Qwen3.8-Flash-Next vs DeepSeek V4 Flash 0731 for Coding Agents
In this article
An agent can write the right function and still leave a broken project. It misses a caller, forgets a type or stops after a failing test.
If DeepSeek V4 Flash 0731 handles automated changes in your workflow, Qwen3.8-Flash-Next deserves a place in the next trial. Its public agent results offer a credible starting point. A migration is worthwhile only if it leaves fewer of those loose ends in your repository.
Where Qwen’s advantage appears
These measurements come from the September 7, 2026 snapshot. Qwen is the AA and LiveBench Flash-Next entry, named Next-Flash on ModelRank. DeepSeek is specifically 0731.
| Test | Qwen3.8-Flash-Next | DeepSeek V4 Flash 0731 |
|---|---|---|
| AA Terminal-Bench 2.1 | 86.1% | 78.7% |
| AA SciCode | 50.6% | 50.3% |
| LiveBench Agentic JavaScript | 68.182 | 63.636 |
| LiveBench Agentic TypeScript | 46.667 | 26.667 |
| LiveBench Agentic Python | 70.000 | 50.000 |
| LiveBench Coding, generation/completion mean | 72.55 | 74.98 |
Qwen is higher on all three Agentic language measurements in this snapshot. DeepSeek has the higher ordinary Coding mean. Treating all of these as one undifferentiated “coding” result would lose useful information.
The TypeScript row has a 20-point difference in the source values. That makes a TypeScript repository a sensible place to investigate. It does not predict 20% more completed tasks in your project. Task composition, sampling uncertainty and configuration differences remain unresolved by the decimal places.
AA identifies DeepSeek’s effort as max and leaves Qwen’s unspecified. This is not an experiment with every variable controlled.
Price needs a channel
ModelRank’s September 6 OpenRouter record for Qwen’s corresponding entry lists $0.15 / $0.47 per million input/output tokens. DeepSeek’s official 0731 prices, rechecked September 8, are $0.22 / $0.66 off-peak and $0.44 / $1.32 peak.
For 1 million input and 200,000 output tokens, that is $0.244 for Qwen, $0.352 for off-peak DeepSeek and $0.704 for peak DeepSeek. The examples exclude cache discounts, extra tool charges and retries. They compare the stated purchasing channels, not two official provider tariffs.
DeepSeek peak hours are weekdays, UTC 01:00–04:00 and 06:00–10:00; all other times are off-peak. Cached requests need a separate calculation. A model change can also change how effectively your application uses caching.
Confirm the Qwen version and route you will actually buy. The word “Flash” is not an interchangeable API identifier.
Choose a task that exposes unfinished work
A useful acceptance task changes an API return type, updates every caller and adds a regression test. It makes the agent locate files, follow references, edit code and run commands.
Inspect what “tests passed” means. Did it run the right tests? Remove assertions to get a pass? Add an unrelated dependency? Can it explain a failure and continue, or does it stop at “fixed”?
Use the same starting commit, permissions and budget for both candidates. Record first-attempt completion separately from total time and cost after repairs. Keep failed tasks in the denominator.
If Qwen proves more reliable on that task class, move that class first. A project that mainly needs completion, and already works well with DeepSeek, does not have to migrate because a separate Agentic result is higher.
What A and B+ mean here
ModelRank currently places Qwen in editorial tier A and DeepSeek 0731 in B+, with reference indices of 105 and 100. Those are an editorial grouping and a two-test index, not capability ratios across every repository.
For a general default, start with the GLM, Qwen and DeepSeek guide. For the current source records, use the Qwen–DeepSeek comparison.
Sources
AA Qwen, AA DeepSeek, LiveBench CSV, Qwen channel entry, DeepSeek official pricing.
The LiveBench filename is a dataset version. Repository tasks above are suggested acceptance checks for a migration trial.