fast hybrid-attention reasoning.
DeepSeek V4 Flash beside its stablemates and nearest rivals. The ◆ marks the best value in each column across every row shown.
| Model | Intelligence | Coding | Speed | Input $/M | Output $/M | Blended $/M | Context |
|---|---|---|---|---|---|---|---|
| DeepSeek V4 Pro | 44.3◆ | 59.4◆ | 70.9◆ | $0.43◆ | $0.87◆ | $0.54◆ | 1M◆ |
| Tencent Hy3 | 41.2◆ | 58.8◆ | 65.2◆ | $0.14◆ | $0.58◆ | $0.25◆ | 262K◆ |
| Inkling | 40.7◆ | 52.1◆ | 56.5◆ | $1◆ | $4.05◆ | $1.76◆ | 256K◆ |
| DeepSeek V4 Flash ◆ | 40.3◆ | 56.2◆ | 122◆ | $0.14◆ | $0.28◆ | $0.18◆ | 1M◆ |
| GLM-5.1 | 40.2◆ | 55.8◆ | 68.7◆ | $1.40◆ | $4.40◆ | $2.15◆ | 200K◆ |
| Step 3.7 Flash | 30.3◆ | 39.6◆ | 399.5◆ | $0.20◆ | $1.15◆ | $0.44◆ | 256K◆ |
| Kimi K3pinned | 57.1◆ | 76.2◆ | 34.5◆ | $3◆ | $15◆ | $6◆ | 1M◆ |
| Kimi K2.7 Codepinned | 41.9◆ | 60.8◆ | 44.4◆ | $0.95◆ | $4◆ | $1.71◆ | 256K◆ |
The Intelligence Index and its sub-scores, ranked against every scored model in the catalog. A metric that has not been measured for DeepSeek V4 Flash has been left empty.
How far a month of credits goes on DeepSeek V4 Flash.
Real coding tasks priced end to end on DeepSeek V4 Flash, from a quick lookup to a full-repo agent run.
| Task | Tokens in · out | DeepSeek V4 Flash | Claude Haiku 4.5 |
|---|---|---|---|
| Quick lookup / one-liner | 8K · 1K | $0.0004 | $0.0065 |
| Review a 500-line PR | 60K · 4K | $0.0038 | $0.04 |
| Fix a bug (agent loop) | 180K · 12K | $0.01 | $0.12 |
| Refactor a module | 320K · 20K | $0.02 | $0.21 |
| Full-repo agent run | 900K · 45K | $0.04 | $0.49 |
$0.14/M input and $0.28/M output, cache reads $0.00/M. In an agent loop most input is cache-read, so the effective input rate is about $0.04/M.cmd --model deepseek/deepseek-v4-flash, or type /model in a session and pick it. You can switch mid-session without losing context.Command Code is the AI coding agent that continuously learns your taste. Start for $1.