OpenAI Cut Its AI Prices 80%. DeepSeek Just Went Even Lower.

OpenAI does not usually blink first. On July 30, the company cut the price of GPT-5.6 Luna, the fastest and cheapest model in its lineup, by 80 percent, dropping input and output costs from $1 and $6 per million tokens down to $0.20 and $1.20. It was a striking move for a lab that had launched the entire GPT-5.6 family barely three weeks earlier. Two days later, DeepSeek answered. The Chinese lab pushed out V4-Flash-0731, an updated version of its agent-focused model, priced roughly 30 percent below OpenAI’s freshly discounted Luna on output tokens. The message was hard to miss: whatever floor OpenAI thought it had just set, DeepSeek was ready to go lower.

Neither move happened in a vacuum. Both labs are chasing the same prize: developers building AI agents that call tools, write code, and run multi-step tasks in a loop rather than answering a single question and stopping. Agentic workloads burn through tokens at a rate a simple chatbot exchange never approached, and whichever company wins the price floor on agent-grade inference stands to become the default backend for a huge share of the AI tooling being built right now.

An 80 Percent Cut That Still Was Not Enough

To understand why OpenAI moved first, it helps to look at what Luna actually is. It sits at the bottom of the GPT-5.6 family, below the mid-tier Terra and the flagship Sol, built for high-volume, latency-sensitive work rather than the hardest reasoning tasks. OpenAI also trimmed Terra’s price by 20 percent, from $2.50 / $15 to $2 / $12 per million tokens, while leaving Sol untouched at $5 / $30. That pattern, cutting the cheap workhorse hardest and leaving the flagship alone, is a pricing playbook aimed squarely at developers who route routine agent calls to whichever model is fast and cheap enough to get the job done.

OpenAI framed the cut as a byproduct of system efficiency gains rather than a defensive reaction. That framing gets harder to sustain given the timing. A CNBC investigation published in early July found that Chinese-origin models had captured 46 percent of US enterprise token usage tracked on OpenRouter, at times overtaking American models entirely. When a large chunk of the market you built is quietly migrating to cheaper alternatives, an 80 percent price cut reads less like an efficiency dividend and more like a company defending share.

DeepSeek’s Answer Arrived in 48 Hours

DeepSeek did not need to build a new model to respond. V4-Flash-0731 keeps the same architecture and parameter count as the earlier V4-Flash preview, a 284 billion parameter mixture-of-experts model with 13 billion active parameters and a 1 million token context window. What changed is the post-training: DeepSeek re-tuned the model specifically for agentic tool use and adopted the Responses API format that OpenAI’s own Codex-style agents rely on, which matters because it lowers the switching cost for anyone already building against that convention. Teams can point their existing agent code at DeepSeek’s endpoint without rewriting how they call tools.

The pricing itself was not new. DeepSeek’s $0.14 input and $0.28 output rates for V4-Flash had been in place before OpenAI’s cut. What DeepSeek did was let those existing prices do the talking at exactly the moment they became a live counterpunch, undercutting Luna’s new rate by roughly 30 percent on output while offering a model tuned harder for the same agentic tasks Luna is meant to handle.

The Performance Gap Is Narrower Than the Price Gap

None of this would matter much if DeepSeek’s model could not actually do the work. On raw intelligence benchmarks it still trails the best Western models. Artificial Analysis puts Claude Opus 4.8 at 57 on its Intelligence Index against roughly 52 for DeepSeek’s V4 Flash line, and Opus 4.8 leads on SWE-bench Verified (87.6 percent versus about 80.6 percent) and SWE-bench Pro (64.3 percent versus 55.4 percent). Those are real, measurable gaps, and anyone routing the hardest coding or reasoning work will still reach for Opus.

But “still behind” and “far behind” are different stories, and the newest DeepSeek numbers tell the second one. On TerminalBench 2.1, a test of how well a model handles extended terminal sessions, DeepSeek’s own reported score for V4-Flash-0731 was 82.7 against Opus 4.8’s 85.0, a gap of just over two points on a benchmark where DeepSeek trailed by a much wider margin a year earlier. DeepSeek also generates output roughly 115 tokens per second against Opus 4.8’s 60, nearly double the throughput, which matters directly for agent loops that need to plan, act, and re-check within a tight budget of wall-clock time.

Put those together and the practical calculus shifts. A model that lands within a few points of the frontier on the benchmarks that matter for agent work, runs close to twice as fast, and costs roughly 13 times less per output token stops being a budget fallback and starts being the default choice for a large share of production workloads, with the frontier model reserved for the smaller slice of tasks where its edge actually pays for itself.

Cheaper Tokens, Bigger Bills

It would be easy to read this price war as straightforwardly good news for anyone building with AI, and in one sense it is. But GeekBlog covered a related paradox back in July: blended frontier token prices have fallen almost 98 percent since 2024, yet corporate AI spending grew an estimated 320 percent over the same stretch. Cheaper tokens do not shrink the bill, they expand what companies are willing to attempt with them. An agent that costs a fraction of a cent per call gets used for tasks that would never have justified the expense a year ago, and the aggregate spend climbs even as the per-token price keeps falling.

That dynamic is likely to repeat here. A developer team that adopts V4-Flash-0731 because it is cheap and fast will not necessarily spend less than it did on GPT-5.6 Luna. It will more likely run more agents, more often, on more tasks, because the marginal cost of trying finally dropped low enough to stop being a consideration at all.

The Distillation Playbook Behind the Speed

DeepSeek’s ability to keep undercutting Western labs on price without falling far behind on capability traces back to a technique the industry has leaned on hard this year: distillation, compressing what a larger model learned into a smaller, cheaper one to run. DeepSeek was showing enterprises the outline of this opportunity as far back as last August, when its early releases demonstrated that a well-distilled model could approach frontier performance at a small fraction of the training and inference cost. Western labs have since adopted the same playbook themselves. Meta’s own Muse Glimmer, a 30 billion parameter agentic model distilled down to run on a single consumer GPU, is effectively the same bet from the other side of the market: keep the expensive flagship closed and monetized, and ship a smaller, faster, cheaper model to compete for the high-volume agentic workloads where price sensitivity is highest.

The result is a market where the meaningful competition is no longer only about who has the smartest model. It is about who can get a model that is smart enough down to a price and speed that make it the obvious default for everyday agent work, and DeepSeek has spent the better part of two years proving it can do that faster than almost anyone else.

What It Means If You Are Building or Buying AI Tools

• Route by task, not by habit. If your agents are doing routine tool calls, code generation, or retrieval work, a model like V4-Flash-0731 is now cheap and fast enough to handle the bulk of it, with a frontier model reserved for the fraction of tasks that genuinely need the extra reasoning headroom.
• Watch total spend, not the per-token rate. A lower price per token has not historically translated into a lower bill industry-wide. Set usage budgets and monitoring before rollout, not after the invoice arrives.
• Check compatibility before switching. DeepSeek’s adoption of the Responses API format lowers the migration cost for teams already built around OpenAI’s agent conventions, but data residency, moderation policies, and support commitments still differ meaningfully between providers.
• Expect more cuts, not fewer. Two aggressive moves in one week suggest neither company sees this as a settled market. Locking into long-term pricing assumptions right now is risky.

The Bottom Line

OpenAI’s 80 percent price cut was meant to be a statement. DeepSeek’s answer, delivered within 48 hours using a model it had already built, turned that statement into an opening bid. What is actually being contested here is not really about Luna or V4-Flash as individual products. It is about who controls the economics of agentic AI as it moves from expensive experiment to everyday infrastructure, and neither side looks ready to let the other set the floor.