OpenRouter — does it really make AI cheaper?
is not a discount store. It is a broker that lets you pick from many companies through one .
Four routes to DeepSeek V4.1 Flash
-
Price listDeepSeek API, directly
The company that made the model sells it
- Price
- input $0.15 · output $0.60 · cache $0.003 (per million tokens) 2× in peak hours
- Fee
- None
- Cache
- 92% of the input hit the
- Menu tool
- 5 of 6 worked
- One answer
- $0.00043 · peak $0.00087
-
Price listOpenRouter → DeepSeek
Same company, same price, plus the top-up fee
- Price
- input $0.15 · output $0.60 · cache $0.003 (per million tokens) 2× in peak hours
- Fee
- 5.5% when you top up
- One answer
- $0.00046 · peak $0.00091 Not measured — worked out from the price list
-
Price listOpenRouter → Open Inference
Another company runs it on its own GPUs (fp4)
- Price
- input $0.02 · output $1.00 · cache $0.02 (per million tokens)
- Fee
- 5.5% when you top up
- Cache
- 0% of the input hit the cache
- Menu tool
- 0 of 6 worked
- One answer
- $0.00044
-
Price listOpenRouter → Sail Research
Another company runs it on its own GPUs (fp4)
- Price
- input $0.10 · output $0.15 · cache $0.0067 (per million tokens)
- Fee
- 5.5% when you top up
- Cache
- 92% of the input hit the cache
- Menu tool
- 4 of 4 worked
- One answer
- $0.00024
What OpenRouter charges
- The provider's price, unchanged
- Top-up fee 5.5%
- What you pay
OpenRouter adds no markup to token prices; it takes a fee when you buy credits. So the same provider costs that much more than calling it directly.
Why it did not get cheaper — four reasons
-
Cache prices differ
charges only $0.003 per million tokens when you resend the same beginning. Open hit 0% cache on short chats, and a hit still costs $0.02.
-
Output is expensive
Open Inference input is cheap at $0.02, but output is $1.00 (DeepSeek $0.60). The longer the answer, the more you lose.
-
The numbers were squeezed
fp4 lowers cost, but the menu tool worked 0 of 6 times, and names and the voice-summary tag were sometimes written wrong.
-
The maker sells it cheapest
DeepSeek charges very little on its own . A company renting GPUs to run the same can hardly beat that.
Lyen chatbot test
We sent the same 12 questions 3 times each with Lyen's own and tools. Cost is worked out from the average of 49 real Lyen chats (input 5,319 · of which cached 3,741 · output 308 tokens).
| DeepSeek , directly | → Open Inference | OpenRouter → Sail Research | |
|---|---|---|---|
| Menu tool | Yes 5 of 6 worked | No 0 of 6 worked | Yes 4 of 4 worked |
| Cache | 92% | 0% | 92% |
| First character | 1 s | 2.2 s | 1.3 s |
| Full answer | 3.3 s | 2.5 s | 5.9 s |
| One answer · off-peak | $0.00043 | $0.00044 | $0.00024 |
| One answer · peak | $0.00087 | $0.00044 | $0.00024 |
An agent like Sonub Code
We imitated 8 steps where the chat keeps growing after a long prompt. One step is worked out as input 15,000 · output 800 tokens.
Fast, cheap models other than DeepSeek
Measured with the same test (6-step long chat · 8 tasks · tool ). Every model finished the tool agent. The tasks were easy, so real coding quality is not settled yet.
When OpenRouter pays off
-
When companies compete
For you can pick the cheapest provider. One answer on Sail Research cost $0.00024, less than DeepSeek direct ($0.00043).
-
DeepSeek peak hours
While DeepSeek doubles its price, most other providers are cheaper.
-
:batchmodelsWork that does not need an answer right away costs about half.
-
:freemodelsFree for tests and light use.
-
Easy to choose
Hundreds of models with one key, another provider when one stops, one bill.
What SONUB decided
Lyen chatbot — DeepSeek direct
The menu tool broke and the cost did not drop, so Lyen stays on DeepSeek direct (2026-10-24). One setting can switch it to OpenRouter.
Sonub Code — OpenRouter
Sonub Code uses OpenRouter so members can choose between models. In our tests DeepSeek direct is cheaper and faster most of the time.
OpenRouter is not "the cheap place" but "the easy place to choose". To save money, pick the provider yourself and always test its and cache.
Sources and how we measured
Prices were checked on 2026-10-24 and change often. We measured by calling OpenRouter and the DeepSeek API directly from the SONUB development server.
- OpenRouter Docs — FAQ (How does OpenRouter make money?)(opens in a new window)
- OpenRouter — Pricing(opens in a new window)
- OpenRouter Docs — Provider selection(opens in a new window)
- OpenRouter Docs — Prompt caching(opens in a new window)
- OpenRouter — DeepSeek V4.1 Flash 1017 providers(opens in a new window)
- DeepSeek API Docs — Models & Pricing(opens in a new window)