跪拜 Guibai
← All articles
DeepSeek · AI Coding · Agent

DeepSeek V4 Pro Tested Across Four Frontend Scenarios: Agent Benchmarks Jump 8.6x, but Real-World Speed Tells a Different Story

By 程序员蜡笔熊 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

At ¥3/¥6 per million tokens — 1/86th the cost of Claude Fable 5 for equivalent agent workloads — DeepSeek V4 Pro resets the price floor for capable agentic coding. The model's strength in logic, bug diagnosis, and cross-file consistency makes it a practical daily driver for solo developers and small teams who previously couldn't justify the cost of frontier coding models.

Summary

DeepSeek V4 Pro lands with a 1M-token context window, 384K-token max output, and a three-tier thinking mode that lets developers trade speed for reasoning depth. Official benchmarks show an 8.6x jump on DeepSWE (from 7.3 to 62.7) and a near-tie with Claude Fable 5 on Terminal Bench, but those scores were achieved with the DeepSeek Harness agent framework, not the bare model.

Four bare-API frontend tests reveal a model that excels at logic-heavy tasks — it correctly implemented TypeScript's DeepPartial with function and array edge cases, and introduced AbortController-based race-condition handling during a React bug-fix scenario without being prompted. Component generation and multi-step agent scaffolding were complete and runnable, with consistent code style across files.

The trade-off is speed: thinking-mode responses ranged from 36 seconds for a focused type-system problem to 182 seconds for a complex bug fix. That makes V4 Pro unsuitable for real-time IDE completion but compelling for deep, multi-step engineering tasks where thoroughness matters more than latency.

Takeaways
DeepSWE score jumped from 7.3 (preview) to 62.7 (official), an 8.6x improvement, putting it ahead of Claude Opus 4.8 and Meta Muse Code.
Terminal Bench 2.1 score of 87.9 essentially ties Claude Fable 5 at 88.0.
Input pricing is ¥3 per million tokens, output is ¥6 — roughly 1/36th and 1/90th of Claude Fable 5 respectively.
The official benchmarks were run with the DeepSeek Harness agent framework, not the bare model; bare-API performance will differ.
Three-tier thinking mode (Low/Medium/High) lets users choose between fast responses and deep reasoning.
In a React bug-fix test, V4 Pro independently introduced AbortController for race-condition handling and added HTTP status-code validation.
TypeScript DeepPartial implementation correctly handled function and array edge cases that many developers miss.
Response times ranged from 36 seconds (type-system problem) to 182 seconds (multi-bug fix), making it unsuitable for real-time IDE completion.
Cross-file code style remained consistent across a 7-step, 10-file agent scaffolding task.
Other reviewers confirm the model is strong on backend logic but weaker on frontend UI rendering and visual effects.
Conclusions

V4 Pro's 8.6x DeepSWE leap is less about raw model intelligence and more about the Harness framework providing the tool-calling and context-management layer that the preview lacked — a reminder that agent benchmarks measure systems, not models.

The model's consistent use of AbortController across unrelated tests suggests it has internalized modern React best practices as a default pattern, not as a prompted afterthought.

At 1/86th the cost of Claude for equivalent agent workloads, the economic argument for DeepSeek is strong enough that many teams will accept the slower thinking-mode latency as a fair trade.

The three-tier thinking mode is a practical design choice, but the current API opacity around mode selection and the sparse documentation mean developers can't yet fine-tune the speed-quality trade-off programmatically.

Cross-reviewer consensus that V4 Pro is 'backend-strong, frontend-weak' suggests its training distribution skews toward logic and systems code over visual and UI generation tasks.

Concepts & terms
DeepSWE
An agentic software engineering benchmark that measures a model's ability to autonomously complete multi-step coding tasks — reading files, writing code, running tests, and iterating — rather than just generating single code snippets.
DeepSeek Harness
An MIT-licensed agent framework released alongside V4 Pro that handles tool calling, file I/O, context management, and error handling. It is the engineering layer that lets the model execute tasks end-to-end; official benchmarks were run with Harness, not the bare model.
AbortController
A Web API mechanism for cancelling asynchronous requests. In React, it is used inside useEffect cleanup functions to abort in-flight fetch requests when a component unmounts or dependencies change, preventing race conditions where stale data overwrites fresh data.
DeepPartial<T>
A recursive TypeScript utility type that makes every property of T optional, including nested objects and array elements. Correct implementations must handle edge cases: functions should not be recursed into (they are objects but should remain callable), and arrays should recurse on their element type, not the array itself.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗