DeepSeek's V4 Flash aces the leaderboards but stumbles on real work, then raises its prices
The AI model everyone called a bargain just got more expensive. New testing shows its real-world performance is patchier than its benchmark scores suggest.
Key points - DeepSeek V4 Flash completed only 53.8% of 240 complex multi-step tasks in independent testing published in August 2026. - DeepSeek raised API prices by as much as 1,100% after V4 Flash became the most-used model on the OpenRouter platform by weekly token volume. - V4 Flash now costs 22 cents per million input tokens off-peak, with peak rates double that; cache-hit pricing rose between 52% and 1,100%. - Despite the increases, DeepSeek remains cheaper than rivals from OpenAI, Anthropic, Cohere, xAI and others across all pricing tiers. - Enterprise adoption is still early, with serious use mostly limited to low-risk, batch-style workloads rather than live customer-facing systems.
DeepSeek V4 Flash arrived in public beta on 31 July 2026 and immediately broke records. Within days it sat at the top of OpenRouter's weekly usage chart, a platform where developers route requests to dozens of AI models. ML researcher Nathan Lambert called it a "total monster." The price was part of the appeal: a fraction of what OpenAI or Anthropic charge for comparable intelligence.
Then two things happened at once.
How did it actually perform on real tasks?
Not as cleanly as the leaderboards implied. Testing firm Composio put V4 Flash through 240 runs across eight different agent harnesses, the software frameworks that let an AI model take real actions inside tools like Gmail, GitHub and Google Sheets, rather than simply answer questions. Only 129 of those 240 runs passed, a 53.8% rate. Of the 30 workflows tested, just six were completed successfully by every framework.
The same model, on identical tasks, produced different results depending on how it was configured. As VentureBeat reported, orchestration factors including retries, caching behaviour and tool configuration shaped outcomes just as much as the model's raw capability.
Meta software engineer Naman Ahuja saw this in his own experiments, conducted independently of his employer. He built a home-automation agent using V4 Flash to coordinate a thermostat, a Ring security system and door locks. His conclusion was blunt: "Once a model can take actions, reliability matters as much as intelligence."
For anyone building on top of this model, that's the real finding. Getting to a passing result isn't just about picking a smart model; it's about what surrounds it.
What is the price increase, and who pays more?
DeepSeek's increases range from 51% to over 1,100% depending on model, token type and time of day. The new rates:
A "token" is roughly three-quarters of a word.
Analyst Sanchit vir Gogia of Greyhound Research framed the structure plainly: "This is not a simple price rise. It is a pricing architecture that makes the timing of inference an economic variable." Work that can wait, summarising documents or generating test data, can stay cheap. Live customer interactions can't.
Tech analyst Carmi Levy called it a risky move for a company still building credibility, but noted DeepSeek remains far cheaper than OpenAI, Anthropic, Google and others even at the new rates. The business case still works; the maths just needs to be tighter. We looked at how American labs responded to exactly this pressure in our 14 August story on the AI price war.
What does this mean if you are thinking about using it?
For most businesses, DeepSeek suits contained, repetitive tasks with clear success metrics, where a wrong answer's easy to catch. Document classification, code review queues, overnight data processing.
Enterprise use for live, high-stakes workflows demands something harder to demonstrate: consistent reliability, data governance and auditability. DeepSeek hasn't fully established those credentials at scale yet.
IT teams will pilot in sandboxed environments before bringing budget requests to senior leadership. That's the realistic timeline here, not a wave of immediate full rollouts.
The one doable takeaway: schedule heavy processing jobs during off-peak hours and start with a workflow where a failed run is recoverable. Build in retries and a fallback model from day one. The model's genuinely capable; the surrounding setup decides whether that shows up reliably.
