DeepSeek V4 Flash makes agent cost a workflow metric

Claude agents show why agent cost is becoming a workflow metric rather than a model sticker price. woshipm cites SemiAnalysis calculating a real blended price of $0.99 per million tokens for an agentic workload, even though Claude 4 Opus was officially priced at $5 per million input tokens and $25 per million output tokens.

The gap comes from the way agents actually run. In the cited workloads, prompt cache hit rate was more than 90%, and cached input was billed at $0.50 per million tokens.

That changes what teams should measure. SemiAnalysis found Claude agents in its workflows used $1.87-$58 in tokens for research tasks that would cost $150-$1000 in equivalent human labor, with ROI ranging from 10.7 times to 80.2 times. The useful question is no longer which model wins a static contest. It is which closed loop can choose the right model, reuse context, complete the task, check the result, and turn mistakes into cheaper future runs.

Measure the task, not the token

A small team should start an agent budget with the finished job, then work backward to tokens. APPSO's unofficial DeepSeek API status monitoring page test shows why. V4 Flash Max made 25 model calls, used 1.22 million input Tokens and 66,995 output Tokens, and cost $0.0758. Ling-3.0-Flash also made 25 model calls, yet cost $0.0402. On the surface, Ling was 40% cheaper. Under the surface, DeepSeek used about 30% more input than Ling's 940,000 Tokens and produced 4.5 times the output, at 66,995 Tokens versus 14,752 Tokens.

That is the trap: the bill says tokens, but the work says task.

The same pattern appears when the task changes. In APPSO's The Odyssey cinema guide case, Ling-3.0-Flash took 17m55s, recorded 137 backend requests, used 3.26 million Tokens, and cost $0.483. Claude Sonnet 4.6 took 16.1 minutes, made 14 tool calls, consumed 1.1 million Tokens, and cost $2.5. The cheaper run used more tokens and far more backend requests. The more expensive run used fewer tool calls and finished faster.

SemiAnalysis's Claude examples, cited by woshipm, push the same accounting into ROI. Claude 4 Opus was officially priced at $5 per million input tokens and $25 per million output tokens, yet SemiAnalysis calculated the real blended price of an agentic workload at $0.99 per million tokens. The reason was structure: agentic workloads had a prompt cache hit rate of more than 90%, and cached input was billed at $0.50 per million tokens.

So the spreadsheet needs task rows, not just token rows. For each run, record model, calls, tool calls or backend requests, input, output, elapsed time, cache behavior, final acceptance, and human review. SemiAnalysis found Claude agents cost $1.87-$58 in tokens for research tasks equivalent to $150-$1000 in human labor, with ROI from 10.7 times to 80.2 times. Hugging Face co-founder clem put the wider spread more bluntly: per-task cost can differ by about 800 times, with leading flagship models averaging more than $31 per task and V4 Flash Max costing only $0.04.

Route models by risk and context type

A practical agent does not have to choose one model for every step. The split can follow risk and context type. ifanr describes DeepSeek V4 Flash as a low-price line that can cover many real tasks, which makes it a fit for broad triage, extraction, rewriting, classification, and first-pass tool calls where errors can be checked before they reach the user.

Vision changes the routing map. DeepSeek's deepseek-v4-flash-vision-exp keeps the text capabilities of DeepSeek-V4-Flash while adding image input, according to ifanr. It can take mixed text and image input, with images supplied through Base64, external URLs, or Files API, and developers can call it through Chat Completions, Messages, and Responses API.

That means screenshots, receipts, UI states, and document images can stay in the cheap lane longer.

The cost logic is concrete. ifanr says deepseek-v4-flash-vision-exp uses the same API pricing system as DeepSeek-V4-Flash, with each image converted into up to 384 tokens and billed under the V4-Flash standard. Cached input costs 0.05 yuan per million tokens off-peak and 0.10 yuan at peak. Uncached input costs 1.5 yuan off-peak and 3 yuan at peak. Output costs 4.5 yuan off-peak and 9 yuan at peak, with peak hours from 9:00 to 12:00 and 14:00 to 18:00 Beijing time.

The premium lane still has a job. woshipm reports Claude Opus 4.5 was priced 3 times cheaper than its predecessor, while Opus fast cost 6 times regular Opus and Mythos cost $25 per million input tokens and $125 per million output tokens, or 5 times regular Opus. woshipm also notes Mythos scored 93.9% on SWE-bench in April, a test of fixing historical bugs inside real GitHub repositories. The routing lesson is simple: cheap models handle volume and visible context; premium reasoning is reserved for failures, high-stakes code changes, and decisions the workflow cannot cheaply verify.

Closed-loop AI work systems compared across model, harness, context, cost, verification, and feedback

The harness is where the feedback loop lives

The harness is the layer that turns an AI model from a reply engine into a work system. Woshipm describes it as the machinery that steers models after they enter real work: breaking tasks down, calling tools, checking results, and turning failures into experience for the next iteration. That definition matters because the useful signal often appears after the model has answered, when a person accepts output, rejects it, edits it, or has to repair the surrounding workflow.

Code is the clearest example. Woshipm notes that programmers' acceptance, rejection, and manual modification of AI-generated code in products such as Claude Code can become valuable feedback signals. The model's first draft is only one event. The harness can observe what survived contact with the repository.

That is why Weng Li's view, cited by woshipm, is useful: near-term AI self-improvement may appear first at the Harness layer, not through models directly rewriting their own weights. A product can improve by learning which files matter, which tools are safe to call, where checks fail, and which human edits recur. Woshipm's author frames the broader loop as better products bringing more users, which create more correction data, which strengthens models and feeds back into better products.

Tencent's WorkBuddy shows the same pattern outside pure programming. According to woshipm, Tencent built the programming assistant CodeBuddy before WorkBuddy appeared publicly. In January 2026, Tencent's team built WorkBuddy 0.01 on the same foundation and first gave it to more than 2,000 non-technical Tencent employees for internal use. WorkBuddy's PC monthly visits reached 8.85 million in March.

The gains were measured inside the harness, not in isolation. After Hy3 Preview was connected to WorkBuddy, first response became 54% faster and total task duration fell by 47%. When the official version of Hy3 was released in July, its task success rate in WorkBuddy scenarios rose from 72% to 90%, while task time was reduced by 34%.

DeepSeek Harness points to the same direction for multimodal agents. Ifanr reported that DeepSeek Harness 0.1.1 added native support for deepseek-v4-flash-vision-exp, letting developers call it directly to add image understanding to existing Agent workflows.

OpenAI's desktop integration fits the pattern too. Woshipm reported that on July 10, Sam Altman released a new version of the ChatGPT desktop app, with the previously standalone Codex integrated as a tab between "Chat" and "Work." The placement is telling: the agent is moving closer to files, tools, checks, and user corrections.

Cheap retries can still fail expensively

Cheap agents change the failure mode. When a model is inexpensive enough to retry, a team can mistake repeated motion for progress. The bill may stay low while the task drifts: wrong file touched, weak tool call accepted, bad assumption copied into the next run. The expensive part is no longer the token charge. It is the delay before a person notices that the workflow should have escalated.

OpenAI's own Codex usage data, cited by ifanr, shows why this matters. In May 2026, 70.2% of users had submitted at least one Codex task requiring at least one hour of human work, and 25.6% had submitted at least one task estimated to require more than eight hours of human work. Once agents operate in that zone, a cheap retry is not a harmless extra click. It is a decision about supervision.

China's model market gives teams more room to make that decision badly or well. woshipm notes that Chinese companies account for 11 of the top 20 laboratories on LMArena's text leaderboard. Artificial Analysis listed 99 models with intelligence scores above 20, from 25 companies, including 12 Chinese companies: DeepSeek, Baidu, Zhipu, Tencent, China Mobile, Moonshot AI, Ant Group, Xiaomi, MiniMax, Alibaba, StepFun, and ByteDance. OpenCode said DeepSeek V4 Flash consumed 8 trillion Tokens through its platform. Efficient models are not theoretical capacity there; they are becoming working inventory.

But enterprise money still clusters around the strongest default suppliers. According to woshipm, OpenAI and Anthropic together accounted for 67% of enterprise LLM API spending in 2025, based on Menlo Ventures data. The Information reported in May that OpenAI and Anthropic took 89% of revenue among 34 leading AI startups it tracked, while Axios cited Ramp data saying Anthropic received more than 73% of spending among companies buying AI products for the first time.

That split creates the practical rule: use cheaper models for recoverable work, not for invisible risk. woshipm's author argues that DeepSeek showed a good model need not require the most expensive chips or highest training costs, and that a low-cost good model can cover 80% of application scenarios. The missing clause is escalation. If the agent cannot verify the result, find the right context, or explain the uncertainty, the workflow should spend more on a stronger model before it spends more time failing.

How to evaluate agent workflows instead of isolated model answers

• You are testing whether a cheaper model can handle a real multi-step task, not just a benchmark prompt. Run the whole workflow and compare task-level cost, calls, tokens, and completion time. In APPSO's unofficial DeepSeek API status monitoring page test, V4 Flash Max made 25 model calls, used 1.22 million input Tokens and 66,995 output Tokens, and cost $0.0758; Ling-3.0-Flash also made 25 model calls and cost $0.0402. That kind of end-to-end measurement is more useful than asking which model is generally 'smarter.'
• You need high-volume agent execution where token cost can dominate product feasibility. Prefer models and harnesses with low per-task cost and strong cache economics, then verify whether the output quality is sufficient. SemiAnalysis calculated that the real blended price of an agentic workload had fallen to $0.99 per million tokens, with prompt cache hit rate of more than 90% and cached input billed at $0.50 per million tokens. The sources also cite research tasks where Claude agents cost $1.87-$58 in tokens versus $150-$1000 in equivalent human labor.
• You are choosing between a flagship model and a Flash-class model for routine agent work. Do not assume the flagship model is the default. The source material describes DeepSeek V4 Flash as able to cover many real tasks while being priced low enough, and cites a claim that per-task cost differences between models are about 800 times, with leading flagship models averaging more than $31 per task and V4 Flash Max costing only $0.04. Use flagship models when the task failure cost justifies it; otherwise test Flash-class models first.
• Your workflow needs images as part of the context. Use a multimodal agent model only if the image path is supported by the workflow. deepseek-v4-flash-vision-exp can process mixed text and image input, accepts images through Base64, external URLs, or Files API, and can be called through Chat Completions, Messages, and Responses API. DeepSeek Harness 0.1.1 added native support for it, so it is a candidate when an existing Agent workflow needs image understanding.
• You want the system to improve from real work rather than from one-off prompting. Instrument the harness layer: task decomposition, tool use, result checking, and failure capture. The sources define Harness as a system for steering models after they enter real work, including task decomposition, tool use, result checking, and turning failures into experience for the next iteration. They also say programmers' acceptance, rejection, and manual modification of AI-generated code in products such as Claude Code can leave valuable feedback signals.

Closed-loop systems are moving beyond office software

Robot deployments make the agent problem visible because the loop cannot end at a model answer. At the World Robot Conference 2026, held at the Beijing Etrong International Exhibition & Convention Center on August 20, a Chinese tech outlet counted more than 300 exhibitors and as many as 3000 robot products. Official data also put 49 central state-owned enterprises on site as groups, with 12 engineering scenarios spanning power grids, aerospace, mining, and metallurgy.

That is a different evaluation surface from a chatbot prompt.

Cowarobot is the clearest example in the ledger. According to the report, executive vice president Zhang Xiaoming said its robots have reached regular operation at the 10,000-unit level across multiple urban scenarios in more than 50 cities and regions worldwide. The same report said Cowarobot has accumulated more than 50PB of real-scenario data, used to improve model decision-making and generalization. Here, quality is measured as deployment, data capture, error return, and model update.

The household demos show the harder edge. Qianxun showed a robot interpreting "help me tidy the living room" and splitting it into actions such as putting bowls in the dishwasher, returning a Coke on the table, putting toys into a basket, and putting trash into a trash can. But the report also reported that multiple industry participants expect household robots to need at least a 2-5 year dark period before expectations and real progress converge.

Practitioners should separate operating evidence from capital-market claims. woshipm reported Anthropic ARR moving from less than $9 billion at the beginning of 2025 to close to $60 billion in the second quarter of 2026, while another row gives 14 billion yuan to 47 billion yuan. It also cites more than 70% inference gross margin, 34.4% enterprise share, nearly 1 trillion dollars in valuation, and a SemiAnalysis claim that December 2025 marked a model-economics turn. Those figures need reconciliation before they are compared with English-language public data.

Watch the places where agents meet the physical world, because they expose the same measurement problem in harder form. According to a Chinese tech media report, Cowarobot says its robots run regularly at the 10,000-unit level in more than 50 cities and regions, backed by more than 50PB of real-scenario data. UBTech's Jiao Jichao says the industry still lacks enough real interaction data and shared training standards. That gap is the warning for office agents too: logs are not exhaust. They are the product input.

Start treating each repeated workflow as a data asset. Track what the agent saw, what it changed, what humans rejected, and what had to be retried.

Use deployment shape as the routing clue. Qianxun's living-room demo breaks one household request into visible subtasks. Xinghaitu's front warehouse targets about 100 orders in 8 hours across tens of thousands of SKUs. Mojia says robot traffic police have reached the hundreds across ten cities, including Wuhu, Changzhou, and Suzhou. The practical question is similar for DeepSeek V4 Flash: which steps are repetitive enough, verifiable enough, and cheap enough to hand to an efficient model, while keeping premium review for the work where one bad answer is expensive?

For readers outside China

• Availability: The source material confirms that deepseek-v4-flash-vision-exp is available on the DeepSeek API platform and can be accessed by setting model='deepseek-v4-flash-vision-exp'. It also says developers can call it through Chat Completions, Messages, and Responses API, and that DeepSeek Harness 0.1.1 added native support. Availability outside China is not disclosed in sources.
• Pricing: For deepseek-v4-flash-vision-exp, cached input costs 0.05 yuan per million tokens during off-peak hours and 0.10 yuan per million tokens during peak hours; uncached input costs 1.5 yuan per million tokens during off-peak hours and 3 yuan per million tokens during peak hours; output costs 4.5 yuan per million tokens during off-peak hours and 9 yuan per million tokens during peak hours. Peak hours are 9:00 to 12:00 and 14:00 to 18:00 Beijing time every day. The sources also state that it uses the same API pricing system as DeepSeek-V4-Flash and that each image is converted into up to 384 tokens and billed according to the V4-Flash pricing standard. For comparison, Claude 4 Opus was officially priced at $5 per million input tokens and $25 per million output tokens; Mythos was priced at $25 per million input tokens and $125 per million output tokens; Claude Opus 4.5 was priced 3 times cheaper than its predecessor. Other subscription or seat pricing is not disclosed in sources.
• Closest Western equivalents: Claude Code-style agentic coding workflows, where user acceptance, rejection, and manual edits can become feedback signals; OpenAI Codex inside the ChatGPT desktop app, which the sources say was integrated as a tab between "Chat" and "Work"; SWE-bench-style evaluation, because the sources describe it as testing whether models can enter real GitHub repositories and fix historical bugs that human programmers previously encountered
• Data residency: The source material does not disclose data residency, cross-border data handling, retention, enterprise isolation, or compliance terms for these services. It does say DeepSeek's Files API supports JPEG, PNG, GIF, and WebP formats, a maximum single file size of 64 MiB, maximum storage of 25 GiB per user, and up to 10000 saved files, but it does not say where those files are stored or how long they are retained.

Sources

• woshipm AI巨头正在争夺人类的错题本 https://woshipm.com/it/6432834.html
• ifanr DeepSeek V4 Flash 之后，大模型开始卷「智效比」了 https://ifanr.com/1675156
• ifanr 突发| DeepSeek 多模态模型上新，鲸鱼终于开「天眼」了 https://ifanr.com/1676067
• 36kr 模型不拼参数拼闭环、讲家庭场景落地，今年的大会不再「性感」丨直击WRC2026 https://36kr.com/p/3948754198117504
• woshipm 完了，国内大模型也卷成"新能源车"了 https://woshipm.com/ai/6427624.html
• woshipm AI产业链利益大洗牌：芯片、储存狂欢退潮，利润正疯狂向模型层涌入 https://woshipm.com/ai/6432642.html