DeepSeek has set its new Flash model a challenge: Can it replace Pro?

Yesterday, an interim version of V4.1 Flash entered closed beta. DeepSeek touted a new model architecture, native multimodality and faster speeds. But in the accompanying feedback survey, it asked users directly: “Do you think this model can fully replace the online DeepSeek V4 Pro?”

DeepSeek is pitting Flash against itself, testing whether it can outperform Flash on quality—and whether workloads can be handed to a cheaper model better suited to high-volume calls.

There is already a substantial price gap between Pro and Flash. During peak hours, Flash charges 9 yuan per million output tokens, compared with 27 yuan for Pro.

Just now, however, DeepSeek announced another Flash price cut on its open platform, effective at noon on September 10. During off-peak hours, cached input will fall to 0.02 yuan per million tokens, uncached input to 1 yuan and output to 4 yuan. Peak pricing will remain double the off-peak rate. The three reductions are 60%, about 33% and about 11%, respectively.

The new Flash is still being tested as a potential Pro replacement, but the barrier to using the Flash series is already coming down.

Developers may occasionally pay three times as much for better results. But in everyday use, particularly for steps that may or may not be necessary, the case for Flash becomes clear.

DeepSeek wants the new version to combine the strengths of both models: Flash’s speed and Pro’s power, making Flash V4.1 both fast and capable.

More than a year ago, R1’s impact made people willing to give AI more time to see what it could come up with. DeepSeek must now answer a different question: How can the reasoning capabilities it has already developed be used more frequently?

“Intelligence Density” Starts to Matter

When R1 took off in early 2025, part of its appeal was that this capability had finally become easy to access. DeepSeek released the model weights while making R1 available through its website, app and API. Ordinary users could activate it by selecting “DeepThink,” while developers could distill their own models from its outputs.

High-level reasoning was no longer something only a handful of labs could demonstrate. It entered chat windows and the product development pipelines of other companies.

When DeepSeek updated R1 in May 2025, it presented “deeper thinking” as the most important advance.

On AIME 2025, the original R1 used about 12,000 tokens per problem on average, while the new version used 23,000. Accuracy rose from 70% to 87.5%. The update made R1 more accurate, but its average reasoning consumption per problem also nearly doubled.

The approach has its logic. Difficult problems have no ready-made answers, so the model must try different methods, check intermediate steps and recalculate after finding errors. If the result is worthwhile, users may accept the extra wait.

But in its V3.2 report, released in December 2025, DeepSeek set itself another goal: increasing the “intelligence density” of its reasoning chains. The reason was clear. To match the output quality of models such as Gemini 3.0 Pro, it typically had to generate a longer reasoning process.

Waiting time for output has become important. The gap shows up both in speed and in the cost of each request.

The issue becomes even more pronounced when dividing work between large and small models. An early V4 model card said Flash could approach Pro’s reasoning performance when given a larger thinking budget, but still lagged in knowledge and the most complex agent workflows.

A small model costs less at each computational step, but it may need more steps to reach the same result as a larger model. How much of the apparent unit-price advantage remains depends on how the task is ultimately completed.

For Flash to replace Pro, it must improve both capability and efficiency. Generating lengthy reasoning faster can reduce some of the wait; finding an effective approach earlier could eliminate later computation altogether. For frequent AI users, the latter matters just as much.

Researchers have begun incorporating user wait times into optimization targets. An EMNLP 2025 study found that the most compute-efficient test-time scaling strategy does not necessarily deliver the lowest latency. Researchers must rearrange parallel computation and speculative decoding to achieve better results within a fixed period.

DeepSeek has also been working to increase generation speed. In its DSpark paper published in July, the team reported that, under live V4-Flash user traffic and at the same throughput, per-user generation speed improved by 60% to 85% compared with an MTP-1 baseline. The team focused on wasted verification in speculative decoding, delivering output to users faster while preserving the service’s overall request capacity.

That matters for DeepSeek as an API provider. Faster request completion is useful only if it does not seriously compromise overall service capacity. Developers need shorter waits, while DeepSeek needs the same pool of computing resources to serve users more efficiently.

“Can it solve the problem?” and “Are users willing to wait for it to solve the problem?” are becoming two separate tests.

They also correspond to two markets. Researchers can devote substantial computing resources to a difficult problem, while everyday assistants must continually respond to new requests. After advancing the former capability, vendors still need to turn it into the latter kind of service before they can win adoption at scale.

The intelligence density proposed in DeepSeek’s report will ultimately be tested here: Can tasks that once required an expensive model and lengthy computation gradually become routine calls?

Agents Must Be Judged by the Cost of Completing the Job

On the day the closed beta began, a LINUX DO user named “Jueqian” happened to need a maintenance page for scheduled downtime, so he connected the new Flash model to his coding tool.

The model’s performance surprised him. It found a requirement in the project guidelines for decision records that had long been deprecated but never removed, then carefully laid out plans for documentation maintenance, code submission and review. After generating the page, he asked it to keep the style consistent with the existing project. The model then found illustrations already in the repository and modified them.

He was pleased with the model’s ability to execute the task, but found it expensive: completing the maintenance page cost 15.5 yuan.

This is where agents differ from ordinary chatbots. A user assigns a single task, but the model may run many rounds in the background: first reading project guidelines, then checking files and modifying code, and finally continuing its review after receiving results from tools. Each call may incur new input and output charges. A low per-token price does not guarantee that the overall job will be cheap.

Making agents cost-effective requires more than lowering model prices. Developers must also determine whether all those calls are necessary. Some steps require model judgment, while others simply involve reading, filtering and organizing material according to established rules. Sending the latter back to the model step by step adds calls and waiting time, while intermediate results keep consuming context.

DeepSeek’s open-source Harness can help orchestrate this execution process. DeepSeek’s website explains the relationship as “Agent = Model + Harness”: the model makes decisions, while Harness supplies tools and manages sessions and the runtime environment, turning those decisions into concrete actions.

Its PTC mode allows the model to write a program that chains together multiple tool operations in one pass. For example, if the model needs to find qualifying content across a group of files, it can have the program read and filter them in batches before returning only the relevant results. The model does not need to decide what to do next after reading each file, nor does it need to receive all the raw content.

This could reduce repeated calls and useless input during task execution. DeepSeek’s official PTC design notes also identify these two types of overhead as problems to address. For developers, the ultimate comparison is whether the same job can be completed well—and what the total bill is afterward.

The process offers opportunities to remove steps that do not require the model’s repeated involvement. When the method for processing a batch of material is already known, a program can execute the full workflow before the model evaluates the information that remains.

DeepSeek’s efforts to reduce repeated computation go back further. Its V3.2 report noted that discarding existing reasoning during consecutive tool interactions forces the model to analyze the entire problem again. The team therefore adjusted context management so earlier reasoning could be retained throughout the process.

Efficiency losses can no longer be traced only to model parameters; they also arise from how the model reasons and how tools execute its decisions.

Harness also records prompts, tool calls and results, allowing developers to review tasks and identify where redundant processing or execution failures occurred.

An official development log dated August 10 shows that DeepSeek corrected the tool and prompt configuration for minimal mode, giving it the same persistent Bash environment used during reinforcement learning so that consecutive operations retain their working state.

Details like these explain why model companies also build execution frameworks. A model may learn how to use tools, but it still needs suitable tools and an appropriate environment at runtime. By opening up one clearly defined execution setup, DeepSeek gives application developers a reference they can directly adopt, inspect and modify.

These improvements are converging on the same goal: faster server-side generation, less repeated reasoning by the model and fewer unnecessary calls through the execution framework. Users once watched AI think; now AI must keep working without allowing costs and waiting times to spiral.

As Everyday Models Improve, Flagships Must Tackle Harder Problems

DeepSeek is not the only AI model company rearranging its premium capabilities.

On July 24, Anthropic released Opus 5, promising near-frontier capabilities at half the price of Fable 5 and making it the default model for Claude Max.

Anthropic even introduced Fable in terms of “cost per task.” In the company’s official CursorBench 3.2 test, Opus 5 at maximum thinking intensity came close to Fable 5’s highest score at roughly half the task cost. In the same launch materials, Cursor co-founder Sualeh Asif described it as near-Fable intelligence at Opus speed and cost.

The direction closely resembles DeepSeek’s: once premium capabilities can be offered more cheaply, the existing pricing tiers must be recalibrated.

The positioning of Fable and Astra could offer a reference point for the next generation of Pro. Anthropic places Fable in coding tasks that take hours, span multiple applications or even continue for days. OpenAI, meanwhile, positions GPT-6 Astra for the hardest end-to-end work, including complex reasoning, coding, research and document creation.

These flagship models are competing for complex work to which users are willing to commit more time and budget. Once given a goal, the model must break down the task, call tools, check results and keep moving forward when problems arise. Its value must be demonstrated by the work it ultimately completes.

Premium models need to assume more complete responsibility. Users hand over a task expecting a result they can inspect and use. The model must stay on course, use tools and continue after failures rather than asking the user to take over every few steps.

Seen from this perspective, one reasonable prediction for a future DeepSeek Pro is that, if Flash can take on much of Pro’s current workload, the next Pro could devote more computing resources to harder, longer tasks and pursue the product position represented by Astra and Fable. If users must give it ample time to think, the results must be good enough to justify the wait.

This also explains why native multimodality and Harness are becoming important. A model expected to complete work independently will eventually need to read charts, view pages and determine whether the interfaces it creates meet requirements. A model limited to processing text and offering suggestions will be constrained in the work it can take on.

As early as the V4 preview released in April, DeepSeek explicitly said it had adapted the model for agent products including Claude Code, OpenClaw, OpenCode and CodeBuddy. Its showcased results included generated PowerPoint slides. The model was already entering code and document workflows inside specific tools.

Now, the new Flash is updating its architecture and multimodal capabilities, while Harness is changing how tasks are executed. Taken together, DeepSeek’s efficiency drive now covers the full process, from a single reasoning pass to an entire task.

For DeepSeek, greater efficiency offers two rewards. Reducing the computation required to complete a task can lower costs; completing more useful reasoning within the same wait time could make harder problems solvable. The latter will push “DeepThink” further, while the former will determine how widely those advances enter everyday use.

V4.1 Flash remains an interim release. Whether it can fully replace Pro, shorten reasoning chains and clarify the positioning of the next Pro will require further evidence. The DSpark speed gains cited in this article came from tests of the earlier V4 service and cannot be counted directly as an achievement of this update.

Still, the questions for the next round of competition have already changed. Flash must make more of this work economical. Pro must find more work worth waiting for.