senn-techDeepSeek V4-Flash: Frontier Performance at One Percent of the Cost
In August 2026, DeepSeek made a quiet move: it retrained V4-Flash on the same base model and pushed its Terminal Bench 2.1 score to 82.7 — ahead of Anthropic's Claude Fable 5 at 80.5. The price is roughly one percent of what Claude Fable 5 charges.
What changed — and what didn't
- Retrained 2026-08-01: DeepSeek sharpened V4-Flash with fresh data and an improved post-training pipeline without renaming the model.
- Same architecture: 284B parameters, around 13B active per token, 1M token context window.
- Sparse MoE design: lots of capacity in the background, low compute per forward pass — the technical reason the price stays low.
- No new API version: the endpoint stays the same for developers, making the upgrade behaviorally stable despite the score jump.
The numbers compared
| Model | Terminal-Bench 2.1 | Input / 1M | Output / 1M |
|---|---|---|---|
| DeepSeek V4-Flash | 82.7 | $0.14 | $0.28 |
| Claude Fable 5 | 80.5 | $10.00 | $50.00 |
| GPT-5.6 Sol | 85.8 | $5.00 | $30.00 |
| Claude Sonnet 5 | 74.5 | $3.00 | $15.00 |
The relevant comparison is the output price, not the input. Agentic workflows, tool use, and longer code completions typically generate far more output than input. At output pricing, Claude Fable 5 is about 178× more expensive than V4-Flash.
Why the price stays so low
DeepSeek continues to rely on a MoE architecture with a small active share and its own infrastructure optimizations. Only a fraction of the 284B parameters is computed per token. That keeps inference cost low even though the total model is large. Combined with Chinese hosting economics and aggressive pricing, you get a model that covers the bulk of production tasks and feels close to free.
The catch
The Terminal Bench score is impressive, but it does not fully predict your own use case. DeepSeek models can sometimes produce longer answers, creative over-engineering, or tool-parameter hallucinations. If you use Claude Fable 5 for complex reasoning, do not switch blindly; benchmark both on your own prompt set. And the usual GDPR caveat applies: sending data to DeepSeek's API is not straightforward for personal or confidential content.
Our take
V4-Flash is currently the clearest demonstration that frontier-level AI performance can become dramatically cheaper. For infrastructure, coding, log analysis, and many agentic workflows, the model is practically unbeatable on price. Those who truly need the last few points of reasoning performance still pay a large premium. The market is splitting more clearly into three tiers: budget, mid-range, and top-shelf.
Further reading
What does 'silently retrained' mean for DeepSeek V4-Flash?+
Between the original V4-Flash release and early August 2026, DeepSeek continued training the model on the same 284-billion-parameter architecture using fresh data and an improved post-training pipeline. The base architecture, 1 million token context window, and sparse MoE design with roughly 13 billion active parameters stayed the same. The result is a significant jump in Terminal Bench 2.1 performance without a new model name or API version.
Why is the output price gap so much larger than the input price gap?+
V4-Flash costs $0.14 input and $0.28 output per million tokens. Claude Fable 5 costs $10 input and $50 output. The output multiplier matters because production agentic or coding workflows usually produce far more output tokens than input tokens. If you generate a lot of code, tool calls, or long explanations, the competitor's bill is driven mostly by output pricing. There, V4-Flash costs a small fraction of one percent of Claude Fable 5.
Should I use V4-Flash in production, or is GPT-5.6 Sol still the safer choice despite the higher price?+
For routine coding, infrastructure scripting, and agentic workflows that need many iterations, V4-Flash is currently the most cost-efficient option on paper. If you need absolute peak performance on complex reasoning tasks, GPT-5.6 Sol remains ahead — but at roughly 100× the cost. Recommendation: run both models in parallel and measure against your own prompt set; leaderboard scores are not a guarantee for your specific use case.