Google DeepMindreleased 30 Sep 2026
Gemini 4 Argon
91.2
B.O.B.
Category profile
- Real use (hard to game)2×93
- Reasoning & knowledge95
- Math88
- Coding85
- Agents & real work91
- Independent indexes94
Reality check
Lab tests
90
Real use (5 tests)
93
Lab tests and real-use tests agree. It does 3 points better on tests that are hard to train for than on the usual lab tests.
What counts as real use★Field reports
What testers found
Hands-on tests by creators and testers. They are linked and dated, and they don't change the score, because they aren't full leaderboards.
- Worse than its benchmarks
Tested on 6 hard apps
Placed fourth overall (about 54% against Opus 5.5's 58%). Code quality 3.6/5 against Opus 5.5's 4.5/5, and it built backend features that didn't really work (keyboard sync, Fitbit sync, Docker downloads) while presenting them as finished. Best on the emulator and Pokémon game tasks, and far cheaper per app.
AICodeKing · YouTube · 14 Oct 2026
- Worse than its benchmarks
Noahbench, a personal coding benchmark
Called it weak at real-world code and suggested Google is aiming at other uses than the coding race.
Noah (Noahbench) · X · 13 Oct 2026
- Worse than its benchmarks
Staff question real-world performance
Staff with early access said it crushes standard benchmarks but struggles with real coding work and front-end website design. Surge AI's founder compared it to top SAT scores without real skills. Google disputes this.
Google employees, via Bloomberg · Report · 1 Oct 2026
- Mixed
Strong on question sets, weaker on real builds
Ranks 1st of 36 on the reasoning and data-analysis question sets, but 27th on design and layout work and 12th on coding.
LiveBench 2.0 suites · Data · 13 Oct 2026
- Better than expected
Box's complex-work evaluation
Box's CEO said it scored 14 percentage points higher on Box's evaluation of complex work.
Box (Aaron Levie) · Company eval
Price per 1M tokens (OpenRouter)
$2.00 in · $10.00 out
Cost per task on the AA index
$1.99
Output speed (Artificial Analysis)
—
§Receipts
Every benchmark result
“Beats” = the share of tracked models this one beat on that test. The raw score is its best reported setting.
| Benchmark | Raw | Beats | Source |
|---|---|---|---|
| Real use (hard to game) · Real use 93 | |||
| LMArena Textas gemini-4-argon-high | 1525 | 100%of 80 | LMArena (blind human votes) |
| Vals Indexas google/gemini-4-argon | 68.9% | 100%of 42 | Vals AI (private test sets) |
| LMArena Agent Arenaas Gemini 4 Argon (High) | 0.136 | 92%of 50 | LMArena (real agent sessions) |
| LMArena WebDevas gemini-4-argon-high | 1678 | 91%of 82 | LMArena (blind human votes) |
| LiveBench 2.0 (real-work suites)as gemini-4-argon | 81.7% | 82%of 39 | LiveBench / Abacus.AI |
| Reasoning & knowledge · Reasoning 95 | |||
| GPQA Diamondas gemini-4-argon_high | 96.4% | 100%of 70 | Epoch AI Benchmarking Hub |
| SimpleQA Verifiedas gemini-4-argon_high | 76.0% | 100%of 57 | Epoch AI Benchmarking Hub |
| Blueprint-Bench 2as gemini-4-argon_unknown | 54.4% | 100%of 31 | Epoch AI Benchmarking Hub |
| Mystery Game Puzzlesas gemini-4-argon_high | 64.0% | 93%of 57 | Epoch AI Benchmarking Hub |
| Chess Puzzlesas gemini-4-argon_high | 59.0% | 90%of 73 | Epoch AI Benchmarking Hub |
| CritPt (physics research)as gemini-4-argon_high | 27.1% | 85%of 89 | Epoch AI Benchmarking Hub |
| Math · Math 88 | |||
| OTIS Mock AIMEas gemini-4-argon_high | 100.0% | 94%of 73 | Epoch AI Benchmarking Hub |
| ProofBenchas gemini-4-argon_unknown | 99.0% | 93%of 68 | Epoch AI Benchmarking Hub |
| FrontierMath Tiers 1-3as gemini-4-argon_high | 86.0% | 84%of 60 | Epoch AI Benchmarking Hub |
| FrontierMath Tier 4as gemini-4-argon_high | 78.0% | 81%of 52 | Epoch AI Benchmarking Hub |
| Coding · Coding 85 | |||
| SciCodeas gemini-4-argon_high | 61.8% | 98%of 86 | Epoch AI Benchmarking Hub |
| CursorBenchas gemini-4-argon_high | 52.9% | 88%of 17 | Epoch AI Benchmarking Hub |
| FrontierSWEas gemini-4-argon_high | 55.0% | 79%of 20 | Epoch AI Benchmarking Hub |
| FrontierCodeas gemini-4-argon_medium | 47.5% | 75%of 45 | Epoch AI Benchmarking Hub |
| Agents & real work · Agents 91 | |||
| APEX-Agentsas gemini-4-argon_unknown | 82.4% | 100%of 48 | Epoch AI Benchmarking Hub |
| Vending-Bench 2as gemini-4-argon_unknown | 13718.2% | 96%of 47 | Epoch AI Benchmarking Hub |
| GDP.pdfas gemini-4-argon_high | 28.4% | 78%of 37 | Epoch AI Benchmarking Hub |
| Independent indexes · Indexes 94 | |||
| Epoch Capabilities Indexas Gemini 4 Argon | 164.6 | 95%of 81 | Epoch AI |
| Artificial Analysis Intelligence Indexas Gemini 4 Argon (high) | 53 | 93%of 54 | Artificial Analysis |