By Model VS · Guide text updated . The comparison is computed from files in this repository: the current arena-latest.json and livebench-latest.json, and the copies stored under /data/archive/ from the snapshot saved on 24 September 2026. We do not fill a missing day with an estimated rank. Tell us if a row here does not match those files.
Every number on the homepage has a timestamp.
Model VS publishes static JSON that the build has validated. The interactive rankings read those files. If a sync fails, the build keeps the last validated benchmark files instead of inventing rows. That policy is in the methodology. The practical effect is that “current” means “the file we last accepted,” which you can see here and on the leaderboard’s published and snapshot-saved labels.
These are the timestamps in the files this page was built with:
- LM Arena snapshot: leaderboard publish date 2026-10-02, saved October 6, 2026 (2026-10-06T09:00:53.013Z), 100 overall rows, dataset lmarena-ai/leaderboard-dataset, license CC BY 4.0.
- LiveBench snapshot: release 2026-06-25, saved October 6, 2026 (2026-10-06T09:00:12.143Z).
- Specialist snapshot: saved October 6, 2026 (2026-10-06T09:00:11.687Z); SWE-bench Verified dataset timestamp August 16, 2026; GPQA dataset timestamp September 21, 2026.
Use the timestamp that matches the claim. The Arena leaderboard publish date is the date on the source board we copied. The saved time is when our file was written. They are not required to be the same day. LiveBench’s release string can stay still for weeks while we refresh the file. A refresh that does not change task cells is still a refresh. Say that, instead of implying every sync is a new exam.
We do not promise minute-level agreement with the upstream websites. Those sites can update after we save a file. If you are about to ship a claim, open the upstream page, write down what you saw, and check whether our file still matches. The essays on Arena intervals and LiveBench categories are about how to read a file, not about freezing the public internet.
Compare two saved lists, not a vibe about the month.
The archive is the overall slice we had on 24 September 2026. The current file is the overall slice we have now. Both are limited to the rows in those files. A model that Arena shows and that we did not save will not appear as an entrant or an exit. A model that left our file may still exist on Arena under a different id. Ids are matched as strings. A renamed configuration is an exit plus an entrant, which is the honest reading of the files, not a claim that the lab deleted a product.
Comparison baseline: arena archive saved September 24, 2026 (2026-09-24T07:37:44.518Z), leaderboard publish date 2026-09-13. Current file saved October 6, 2026, publish date 2026-10-02. Both files are Model VS copies of the public dataset. They are not a minute-by-minute feed, and the list is the saved overall slice (100 rows then, 100 rows now), not every model Arena has ever shown.
Models in the current file that were absent from that archive: 12. Models in the archive that are absent now: 12. Models present in both whose raw rank changed: 85.
| Source model id | Current raw rank | Current rating | Votes |
|---|---|---|---|
| gemini-4-argon-high | 1 | 1525.22 | 4,932 |
| claude-fable-5-high | 3 | 1504.27 | 38,387 |
| claude-opus-5.5-high | 4 | 1503.70 | 4,552 |
| gpt-6.1-sol-max | 21 | 1483.21 | 3,071 |
| mimo-v2.6-pro | 26 | 1480.08 | 4,056 |
| deepseek-v4.1-flash-max | 38 | 1474.44 | 8,728 |
| claude-sonnet-5.5-xhigh | 45 | 1471.01 | 3,145 |
| gpt-6-sol-max | 64 | 1457.21 | 8,787 |
| Step 5 Preview | 74 | 1453.68 | 2,701 |
| mimo-v2.6-flash | 77 | 1451.78 | 6,053 |
| gpt-6-luna-max | 90 | 1442.88 | 10,014 |
| grok-4.7-xhigh | 91 | 1442.46 | 6,405 |
| Source model id | Archive raw rank | Archive rating |
|---|---|---|
| claude-fable-5 | 1 | 1505.68 |
| gpt-5.2-high | 89 | 1437.58 |
| gpt-5.2 | 91 | 1435.84 |
| deepseek-v4-flash | 92 | 1435.57 |
| longcat-flash-chat-2602-exp | 93 | 1435.49 |
| qwen3-max-preview | 94 | 1434.68 |
| gpt-5-high | 95 | 1434.65 |
| mimo-v2.5 | 96 | 1433.61 |
| glm-5v-turbo | 97 | 1432.68 |
| gemini-3.1-flash-lite-preview | 98 | 1432.36 |
| o3-2025-04-16 | 99 | 1431.55 |
| kimi-k2.5-instant | 100 | 1430.91 |
| Source model id | Archive rank | Current rank | Place change | Rating change |
|---|---|---|---|---|
| glm-5.2-max | 38 | 32 | +6 | +3.67 |
| gemini-3.6-flash-high | 23 | 22 | +1 | +3.12 |
| gemini-3.8-flash-high | 9 | 8 | +1 | +1.77 |
| kimi-k3-max | 17 | 16 | +1 | +3.59 |
| Source model id | Archive rank | Current rank | Place change | Rating change |
|---|---|---|---|---|
| grok-4.3 | 81 | 94 | -13 | -1.24 |
| gemini-3.5-flash-lite | 59 | 71 | -12 | -0.91 |
| gemma-4-26b-a4b | 88 | 100 | -12 | -0.46 |
| glm-5.3-flash | 29 | 41 | -12 | -2.40 |
| minimax-m3 | 84 | 96 | -12 | -1.19 |
LiveBench release stayed 2026-06-25 in the 24 September archive and 2026-06-25 in the current file. Ids added since that archive: claude-sonnet-5-5-max-effort, claude-sonnet-5-5-xhigh-effort, gpt-6.1-sol-max, gpt-6.1-sol-xhigh. Ids removed: union-alpha. For ids present in both LiveBench files, this comparison found 0 task-cell changes. A roster change is not a score change.
A long drop on a fixed list is often displacement.
The saved overall slice is about one hundred rows. When new ids enter and old ids leave, everyone who remains can shift even if their rating barely moves. The tables put the place change next to the rating change so you can see the difference. A model that falls many places while its rating changes by about a point did not “collapse.” It was passed by rows that entered, or it was rearranged among rows whose intervals may still overlap. Read the interval essay before you turn a place change into a performance story.
The reverse is also true. A model that climbs because a crowded band reshuffled is not a new champion. Look at the rating change. If you need to know whether two current neighbors are separable, use the current intervals, not the archive rank. The archive is context for why the homepage looks different. The current file is the file the homepage uses.
We do not publish a narrative of the form “provider X won October.” The files do not say that. They say which ids entered, which ids left, and how the shared ids’ raw ranks and ratings moved between two timestamps. That is the whole claim.
A new id is not a rewritten score.
The generated LiveBench paragraph reports whether the release string changed, which ids were added, which ids were removed, and how many task cells changed among ids that exist in both files. A cell change of zero means the shared rows still contain the same task text as the archive. Added ids can still change who leads Overall, because Overall is recomputed from the rows that are present. Removed ids drop out of that ordering. Neither event edits the historical cells of a model that stayed.
If you saved a screenshot of a category rank in September, check whether that id is still in the file before you repeat the rank. Category rank is a position among the rows we have, not a property stamped on the model forever. The category guide explains how we compute the average. This log only says whether the underlying cells moved between the two files.
Writing changes are not score changes.
On 31 October 2026 the site gained this insights series and a route repair. These are editorial and navigation changes. They do not alter a source rating.
- New indexable essays: Arena intervals, LiveBench categories, SWE-bench flags, and this changelog. They are listed from /insights/ and included in the sitemap. They are meant to be read as arguments about evidence, with numbers inserted from the files.
- /rankings now redirects to the homepage leaderboard. /compare redirects to the on-page compare tool. /radar redirects to /projects/, the AI Radar index. The navigation bar links to those destinations directly, so a menu click does not depend on the redirect.
- AI Radar and Skills stay outside the sitemap and keep
noindex. They are source indexes. Popularity there is not a quality score, scripts are not executed, and they are not the essays we ask a reviewer to count as articles. This paragraph is the indexable explanation. The indexes themselves remain tools. - The homepage and the rankings header continue to show source, publish, and snapshot-saved times from the files. This page is the place where an older file is still visible, so a change in the table can be audited.
Older methodology text was not rewritten to pretend these essays had always been there. How we rate remains the protocol for a private bake-off. The new essays point at it instead of replacing it.
A changelog is not a forecast.
Two files cannot tell you who will lead next month. They cannot tell you the best model for a job we did not test. They can tell you that we did not silently swap a number. If a future month needs another note, it should name its own archive file and its own sync times. Until that file exists, this page should not be updated with guessed movement.
The Chinese toggle on the homepage translates the interface. It does not translate these essays. Linking here from the Chinese UI is a pointer to the English record, not a claim of a second editorial version.