.32 / 1M tokens (output) Historical 2026-08-16 → 2026-09-09
Off-peak input tokens
Retired — no longer sold · last price $0.22 / 1M tokens (input), until 2026-09-09
- $0.22 / 1M tokens (input) Historical 2026-08-16 → 2026-09-09
Off-peak output tokens
Retired — no longer sold · last price $0.66 / 1M tokens (output), until 2026-09-09
- $0.66 / 1M tokens (output) Historical 2026-08-16 → 2026-09-09
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how DeepSeek V4-Flash fits into a cost-aware routing setup
See how →Capability profile
How DeepSeek V4-Flash rates across core capability dimensions, with the task-level evidence behind each rating.
1,000,000-token default context — the headline of the V4 generation and the subject of its tech report.
Thinking mode available. Independent GA coverage describes "production-tuned agent, math, and coding performance" as the headline improvement over the preview build; DeepSeek did not publish a Flash-specific benchmark table cross-checked here (the widely-cited V4 benchmark figures are for the larger V4-Pro / V4-Pro-Max configs).
`supports_coding: true` in the registry; positioned for agentic coding via the Responses API. No Flash-specific coding-benchmark score verified.
The model backing DeepSeek's agent-oriented Responses API; integrates with agent frameworks (Claude Code, OpenCode) per the release notes.
Strong Chinese and English; the DeepSeek line is not especially broad beyond those.
~13B active parameters and the compressed-attention design make it fast and cheap at long context — the point of the 'Flash' tier.
$0.14 / 1M cache-miss input, $0.28 / 1M output (DeepSeek's own pricing page; a 50x cache-hit discount exists but isn't modelled here). Very cheap for a 1M-context model, and MIT open weights for self-hosting. An upcoming peak/off-peak pricing policy is flagged in DeepSeek's docs but not yet effective.
Ratings are estimated — limited independent data is available for this model.
Operator guidance
The default V4 choice for cost- and latency-sensitive long-context and agentic work. Step up to V4-Pro when reasoning/coding ceiling matters more than cost. Against GLM-5.2 (the other MIT Chinese-lab flagship added in the same registry batch), V4-Flash is cheaper with the same 1M context; GLM-5.2 has more (vendor-reported) coding-benchmark signal.
Use cases
- Very-long-context tasks (whole codebases, large document sets) at low cost — the V4 generation's core pitch
- Agentic coding and tool-use workflows via the Responses API
- High-volume production workloads where the cache-hit discount and low token price compound
- Self-hosted deployment under MIT for data-control requirements
Limitations
- DeepSeek did not publish a V4-Flash-specific benchmark table; capability ratings lean on independent coverage and the V4 tech report's general claims, not verified Flash-tier numbers
- Architecture details (active-parameter count, Compressed Sparse Attention) are from independent technical coverage, not DeepSeek's own docs
- GA-date and snapshot naming vary across sources by a few weeks (V4-Flash-0731 etc.)