DeepSeek V4.1 Flash Undercuts Older Flash on Price and Beats Pro on Agentic Tests

Three models, one endpoint

Novita AI says its V4 family now holds three current models: DeepSeek V4 Pro 0813, DeepSeek V4 Flash 0731 and DeepSeek V4.1 Flash. All three share a 1,048,576-token context window, a maximum output of 393,216 tokens and the same OpenAI-compatible endpoint, so switching between them means changing only the model value in a request.

The preview-era V4 Pro has been superseded by the GA release, listed as deepseek/deepseek-v4-pro-0813. According to Novita AI, it adds a DSpark speculative-decoding module, ships under an MIT license, and its pricing changed at GA, so figures in older posts no longer apply. The original unsuffixed V4 Flash is replaced by the dated deepseek/deepseek-v4-flash-0731.

The newest model is DeepSeek V4.1 Flash, described as the September 2026 successor to V4 Flash. It is the first in the family to accept image input alongside text, and it uses a causal encoder-decoder design with a smaller KV cache. The other two models are text to text.

Pricing on Novita AI

Novita AI lists the following prices per 1M tokens:

• V4 Pro 0813: $1.32 input, $0.044 cache read, $3.96 output
• V4 Flash 0731: $0.44 input, $0.028 cache read, $1.32 output
• V4.1 Flash: $0.30 input, $0.006 cache read, $1.20 output

The newer Flash model is therefore cheaper than its predecessor on input, output and cache reads. For a workload of 1M input tokens and 1M output tokens with no cache hits, Novita AI calculates $5.28 for V4 Pro 0813, $1.76 for V4 Flash 0731 and $1.50 for V4.1 Flash. Pro costs roughly 3.5x V4.1 Flash for the same volume.

Novita AI notes that the cheaper model is not automatically the right default. If a task needs the higher knowledge-reasoning scores of Pro 0813, the extra cost can still be lower than the cost of a wrong answer. The prices were checked on October 13, 2026, and the platform advises confirming them on the model pages before building a cost model.

Benchmarks: agentic work versus reasoning

The comparison draws on DeepSeek's DeepSeek-V4.1-Flash model card, using maximum reasoning effort. V4.1 Flash leads on the agentic tests, while Pro 0813 leads on GPQA Diamond.

• Terminal-Bench 2.1 (Pass@1): 87.9 for Pro 0813, 82.7 for Flash 0731, 90.6 for V4.1 Flash
• Terminal-Bench 3.0 (Pass@1): 11.8, 7.6 and 30.0
• DeepSWE v1.1 (Resolved): 62.7, 54.4 and 74.2
• AutomationBench (Pass@1): 43.2, 37.7 and 54.8
• GPQA Diamond (Pass@1): 92.4, 89.9 and 90.9

Novita AI singles out the Terminal-Bench 3.0 gap, 30.0 against 11.8 for Pro, as large enough to change whether a long tool-use loop finishes at all. V4 Flash 0731 trails on both axes, which is why the platform describes it as a fallback rather than the budget pick. It also advises teams to treat the numbers as a starting point and to run their own prompt set, comparing accepted answers, retry rate, latency, structured-output reliability and total token cost.

What it means for developers

Novita AI recommends V4.1 Flash for coding agents, terminal automation, multi-step tool use and any workload that needs image, screenshot or document-page input. It recommends V4 Pro 0813 for knowledge- or reasoning-heavy tasks with a high failure cost. V4 Flash 0731 is worth keeping only for high-volume, text-only workloads already tuned to its behavior and still matching in the team's own evaluation.

For mixed traffic, the platform suggests routing by tier: baseline requests go to V4.1 Flash, and hard reasoning escalates to V4 Pro 0813.

Its migration advice is to confirm model IDs and pricing before shipping, since the family has changed IDs and prices more than once this year. It also suggests keeping the model ID, context limit and pricing in configuration so a rollback is easy, and updating message construction to include image parts when adopting V4.1 Flash for image input. Teams should run a representative regression set and compare quality by task type, not only aggregate win rate.