OpenRouter AI model rankings 2026: this week’s GPT, Claude, Gemini, and DeepSeek

In one line: the name leading usage and the name leading output price are unlikely to be the same model this week. The weekly board sums tokens, the catalog prices each million tokens, and request share is a third number. Fold the three together and it is easy to read them as “this lab already won.”

What follows lines up what OpenRouter’s “this week” counts and what the catalog price counts, then puts GPT, Claude, Gemini, and DeepSeek on one comparison: tokens and week-over-week change inside the published top 10, plus current input and output prices.

What the two numbers count

The OpenRouter rankings page is labeled “Usage data through Oct 18, 2026.” “This week” is the seven days ending on the latest complete UTC day, and the change compares those seven days with the seven before them. With that cutoff, the window is 12–18 October. Price is not a column on that board. It is the list price read from the public model catalog, /api/v1/models, on 2026-10-19, in US dollars per million tokens, input and output separate.

Keep three limits next to every table below:

• Not global share. First-party usage in ChatGPT, claude.ai, and the Gemini app is not here.
• Tokens do not share one ruler across labs. Counts come from each provider’s tokenizer. The “T” below is there so you can match the page.
• “Top 10” means the ten rows on the published weekly board. A model outside those ten is cited only when it appears on Today, This Month, or New & Trending. No rank is invented for a row the page did not show.

This week’s top 10: the four labs fill four cells

Of the four labs, four cells make this top 10: two DeepSeek rows, one GPT, one Claude. Claude Opus 5.5 is 6.18T. No Gemini row. The leader is not a new model dropping in from nowhere. It is a Flash snapshot that is still first and down on the week.

Usage next to price

The next table adds only the rows from these four labs inside the published weekly top 10. The prior week is reversed from the rounded tokens and the percentage on the page, kept at the 0.1T magnitude. “1M in + 1M out” multiplies the catalog input price and output price by one million tokens each, so the prices share a column. It is not a quote for a real session. Real sessions are usually heavier on output, which widens the gap.

Those two DeepSeek rows total 37.5T, still the largest block of the four, and the direction inside this top 10 is down, almost all of it on V4.1 Flash. GPT-6 Luna and Claude Opus 5.5 are close in tokens and about 40× apart on output price: $0.50 versus $20. Opus output is about 167× the V4.1 Flash output price.

Models outside the top 10 that still show up on the other tabs, or that people still pin, spread the catalog further. Input prices are in the same units, US dollars per million tokens.

DeepSeek: the week leader and the day leader are not the same id

The weekly number one is still the undated DeepSeek V4.1 Flash, 32.8T, −20%. Reversed from the rounded page figures, the prior week is about 41T, roughly 8.2T lower now. V4 Flash 0731 is still tenth, 4.7T, −4%, about 0.2T lower. Inside this top 10 the family is net down. The new snapshot has not yet overwritten the old row.

On the day board, the latest complete day, number one is already DeepSeek V4.1 Flash 1017, 3.04T, +345%. The undated snapshot is second, 2.62T, −36%. New & Trending marks 1017 as new, at 3.72T for the week. 3.04T is about 82% of 3.72T, so most of that new snapshot’s week landed on the latest day. It has not yet rewritten the weekly number one.

Price does not follow “same family.” 1017 is $1.20 per million output tokens. The undated Flash is $0.12, a 10× output gap. Input is $0.30 versus $0.0472, about 6.4×. If a route slides from the undated id onto 1017, the usage line still says DeepSeek and the output bill is about 10×.

GPT: only Luna is in the top 10

The published weekly top 10 contains one OpenAI row, GPT-6 Luna, 6.21T, −12%. The reversed change is about 0.85T down. On the day board it is eighth, 644B, −2%: same direction as the week, a smaller one-day move. It is still OpenAI’s only row in this top 10.

On the month board GPT-6 Luna is 22.3T, marked new. GPT-5.6 Luna is 18.8T, −61%, and it is not in this week’s published top 10. The more expensive GPT-6.1 Sol, $10 output and $12 for 1M in + 1M out, is not in the top 10 either. “GPT fell this week” is supported if it means Luna. It is not supported as a statement about every OpenAI row on the price list, because this top 10 does not contain those rows.

Luna’s output price is $0.50, the same catalog price as Claude Haiku 5.5, and about 4.2× the V4.1 Flash output price. 1M in + 1M out is $0.60, against $0.167 for Flash. GPT-5.6 Luna is dearer, $1.20 output and $1.40 for the same 1M in + 1M out. Its 30-day volume is still large. It just did not make this week’s top 10.

Claude: the one in the top 10 that is still up

The top 10 has one Claude row, Claude Opus 5.5, 6.18T, +23%. Reversed from the rounded figures, that is about 1.2T more. Among the four labs’ rows in this top 10, it is the only one still up. On the day board it is sixth, 822B, +3%. On the request-share chart Anthropic is 4.7% of text requests, with request count up 25%. The share is still small. The direction matches the tokens.

Claude Sonnet 5.5 and Claude Haiku 5.5 are not in this top 10. Output prices: Opus $20, Sonnet $10, Haiku $0.50. For the same 1M in + 1M out, Opus is $24 and Haiku is $0.60, which lands on GPT-6 Luna’s catalog price. Each extra 1T of Opus tokens is, on output price alone, about 167× the bill of the same 1T on Flash.

+23% is this model’s own week. It is not “Claude has caught DeepSeek.” The two DeepSeek rows in the top 10 total 37.5T. Opus is 6.18T. A Sonnet-class workload does not get a migration order from this table.

Gemini: absent from the token top 10, second by requests

The published weekly top 10 has no Google model. Neither does the month top 10. On New & Trending, Gemini 4 Argon is marked new, at 767B this week. Catalog price is $2 input and $10 output, $12 for 1M in + 1M out, the same band as Claude Sonnet 5.5 and GPT-6.1 Sol, not the Flash band. Gemini 3.8 Flash is still in the catalog, $0.825 input and $4.125 output, $4.95 for the same pair, and this week’s top 10 gives it no token row.

Request share is a different chart. The week beginning 12 October is the same calendar stretch as those seven days, and it counts text requests, not tokens and not spend. Google is 19.7%, with request count up 14%, behind DeepSeek (24.4%, requests up 3%) and ahead of OpenAI (18.5%, requests up 3%). Anthropic is 4.7%.

In that same week, the request table for the 10K context bucket still shows google/gemini-2.5-flash-lite at 150 million requests. That bucket does not explain all of the 19.7%. It is enough to separate two sentences: “no Gemini in the token top 10” and “Google’s requests disappeared.”

Which id to check before you reroute

• Match the snapshot before the lab. DeepSeek’s weekly leader and daily leader are no longer the same id, and the output price differs by 10×. If the default is still the undated V4.1 Flash, check whether traffic is sliding onto 1017.
• Luna’s −12% describes Luna. Do not use it to move GPT-6.1 Sol or GPT-5.6 Luna. Neither is in this top 10, and the month board still shows 5.6 Luna’s 30-day volume shrinking.
• Opus at +23% is about +1.2T. Output price is about 167× Flash. More tokens are not, by themselves, an instruction to make Opus the default.
• Gemini is missing from the token top 10 and still second by requests. Whether it stays on the routing table depends on task hit rate and price, not on this week’s absence. Argon’s 767B is the new-model board, and the list price is already in the Sonnet band.

The 30 September weekly note covers a different week and a top 20. Why price pulls volume apart is in the OpenRouter price note. How a token is counted is in the token and unit-price comparison.

Do not subtract the 30 September piece

That piece closed a week ending 29 September, and its family totals used a top 20. Gemini 3.8 Flash was still in that top 20, and so was GPT-5.6 Luna. This published weekly board stops at ten, and neither model is in it. Do not subtract that week’s 22T from today’s 32.8T and call the difference growth. The window, the cutoff, and the snapshots differ.

What can sit side by side are the facts each piece states. In the 30 September week, GPT-6 Luna was 4.38T inside the top 20. In this week’s published top 10 it is 6.21T. Weeks in between are not in either piece, so this is not a week-over-week change.

Questions

Which model led usage this week? The published weekly board’s first row is DeepSeek V4.1 Flash, 32.8T, −20%. The latest complete day is led by DeepSeek V4.1 Flash 1017, 3.04T.

Which of these four has the highest and lowest output price? In the tables here, the highest output price is Claude Opus 5.5, $20 per million tokens. The lowest is the undated DeepSeek V4.1 Flash, $0.12. The gap is about 167×.

Did GPT rise or fall this week? The top 10 contains only GPT-6 Luna, 6.21T, −12%. On the month board GPT-5.6 Luna is −61%, which is a 30-day figure, and that model is not in this week’s top 10.

Did Claude’s gain catch DeepSeek? No. Opus 5.5 is 6.18T. DeepSeek’s two rows in the top 10 total 37.5T. By request share Anthropic is 4.7% and DeepSeek is 24.4%.

How did Gemini do? Neither the weekly token top 10 nor the month top 10 contains a Google model. Request share is 19.7%, with request count up 14%. Gemini 4 Argon is 767B on the new-model board, marked new.

Does this board stand in for ChatGPT, claude.ai, or the Gemini app? No. It counts tokens routed through OpenRouter. The share chart counts requests, not spend.