- Token growth: DeepSeek. Usage runs through October 9, 2026. On the OpenRouter rankings, DeepSeek V4.1 Flash processed 37.6T tokens in the trailing seven days, up 56%, first on the board. GPT-6 Luna was 6.94T, up 17%, in sixth.
- Request growth: Google. In the week beginning September 28, Google had 21.5% of text requests, and the request count rose 23%. OpenAI was 18.2%, up 20%. DeepSeek still had the largest share, 21.8%, but requests rose only 5%.
- Money: Gemini costs the most, Luna the least. On the makers’ standard endpoints on October 10, the same 880 calls (17.6M in, 3.52M out) cost $3.52 on GPT-6 Luna, $4.75 off-peak and $9.50 at peak on DeepSeek, and $26.40 on Gemini 3.8 Flash. Gemini is 7.5× Luna.
What the board measures
OpenRouter ranks each model by tokens processed through its API. Prompt and completion both count. Usage is bucketed by UTC day. Requests a user or app marks private are dropped before aggregation. Week, day, and month are the trailing 7 days, 1 day, and 30 days, ending on the latest complete day.
Market share is a different chart. It counts text requests, not tokens. OpenRouter also says these boards do not measure users, spend, or accuracy. Each lab tokenizes its own traffic, so 37.6T and 6.94T are not a word-count comparison. The budget still multiplies each lab’s own token price.
The comparison below uses the volume model for each lab: deepseek/deepseek-v4.1-flash, openai/gpt-6-luna, and google/gemini-3.8-flash. GPT-6 Astra and GPT-6 Sol are not in this week’s token top 10. Their prices are listed separately. Anonymous Space Bunny Alpha is second on the weekly board (16.9T, down 49%) and has no public price, so it stays out of the cost table.
Whose tokens grew fastest
Trailing seven days
Usage runs through October 9, 2026. Change is those seven days against the seven before them.
| Model | Weekly rank | Tokens, 7 days | Change |
|---|---|---|---|
| DeepSeek V4.1 Flash | 1 | 37.6T | +56% |
| GPT-6 Luna | 6 | 6.94T | +17% |
| Gemini 3.8 Flash | Not in the top 10 | Below 3.42T | Not listed in this window’s top 10 |
Tenth place is GLM 5.3, at 3.42T, up 34%. Gemini is absent, so its tokens for these seven days are below 3.42T. Of the three headline models, DeepSeek V4.1 Flash has both the most tokens and the fastest growth.
GPT-6 Luna is not the fastest model in this window. In the previous complete week, September 28 through October 4, Teamday’s October 5 fetch of the OpenRouter week showed Luna at 6.18T, up 116%, DeepSeek V4.1 Flash at 25.6T, up 31%, and Gemini 3.8 Flash at 2.18T, up 1%, then 14th. By the window ending October 9, Luna’s volume had risen to 6.94T while its growth rate fell to 17%. DeepSeek rose to 37.6T and its growth rate rose to 56%. Luna’s doubling happened in the last week of September. Early October cooled it.
The latest complete day: DeepSeek V4.1 Flash at 5.65T, up 2% on the day, still first. GPT-6 Luna at 1.26T, down 1% on the day, sixth. Gemini is absent from that day’s top 10 as well.
Trailing 30 days
On the same page, the 30-day board puts DeepSeek V4.1 Flash at 90.5T, marked new. There is no prior 30-day baseline, so there is no percentage. GPT-6 Luna is not in the monthly top 10. Fifth place is still the previous tier, GPT-5.6 Luna, at 37.3T, up 5%. In the week of September 28 to October 4, GPT-5.6 Luna was 4.37T, down 49%, and it has fallen out of the October 9 weekly top 10. A large part of the GPT growth is traffic moving from 5.6 Luna to 6 Luna, not the whole GPT family doubling.
DeepSeek has the same handoff. DeepSeek V4 Flash 0731, eighth on the week, is 5.21T, down 19%. On the month, 0731 is 34.6T, down 32%, and 0423 is 15.2T, down 31%. Inside the family, V4.1 Flash is the one going up. The older snapshots are going down.
Whose requests grew fastest
Market share uses the week beginning September 28. The change is request count against the week before it. That is the latest complete week written on the OpenRouter page.
| Lab | Share of text requests | Change in requests |
|---|---|---|
| DeepSeek | 21.8% | +5% |
| 21.5% | +23% | |
| OpenAI | 18.2% | +20% |
By request count, Google grew fastest of the three, just ahead of OpenAI. DeepSeek still has the largest share, and its request count barely moved. The shares are close. The token totals are not. DeepSeek’s calls are longer, or more of them land on V4.1 Flash. Google’s requests are spread out, and no single Gemini model enters the October 9 weekly token top 10.
“Who is growing fastest” depends on the unit. Count tokens, and the answer is DeepSeek V4.1 Flash. Count requests, and the answer is Google. Reading Google’s +23% as “Gemini 3.8 Flash tokens rose 23%” does not match these two tables.
The three list prices
Prices are the maker’s own standard endpoint from the endpoints API on October 10, 2026, in dollars per million tokens. Third-party routes are cheaper or more expensive, and the quantization differs, so they stay out of the main comparison. OpenRouter’s pricing page still says standard endpoint prices match the maker. The extra on Standard is 5.5% when you buy credits, not a second markup on each token.
| Endpoint | Input | Output | Output vs Luna |
|---|---|---|---|
| openai/gpt-6-luna · OpenAI standard | $0.10 | $0.50 | 1× |
| deepseek/deepseek-v4.1-flash · DeepSeek off-peak | $0.15 | $0.60 | 1.2× |
| Same model · weekday peak | $0.30 | $1.20 | 2.4× |
| google/gemini-3.8-flash · AI Studio standard | $0.75 | $3.75 | 7.5× |
Gemini’s discount field is 0.5. The price written on the endpoint, $0.75 / $3.75, is what you pay. The half-off is already inside that number. It is not another 50% cut on $0.75. When the promotion ends, the number moves.
Faster and slower tiers sit on the same pages and are not in the table. GPT-6 Luna on OpenAI flex is $0.05 / $0.25, and fast is $0.20 / $1.00. Above about 272,000 prompt tokens, the standard input and output prices double. Gemini on AI Studio flex is $0.375 / $1.875, and priority is $1.35 / $6.75. Both still carry a 0.5 discount. GPT-6 Sol on OpenAI’s standard endpoint is $2 / $10. GPT-6 Astra is $10 / $50. Pricing “GPT” as Astra blows the bill up by an order of magnitude.
DeepSeek’s own model listed more than thirty routes that day. The cheapest input quote was $0.016. Several hosts sat on the peak pair, $0.30 / $1.20. Fireworks’ US route was $0.45 / $1.80. Those are other providers. To compare an official call with GPT and Gemini, use DeepSeek’s own endpoint.
What 880 calls cost
The workload matches the September 30 pricing guide: 880 calls, 20,000 input tokens and 4,000 output tokens each, 17.6M in and 3.52M out. Model cost = million input × input price + million output × output price. The table excludes cache reads and the 5.5% fee.
| Endpoint | 880 calls | Vs Luna | Credits ×1.055 |
|---|---|---|---|
| GPT-6 Luna, standard | $3.52 | 1× | $3.71 |
| DeepSeek V4.1 Flash, off-peak | $4.75 | 1.35× | $5.01 |
| DeepSeek V4.1 Flash, peak | $9.50 | 2.7× | $10.03 |
| Gemini 3.8 Flash, standard | $26.40 | 7.5× | $27.85 |
Off-peak DeepSeek is only 35% more than Luna ($4.75 versus $3.52). Peak is exactly double off-peak. Gemini’s input and output prices are both 7.5× Luna, so the whole bill is 7.5×: $22.88 more. Against DeepSeek off-peak, Gemini is 5.6× ($26.40 / $4.75). On output alone, Gemini’s $3.75 is 6.25× DeepSeek’s off-peak $0.60 and 3.125× the peak $1.20.
On flex, the same 880 calls are $1.76 for Luna and $13.20 for Gemini, still about 7.5×. The ratio comes from the unit price, not from the tier name.
BYOK skips the 5.5% while monthly list-price usage stays under $25,000. Use the left column. Above that, the pricing page says 5%. Business is 8%. The free tier is about 50 requests a day, with no platform fee, and it is the wrong place to build a monthly budget.
DeepSeek’s peak window
On October 10, DeepSeek’s own endpoint splits the price in overrides, in UTC. Both weekend days are off-peak. On weekdays:
- 00:00–01:00, 04:00–06:00, and 10:00–24:00 UTC: $0.15 / $0.60
- 01:00–04:00 and 06:00–10:00 UTC: $0.30 / $1.20
When a local midnight crosses a UTC day, use the endpoint JSON. Do not treat the local word “weekday” as the switch.
Cache reads double with the peak: $0.003 per million tokens off-peak, $0.006 at peak. Luna’s standard cache read is $0.01. A long prompt with a repeated prefix will move the bill. The table above leaves that out.
The board is not a model pick
Three things do not follow from this board:
- Volume is not quality. V4.1 Flash is first because it processed the most tokens on OpenRouter. Hard cases and a final review sit on GPT-6 Sol ($2 / $10) or Astra ($10 / $50), and neither is in the weekly top 10. The top 10 says who is taking the bulk of the calls.
- Request count is not the bill. Google’s requests rose 23%, and Gemini 3.8 Flash’s standard output is still 7.5× Luna. More, shorter calls can leave the token board almost still. The invoice still follows tokens.
- The model bill does not include a Mac. Inference is on OpenRouter. Xcode, the simulator, and signing are another line. Do not fold a 16GB Cloud Mac into the $3.52 or the $26.40.
For bulk extraction, classification, and routing, Luna and off-peak DeepSeek are only about a third apart. Let your own pass rate decide. Put Gemini 3.8 Flash in the same eval when you need the multimodal path and can accept about 7.5× the model fee. Peak prices DeepSeek at 2.7× Luna. Batch work that can move should not sit in that UTC window.
Where the numbers come from
- OpenRouter rankings, usage through October 9, 2026: the week, the day, the 30-day board, and lab request share for the week beginning September 28. Attribute the rankings to OpenRouter if you republish them.
- Endpoint prices on October 10, 2026: DeepSeek V4.1 Flash, GPT-6 Luna, Gemini 3.8 Flash. The peak windows are that DeepSeek endpoint’s
overrides. - OpenRouter pricing: Standard 5.5%, Business 8%, BYOK with no extra fee up to $25,000 of list-price inference per month.
- Teamday’s October 5 write-up of the weekly board: Gemini 3.8 Flash at 2.18T, up 1%, for September 28–October 4, and GPT-6 Luna at +116% that week. The October 9 top-10 text does not include that row.
FAQ
Which of GPT, DeepSeek, and Gemini is growing fastest?
By tokens, DeepSeek. Through October 9, 2026, DeepSeek V4.1 Flash was at 37.6T for the trailing seven days, up 56%. By requests, Google, up 23% in the week of September 28, ahead of OpenAI at 20% and DeepSeek at 5%. GPT-6 Luna was 6.94T, up 17%, over those seven days.
How far apart are the prices for the same calls?
On the makers’ standard endpoints on October 10, 880 calls of 20,000 input and 4,000 output tokens cost $3.52 on Luna, $4.75 off-peak and $9.50 at peak on DeepSeek, and $26.40 on Gemini 3.8 Flash. Gemini is 7.5× Luna and 5.6× DeepSeek off-peak. Buying credits multiplies by 1.055.
Why did Google’s requests rise while Gemini is missing from the token top 10?
The week board ranks tokens. Market share counts requests. Google’s requests are close to DeepSeek’s, and they are not concentrated on one model. The October 9 weekly top 10 starts at 3.42T. The previous complete week had Gemini 3.8 Flash at 2.18T, up 1%.
When is DeepSeek at peak price?
Weekdays 01:00–04:00 and 06:00–10:00 UTC, at $0.30 / $1.20. The other listed weekday hours and the weekend are $0.15 / $0.60. Cross a midnight boundary in UTC.
Does the 5.5% platform fee belong in this table?
It is not added to the token price. It is added when you buy Standard credits. BYOK under $25,000 of list-price usage per month skips it. Above that, the fee is 5%. Business is 8%.
Is the token board a quality ranking?
No. It counts tokens on OpenRouter. It does not rank accuracy. Tokenizers differ, so the totals are not a word count. Test your own task.
Close
Three sentences: in the week through October 9, the fastest token growth was DeepSeek V4.1 Flash (37.6T, up 56%); the fastest request growth was Google (up 23%); the same 880 calls price Gemini 3.8 Flash at 7.5× GPT-6 Luna, while off-peak DeepSeek is only 35% more and peak DeepSeek is double its off-peak price.
Endpoint prices move. Before you write a budget, open the three endpoint URLs, check the provider you will actually hit, and multiply your own tokens.
The model bill and the Mac are two lines
The $3.52 to $26.40 above is OpenRouter inference only. Xcode, the simulator, and signing still need a Mac that stays awake. Write the Cloud Mac rent on its own line. Do not fold it into the model month.