State of the AI Race · 01 / 13
The archive opens with OpenAI ahead
OpenAI has the highest provider-frontier rating in this selected artifact at 1274.2. 3 providers are represented; missing observations remain missing.
State of the AI Race · Q3 2026
Anthropic leads the current field.
An eight-part, source-backed briefing on who leads, how close the field is, what the historical archive can tell us, and what operators should verify next.
Finding 01 · Current field
1507.5 leads by 15.9.
Claude Fable 5 is Anthropic’s highest-scoring model in LMArena’s text style-control overall provider-frontier snapshot. Muse Spark 1.1 places Meta second.
This is a point-in-time view dated July 12, 2026. It answers who leads this specific table, not which model is best for every task.
Finding 02 · Uncertainty
The lead is visible. The intervals still touch.
Anthropic’s published confidence interval is 1500.0 to 1515.0. Meta’s is 1482.5 to 1500.7.
Those intervals overlap between 1500.0 and 1500.7. The point estimate ranks Anthropic first, but the uncertainty is part of the result and should travel with it.
Battle counts: 7,856 for Anthropic’s frontier model and 4,498 for Meta’s.
Finding 03 · Chase group
Three providers fit inside 7.0 points.
Meta at 1491.6, Google at 1485.9, and OpenAI at 1484.6 form the immediate chase group.
The small point-estimate spread makes rank alone a weak buying shortcut. Confidence bounds, battle volume, task performance, product access, and operational constraints need separate review.
Finding 04 · Historical archive
LMArena’s lead changed. Its metric changed too.
The replay covers eleven selected snapshots from May 2023 through August 2025. It is historical context, not a continuation of the current table.
11 / 13 · archive ·
Google leads the final archive sample
Google has the highest provider-frontier rating in this selected artifact at 1466.2. 7 providers are represented; missing observations remain missing.
Source: LMArena legacy artifact
Chatbot Arena archive
Sampled provider frontier ratings, May 2023 to August 2025
Meta highlighted. 827 to 1288 published rating; sampled peak 1396.
Line ends, top to bottom:Google 1466Mistral 1433xAI 1431OpenAI 1429DeepSeek 1424Anthropic 1417Meta 1288
Read the race transcript
archive ·
The archive opens with OpenAI ahead
OpenAI has the highest provider-frontier rating in this selected artifact at 1274.2. 3 providers are represented; missing observations remain missing.
1. OpenAI 1274.2 · 2. Anthropic 1224.7 · 3. Meta 827.0
Source: LMArena legacy artifact
archive ·
Anthropic leads as Google enters
Anthropic has the highest provider-frontier rating in this selected artifact at 1200.7. 4 providers are represented; missing observations remain missing.
1. Anthropic 1200.7 · 2. OpenAI 1184.9 · 3. Google 989.0 · 4. Meta 865.2
Source: LMArena legacy artifact
archive ·
Anthropic holds the sampled lead
Anthropic has the highest provider-frontier rating in this selected artifact at 1203.7. 5 providers are represented; missing observations remain missing.
1. Anthropic 1203.7 · 2. OpenAI 1149.6 · 3. Meta 1029.3 · 4. Mistral 1014.0 · 5. Google 977.0
Source: LMArena legacy artifact
archive ·
OpenAI returns to the sampled lead
OpenAI has the highest provider-frontier rating in this selected artifact at 1228.3. 5 providers are represented; missing observations remain missing.
1. OpenAI 1228.3 · 2. Anthropic 1085.4 · 3. Google 1073.0 · 4. Mistral 1068.3 · 5. Meta 1027.2
Source: LMArena legacy artifact
archive ·
OpenAI leads as DeepSeek enters
OpenAI has the highest provider-frontier rating in this selected artifact at 1211.4. 6 providers are represented; missing observations remain missing.
1. OpenAI 1211.4 · 2. Anthropic 1130.3 · 3. Google 1096.5 · 4. Mistral 1080.7 · 5. Meta 1024.6 · 6. DeepSeek 1008.0
Source: LMArena legacy artifact
archive ·
OpenAI leads a closer sampled field
OpenAI has the highest provider-frontier rating in this selected artifact at 1286.7. 6 providers are represented; missing observations remain missing.
1. OpenAI 1286.7 · 2. Google 1267.1 · 3. Anthropic 1248.4 · 4. Meta 1207.2 · 5. DeepSeek 1179.0 · 6. Mistral 1156.9
Source: LMArena legacy artifact
archive ·
OpenAI leads as xAI enters
OpenAI has the highest provider-frontier rating in this selected artifact at 1316.2. 7 providers are represented; missing observations remain missing.
1. OpenAI 1316.2 · 2. Google 1300.9 · 3. xAI 1294.8 · 4. Anthropic 1270.1 · 5. Meta 1266.2 · 6. Mistral 1249.8 · 7. DeepSeek 1219.2
Source: LMArena legacy artifact
archive ·
OpenAI remains first in the sample
OpenAI has the highest provider-frontier rating in this selected artifact at 1361.3. 7 providers are represented; missing observations remain missing.
1. OpenAI 1361.3 · 2. Google 1343.3 · 3. xAI 1289.1 · 4. Anthropic 1282.6 · 5. Meta 1267.2 · 6. DeepSeek 1257.5 · 7. Mistral 1251.4
Source: LMArena legacy artifact
archive ·
Google takes the sampled lead
Google has the highest provider-frontier rating in this selected artifact at 1383.8. 7 providers are represented; missing observations remain missing.
1. Google 1383.8 · 2. OpenAI 1365.3 · 3. DeepSeek 1361.4 · 4. xAI 1287.7 · 5. Anthropic 1283.2 · 6. Meta 1268.0 · 7. Mistral 1251.6
Source: LMArena legacy artifact
archive ·
Google leads as Meta reaches its sampled peak
Google has the highest provider-frontier rating in this selected artifact at 1446.0. 7 providers are represented; missing observations remain missing.
1. Google 1446.0 · 2. OpenAI 1409.1 · 3. xAI 1399.4 · 4. Meta 1395.9 · 5. DeepSeek 1368.5 · 6. Mistral 1339.0 · 7. Anthropic 1296.5
Source: LMArena legacy artifact
archive ·
Google leads the final archive sample
Google has the highest provider-frontier rating in this selected artifact at 1466.2. 7 providers are represented; missing observations remain missing.
1. Google 1466.2 · 2. Mistral 1433.4 · 3. xAI 1430.5 · 4. OpenAI 1429.4 · 5. DeepSeek 1424.4 · 6. Anthropic 1416.6 · 7. Meta 1288.0
Source: LMArena legacy artifact
break ·
The historical line stops here
The selected snapshots cross LMArena rating-method changes, including Bradley-Terry, style control, and frequency reweighting. Endpoint differences are not normalized like-for-like performance changes. The July 2026 table measures something different, so we never draw a line or calculate change across this boundary.
Source: LMArena current dataset
current ·
Anthropic leads the current field
Anthropic has the highest provider-frontier point estimate in the separately reported LMArena text style control overall snapshot at 1507.5. This is not a continuation of the historical series.
1. Anthropic 1507.5 · 2. Meta 1491.6 · 3. Google 1485.9 · 4. OpenAI 1484.6 · 5. xAI 1474.4 · 6. DeepSeek 1457.1 · 7. Mistral 1427.2
Source: LMArena July 2026 artifact
State of the AI Race · 02 / 13
Anthropic leads as Google enters
Anthropic has the highest provider-frontier rating in this selected artifact at 1200.7. 4 providers are represented; missing observations remain missing.
State of the AI Race · 03 / 13
Anthropic holds the sampled lead
Anthropic has the highest provider-frontier rating in this selected artifact at 1203.7. 5 providers are represented; missing observations remain missing.
State of the AI Race · 04 / 13
OpenAI returns to the sampled lead
OpenAI has the highest provider-frontier rating in this selected artifact at 1228.3. 5 providers are represented; missing observations remain missing.
State of the AI Race · 05 / 13
OpenAI leads as DeepSeek enters
OpenAI has the highest provider-frontier rating in this selected artifact at 1211.4. 6 providers are represented; missing observations remain missing.
State of the AI Race · 06 / 13
OpenAI leads a closer sampled field
OpenAI has the highest provider-frontier rating in this selected artifact at 1286.7. 6 providers are represented; missing observations remain missing.
State of the AI Race · 07 / 13
OpenAI leads as xAI enters
OpenAI has the highest provider-frontier rating in this selected artifact at 1316.2. 7 providers are represented; missing observations remain missing.
State of the AI Race · 08 / 13
OpenAI remains first in the sample
OpenAI has the highest provider-frontier rating in this selected artifact at 1361.3. 7 providers are represented; missing observations remain missing.
State of the AI Race · 09 / 13
Google takes the sampled lead
Google has the highest provider-frontier rating in this selected artifact at 1383.8. 7 providers are represented; missing observations remain missing.
State of the AI Race · 10 / 13
Google leads as Meta reaches its sampled peak
Google has the highest provider-frontier rating in this selected artifact at 1446.0. 7 providers are represented; missing observations remain missing.
State of the AI Race · 11 / 13
Google leads the final archive sample
Google has the highest provider-frontier rating in this selected artifact at 1466.2. 7 providers are represented; missing observations remain missing.
State of the AI Race · 12 / 13
The historical line stops here
The selected snapshots cross LMArena rating-method changes, including Bradley-Terry, style control, and frequency reweighting. Endpoint differences are not normalized like-for-like performance changes. The July 2026 table measures something different, so we never draw a line or calculate change across this boundary.
State of the AI Race · 13 / 13
Anthropic leads the current field
Anthropic has the highest provider-frontier point estimate in the separately reported LMArena text style control overall snapshot at 1507.5. This is not a continuation of the historical series.
Finding 05 · Meta case study
827.0 → 1395.9 → 1288.0
Meta’s provider frontier begins at 827.0 in the selected May 2023 artifact, reaches a sampled peak of 1395.9 in May 2025, and finishes the archive at 1288.0 in August 2025.
The final decline is real in the selected archive. When a model's ownership is ambiguous we leave it out, so third-party derivatives do not become Meta through name matching.
These are legacy sampled ratings across methodology changes. They are not comparable to the July 2026 current snapshot.
Finding 06 · Operator implications
Use the table to form a shortlist, not end the decision.
Start with task fit
Test the shortlisted models on representative prompts, failure cases, languages, and output formats from your own workflow.
Price the full system
Compare input, output, caching, tooling, retries, human review, and migration costs. Arena Score does not encode them.
Verify operations
Check latency, rate limits, region availability, data handling, uptime terms, model stability, and exit options before production use.
Recheck the evidence
Use confidence bounds and battle counts, then run an internal evaluation. A close public rank is a reason to test more, not less.
Finding 07 · Limitations
What this report does not claim.
- The current table is a single LMArena text style-control overall snapshot, not a universal model ranking.
- The legacy archive is eleven selected artifacts across methodology changes, not a continuous live series.
- The current snapshot is never appended to the legacy line or used to calculate cross-window change.
- We do not extend the historical line toward the present, because the newer data does not measure the same thing as the archive.
- No composite score is published. Catalog and operational facts do not alter Arena Score.
- The report does not measure price, latency, safety, reliability, access, or performance on a reader’s private tasks.
Read the complete methodology and inspect every receipt in the source registry.
Finding 08 · Corrections and reuse
Built to be checked, cited, and reused.
Every chapter has a stable fragment. Two linked SVGs package the validated chart data for presentations and coverage.
Current provider frontier
Point-in-time snapshot. Do not join this asset to the historical series.
Download SVG
Selected historical race
Legacy sampled archive through August 2025. Methodology changed during the window.
Download SVGCorrection history
- July 14, 2026
- Corrected the Meta legacy account to 827.0, a 1395.9 sampled peak, and a 1288.0 final value. Superseded claims using 1339, +512, or -57 are not used.
- July 14, 2026
- Clarified that the July 2026 point-in-time snapshot is not continuous with the legacy archive and cannot be used for cross-window change.
Report source ledger
12 pinned artifacts support this report.
| Date | SHA-256 | Artifact |
|---|---|---|
| 2026-07-12 | 8884336d5d6307191bacac5f7bab560c9f597908f549b1b542b861f323319abe | Open source |
| 2023-05-08 | 50dbde9cbbadd4939233edf2a25ad2ee345e0ce5af4e310000846bf67a410f5d | Open source |
| 2023-07-17 | ae6feeba2538ea2593da1971df50eaa436ab114ba3d91be39de3cf5c692d293f | Open source |
| 2023-10-02 | d5b3ed8668fd1321e28623fc434269274a70ed011b67513cb3ce785d3309f458 | Open source |
| 2023-12-15 | 0848b15e0a15df023703734666cbd59a50a621b865aa66aefd0b7e066eccbd6d | Open source |
| 2024-03-07 | cd51f95d6c872755bfe782b9a4ccc74a44fe57428776341a315ae1cabf611cd8 | Open source |
| 2024-06-21 | b098f9f69009376d87ef4d317c81c648d0558f7c912f693e285de5d3115e309b | Open source |
| 2024-08-28 | b7da13b5f061a7a5a112e5ca45ff707d6cf6259c8a01b40ea5b77bbd5bd3d5b0 | Open source |
| 2024-11-20 | 3e5f7d30d02d57f07d0506d385d2007fb7ae4a33a24a2e8e62fc95e9434c565e | Open source |
| 2025-02-05 | 53603d4d811e02261964a97bc8cbb46be4380f914388667974ba483e94262e3e | Open source |
| 2025-05-19 | e6c7227bfda19c7aa521edb19907168e211b8748fe113f7a4536a4d2c2beb48c | Open source |
| 2025-08-29 | a5c6c2b0143d1a3bfe59a2a970686afae314d8b32321c294c1b6dd9fe63e2417 | Open source |