The Short Version
Two things happened a month apart and almost nobody has put them together.
In August 2026, DeepSeek raised API prices by 2.3x to 4.5x depending on the model. On 22 September, OpenAI halved GPT-6 Luna and Sol against their GPT-5.6 predecessors.
The result is that DeepSeek, the company whose entire reputation was built on being the cheap option, is no longer the cheap option.
Current List Prices
| Model |
Input / 1M |
Output / 1M |
Cached input |
| GPT-6 Luna |
$0.10 |
$0.50 |
$0.01 |
| DeepSeek V4.1 Flash (peak) |
$0.30 |
$1.20 |
$0.006 |
| DeepSeek V4.1 Flash (off-peak) |
$0.15 |
$0.60 |
$0.003 |
| Grok 4.7 |
$2.00 |
$6.00 |
- |
| DeepSeek V4 Pro (peak) |
$1.32 |
$3.96 |
- |
The Comparison People Are Still Making From Memory
The retired DeepSeek V4 Flash 0731 cost $0.14 in and $0.28 out. That is the number most people still have in their head, and it is the number that built the reputation.
Its replacement, V4.1 Flash, costs $0.30 and $1.20 at peak. That is roughly 2.1x on input and 4.3x on output - consistent with the 2.3x to 4.5x range DeepSeek applied across the line.
Against GPT-6 Luna at $0.10 and $0.50, DeepSeek V4.1 Flash at peak is 3x the input cost and 2.4x the output cost.
On an output-heavy agent workload - which most agent workloads are - that is not a rounding difference.
Where DeepSeek Still Wins
Cached input, decisively. $0.006 per million at peak and $0.003 off-peak, against Luna's $0.01. If your pipeline reuses a large stable prompt across many calls, cached reads may dominate your bill and DeepSeek is the cheaper read.
Off-peak, if your work can wait. $0.15 in and $0.60 out off-peak narrows the gap to 1.5x and 1.2x. For overnight batch work that is close enough that other factors decide.
Self-hosting. Open weights remain open weights. If you run it yourself, list price is irrelevant and this whole comparison is moot.
Where Grok 4.7 Sits
At $2 in and $6 out it is not competing in this tier at all, and should not be read as expensive for it. Grok 4.7 scores 71.0% on DeepSeek v1.1 and 46.3% on CursorBench 4.0 - it is a mid-tier coding model priced against Claude and GPT frontier tiers, not against budget models.
Included here because people bracket it with "the cheap ones", and on current prices it is 20x Luna's input cost.
The Part That Should Give You Pause
DeepSeek reports an annualised revenue run rate of $1 billion, up from under $500 million a few months earlier, and is finalising a raise of roughly $7.5 billion at a $74 billion valuation.
Booked revenue for the first seven months of 2026 was $70.7 million. API gross margin is 82.9%, overall gross margin 44.6%, and the company reported a net loss for the period.
Founder Liang Wenfeng told investors the price increase "had not reduced the customer base." That may be true and it does not tell you how much of the revenue jump came from price rather than growth - raising prices 2.3x to 4.5x while keeping customers produces a revenue chart that looks like demand.
The practical read: this is a company that has demonstrated it will reprice substantially and that its customers mostly stayed. Price it into your planning.
Decision Framework
- High-volume, well-specified, output-heavy work - GPT-6 Luna. Cheapest on both axes and it is not close at peak.
- Heavy prompt reuse, cache-dominated bill - DeepSeek V4.1 Flash. $0.006 cached reads beat everything here.
- Batch work that can run off-peak - DeepSeek at $0.15 / $0.60 is competitive again. Check whether your scheduler can actually target the window.
- Data residency or self-hosting requirements - DeepSeek, on open weights. The pricing conversation does not apply.
- Anything agentic with tool calls in a loop - neither. Luna drops to 20.7% on AutomationBench. Step up to Sol.
- You chose DeepSeek on price before August - re-run the numbers today. The decision you made was made against prices that no longer exist.
Verdict
GPT-6 Luna is the cheapest credible option on list price as of today, by 3x on input and 2.4x on output against DeepSeek V4.1 Flash at peak. DeepSeek keeps the cached-read crown and the self-hosting escape hatch, and its off-peak tier is still competitive.
The wider lesson is about how quickly this market reprices. DeepSeek went up 2.3x to 4.5x in August, OpenAI came down 50% in September, and anyone who set a model choice in July and did not revisit it is now on the wrong side of both moves. Re-check quarterly, not annually.