# Gemini 4 Argon

Google · Proprietary · rank 10 · bench-align-v5

> Every figure below is reproduced as its upstream source published it: nothing is modelled, estimated, interpolated or converted. `Unavailable` means no source published the value — it is never a zero. Each value carries its evidence state and the date it was observed.

Page: https://modelchoice.dev/models/gemini-4-argon  
JSON: https://modelchoice.dev/api/model/gemini-4-argon

## Facts

| Field | Value | Evidence |
| --- | --- | --- |
| Canonical id | `gemini-4-argon` | — |
| Overall score | 79 | Observed 2026-11-21 · source benchlm:models |
| Verified score (BenchLM): a separate BenchLM figure, not the overall score | 77 | Observed 2026-11-21 · source benchlm:models |
| Context window | 1M tokens | Observed 2026-11-21 · source benchlm:models |
| Max output (models.dev) | 32,768 tokens | Observed 2026-11-21 · source modelsdev:catalog |
| Modalities (models.dev) | Input: text, audio, image, PDF, video · Output: text | Observed 2026-11-21 · source modelsdev:catalog |
| Release date | 2026-09-30 | Observed 2026-11-21 · source benchlm:models |
| Access type | Proprietary | — |
| Blended $/1M (75% input / 25% output) | $4.00 | Derived from the input and output rates below |

## Capability evidence

Seven axes from the ranking source. An axis the source did not score is unavailable, not zero. Beside each score is BenchLM's separate verified score for that axis; the two are published side by side and are not interchangeable.

| Axis | Score | Evidence | Verified (BenchLM) | Verified evidence |
| --- | --- | --- | --- | --- |
| Agentic | 80.6 | Observed 2026-11-21 · source benchlm:models | 80.8 | Observed 2026-11-21 · source benchlm:models |
| Coding | 73.9 | Observed 2026-11-21 · source benchlm:models | Unavailable | Unavailable · source benchlm:models |
| Knowledge | 69.1 | Observed 2026-11-21 · source benchlm:models | 69.1 | Observed 2026-11-21 · source benchlm:models |
| Reasoning | 82.6 | Observed 2026-11-21 · source benchlm:models | 82.6 | Observed 2026-11-21 · source benchlm:models |
| Multimodal & Grounded | Unavailable | Unavailable · source benchlm:models | Unavailable | Unavailable · source benchlm:models |
| Instruction Following | Unavailable | Unavailable · source benchlm:models | Unavailable | Unavailable · source benchlm:models |
| Math | 85.6 | Observed 2026-11-21 · source benchlm:models | 85.6 | Observed 2026-11-21 · source benchlm:models |

## Runtime service evidence

Measured values with the date they were observed. Nothing is inferred from a sibling model or a provider claim. Regional or per-endpoint measurements appear only when the API supplies them; none are modelled.

| Measurement | Value | Observed | Last good | Evidence |
| --- | --- | --- | --- | --- |
| Time to first token | 44.54 s | 2026-11-21 | 2026-11-21 | Observed 2026-11-21 · source benchlm:speed |
| Throughput | 69 tok/s | 2026-11-21 | 2026-11-21 | Observed 2026-11-21 · source benchlm:speed |

## Endpoint and price matrix

Every published price component, including cache reads and writes.

| Component | USD | Evidence |
| --- | --- | --- |
| Input / 1M tokens | $2.00 | Observed 2026-11-21 · source benchlm:pricing |
| Output / 1M tokens | $10.00 | Observed 2026-11-21 · source benchlm:pricing |
| Cache read / 1M tokens | $0.10 | Observed 2026-11-21 · source benchlm:pricing |
| Cache write / 1M tokens | $0.19 | Observed 2026-11-21 · source openrouter:pricing |
| Blended / 1M (75% input / 25% output) | $4.00 | Derived — from the input and output rates above; it has no source record of its own |
| Cost per successful task (LiveBench) | Unavailable | Unavailable · source livebench:table |
| Batch input / 1M tokens (OpenRouter batch) | $1.00 | Observed 2026-11-21 · source openrouter:pricing |
| Batch output / 1M tokens (OpenRouter batch) | $5.00 | Observed 2026-11-21 · source openrouter:pricing |
| Batch cache read / 1M tokens (OpenRouter batch) | $0.05 | Observed 2026-11-21 · source openrouter:pricing |
| Batch cache write / 1M tokens (OpenRouter batch) | $0.19 | Observed 2026-11-21 · source openrouter:pricing |

Batch API rate (asynchronous jobs), from OpenRouter's batch listing.

**Pricing note.** Google's Gemini API pricing page lists gemini-4-argon on the paid tier, not the Free tier. Standard input/output is $2.00 / $10.00 per million tokens through November 30, 2026, then $4.00 / $20.00 starting December 1; implicit cached input is $0.10 / $0.20 across those dates, and explicit caching is unsupported. Batch and Flex input/output is $1.00 / $5.00 through November 30, then $2.00 / $10.00 from December 1, with implicit cached input at $0.05 / $0.10. Priority input/output is $3.60 / $18.00 through November 30, then $7.20 / $36.00 from December 1, with implicit cached input at $0.18 / $0.36. An invite-only Contributor tier, whose content is used to improve Google's products, lists $1.00 / $5.00 with $0.05 cached input.

## Workload-aware monthly cost example

10 conversations per day × 8 messages × 22 active days, 1200 input and 400 output tokens per message, no cache. Derived here from the published rates above by this site's own calculator — not a figure any source published.

| Field | Value |
| --- | --- |
| Modelled monthly cost | $11.26 |
| Modelled tokens | 2.82M |

## Benchmark record

77 matched benchmark rows with their published value, unit, and provenance.

### Knowledge

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| LABBench2 (LABBench2: An Improved Benchmark for AI Systems Performing Biology Research) | 88.8 | Nearly 1,900 biology-research tasks | Real-world biology research | [LABBench2: An Improved Benchmark for AI Systems Performing Biology Research](https://arxiv.org/abs/2604.09554) |
| Artificial Analysis Intelligence Index | 50.4 | Cross-benchmark intelligence index | Display-only external reference | [Artificial Analysis](https://artificialanalysis.ai/) |
| AA-HLE (Artificial Analysis Humanity's Last Exam) | 57.1 | Expert-level questions | Frontier expert reasoning | [Artificial Analysis Humanity's Last Exam Benchmark Leaderboard](https://artificialanalysis.ai/evaluations/hle) |
| AA-Omniscience Index (Artificial Analysis Omniscience Index) | 46.1 | Knowledge questions | Broad factual knowledge | [AA-Omniscience: Knowledge and Hallucination Benchmark](https://artificialanalysis.ai/evaluations/omniscience) |
| AA-Omniscience Accuracy (Artificial Analysis Omniscience Accuracy) | 48.3 | Knowledge questions | Broad knowledge | [Artificial Analysis model benchmarks](https://artificialanalysis.ai/models/grok-4-3) |
| AA-Omniscience Hallucination Rate (Artificial Analysis Omniscience Hallucination Rate) | 4.3 | Knowledge questions | Factuality | [Artificial Analysis model benchmarks](https://artificialanalysis.ai/models/grok-4-3) |
| Morphology-to-molecule matching (Axiom Bio morphology-to-molecule matching) | 25.2 | Four multi-channel cell-imaging perturbation-matching tasks | Biomedical image and molecular reasoning | [Claude Opus 5.5 System Card](https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf) |

### Coding

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| Vibe Code Bench (Vibe Code Bench v1.1) | 91.90 | End-to-end web application builds | End-to-end software delivery | [Vibe Code Bench: Evaluating AI Models on End-to-End Web Application Development](https://www.vals.ai/benchmarks/vibe-code) |
| FrontierSWE v2 | 55.0 | 34 ultra-long-horizon engineering and research tasks | Ultra-long-horizon frontier software engineering | [FrontierSWE v2](https://www.frontierswe.com/blog/v2) |
| Bug Hunt Bench | 40.7 | 105 planted bugs across two production TypeScript repositories | Blind production-repository bug finding and repair | [Bug Hunt Bench method, caveats, and definitions](https://bughunt.productcompass.pm/method) |
| AA-SciCode (Artificial Analysis SciCode) | 61.8 | Scientific coding subproblems | Scientific programming | [Artificial Analysis SciCode Benchmark Leaderboard](https://artificialanalysis.ai/evaluations/scicode) |
| DeepSWE | 77.9 | 113 software engineering tasks across 91 repositories and 5 languages | Long-horizon software engineering | [DeepSWE benchmark blog](https://deepswe.datacurve.ai/blog) |
| PostTrainBench v1.1 | 45.3 | Post-training Qwen3 1.7B and 4B, SmolLM3 3B, and Gemma3 4B | Frontier agent evaluation | [PostTrainBench v1.1 leaderboard and methodology](https://posttrainbench.com/?version=v1.1) |
| FrontierCode 2.0 | 27.6 | 132 private software-engineering tasks | Frontier coding-agent quality | [FrontierCode 2.0](https://cognition.com/blog/frontier-code-2.0) |
| ScarfBench (ScarfBench: Cross-Framework Application Migration in Enterprise Java) | 75.2 | 204 framework migration tasks across 34 enterprise Java applications | Whole-application framework migration | [ScarfBench: A Benchmark for Cross-Framework Application Migration in Enterprise Java](https://arxiv.org/abs/2605.06754) |
| PostTrainBench (PostTrainBench v1.2) | 40.4 | Post-training four base models on six benchmarks | Autonomous ML engineering | [PostTrainBench](https://posttrainbench.com/) |

### Mathematics

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| FrontierMath v2 (Tiers 1-3) (FrontierMath v2 Tiers 1-3) | 85.965 | 285 private advanced mathematics problems | From olympiad-plus to early research mathematics | [FrontierMath v2 benchmark hub](https://epoch.ai/benchmarks/frontiermath-tiers-1-3-v2) |
| FrontierMath v2 (Tier 4) (FrontierMath v2 Tier 4) | 78.049 | 41 private extreme-difficulty mathematics problems | Research-level mathematics requiring hours or days of expert work | [FrontierMath Tier 4 v2 leaderboard](https://epoch.ai/benchmarks/frontiermath-tier-4-v2?view=graph&tab=leaderboard) |

### Reasoning

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| ARC-AGI-1 (ARC-AGI-1 Semi-Private Evaluation) | 98.50 | Semi-private ARC-AGI-1 evaluation set | Abstract visual reasoning | [ARC Prize leaderboard](https://arcprize.org/) |
| Graphwalks BFS 128K (Graphwalks BFS 0K-128K) | 99.7 | Graph traversal tasks | Algorithmic long-context reasoning | [Introducing GPT-5.4 mini and nano](https://openai.com/index/introducing-gpt-5-4-mini-and-nano/) |
| ARC-AGI-2 (Abstraction and Reasoning Corpus for AGI v2) | 91.25 | Visual pattern completion and abstract reasoning | Expert-level — hardest public reasoning benchmark | [ARC-AGI-2: A Harder General Intelligence Benchmark](https://arcprize.org/arc-agi/2/) |
| ARC-AGI-3 (Abstraction and Reasoning Corpus for AGI v3) | 46.265668 | Interactive game-like tasks with hidden rules | Frontier agentic reasoning | [ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence](https://arcprize.org/media/ARC_AGI_3_Technical_Report.pdf) |
| AA-LCR (Artificial Analysis Long Context Reasoning) | 79.7 | Long-context reasoning tasks | Long-context reasoning | [Artificial Analysis model benchmarks](https://artificialanalysis.ai/models/grok-4-3) |
| CritPt (Critical Physics Tasks) | 27.1 | Research-level physics questions | Research-level physics reasoning | [CritPt Benchmark Leaderboard](https://artificialanalysis.ai/evaluations/critpt) |
| GraphWalks BFS 256K–1M | 84.2 | Breadth-first-search graph traversal on 200 problems with context lengths from 256K to 1M tokens. | Frontier agent evaluation | [Gemini 4 Argon evaluation methodology](https://deepmind.google/models/evals-methodology/gemini-4-argon) |
| STEMBench (STEMBench120: six-domain scientific research progress) | 2.44 | 120 research tasks, 20 in each of six STEM domains | Open research targets; no run has reached a full-target success | [STEMBench: six-domain model evaluation](https://www.ulam.ai/stembench) |
| ARC-AGI-3 Provider Adapter (ARC-AGI-3 Semi-Private: Provider Adapter harness) | 97.30 | — | — | [ARC Prize verified ARC-AGI-3 results](https://arcprize.org/leaderboard) |

### Agentic

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| Terminal-Bench 4.0 | 57.40 | 66 professional computer-work tasks | Frontier autonomous knowledge work | [Terminal-Bench 4.0](https://www.tbench.ai/news/terminal-bench-4-0) |
| AA Briefcase (Artificial Analysis Briefcase) | 1486 | Professional knowledge-work tasks | Professional work | [Artificial Analysis Briefcase Benchmark Leaderboard](https://artificialanalysis.ai/evaluations/aa-briefcase) |
| AA AutomationBench (Artificial Analysis AutomationBench) | 77.5 | Business-process automation tasks | Agentic automation | [Artificial Analysis AutomationBench Benchmark Leaderboard](https://artificialanalysis.ai/evaluations/automationbench-aa) |
| GDPval-AA | 1624 | Agentic real-world work tasks | Professional agentic workflows | [DeepSeek-V4 Technical Report](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf) |
| GDPval-AA (GDPval-AA normalized) | 56.2 | Economically valuable tasks | Professional agentic workflows | [Artificial Analysis model benchmarks](https://artificialanalysis.ai/models/grok-4-3) |
| AA Terminal-Bench 4.0 (Artificial Analysis Terminal-Bench v4.0) | 57.1 | Terminal-based agent tasks | Agentic software engineering | [Artificial Analysis Terminal-Bench v4.0 Benchmark Leaderboard](https://artificialanalysis.ai/evaluations/terminalbench-v4-0) |
| GDP.pdf (Artificial Analysis GDP.pdf) | 21.8 | Professional document-production tasks | Professional knowledge work | [GDP.pdf Benchmark Leaderboard](https://artificialanalysis.ai/evaluations/gdp-pdf) |
| AA-AnalystAgent (Artificial Analysis AnalystAgent) | 48.8 | Spreadsheet and document analysis questions | Business and data analysis | [AA-AnalystAgent Benchmark Leaderboard](https://artificialanalysis.ai/evaluations/aa-analyst-agent) |
| OSWorld 2.0 | 69.2 | 108 long-horizon computer-use workflows | Long-horizon professional workflows | [OSWorld2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks](https://arxiv.org/abs/2606.29537) |
| ExploitGym | 39.5 | 898 exploitation tasks | Advanced cybersecurity exploitation | [ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?](https://arxiv.org/abs/2605.11086) |
| AutomationBench | 51.3 | 600 public automation tasks | Long-horizon automation | [Kimi K3: Open Frontier Intelligence](https://www.kimi.com/blog/kimi-k3) |
| Agents' Last Exam | 39.5 | Agent tasks | Advanced agentic work | [DeepSeek V4 Flash 0731 update](https://api-docs.deepseek.com/zh-cn/updates/) |
| Finance Agent v2 | 65.4 | Financial analyst task categories | Professional expert-task agent workflow | [Finance Agent v2](https://www.vals.ai/benchmarks/fabv2) |
| Terminal-Bench 2.1 (Vals) (Terminal-Bench 2.1, Vals AI run) | 57.7 | Difficult terminal tasks | Frontier agentic | [Vals AI Terminal-Bench 2.1, Vals AI run leaderboard](https://www.vals.ai/benchmarks/terminal-bench-2-1) |
| Terminal-Bench-Science 0.1 (6x verifier timeout) (Terminal-Bench-Science 0.1 with sixfold verifier timeout) | 57.6 | 70 expert-curated scientific research workflows | Frontier scientific research workflows | [Gemini 4 Argon evaluation methodology](https://deepmind.google/models/evals-methodology/gemini-4-argon) |
| CWE-bench v1 | 68.0 | Audit and patch held-out C/C++, Go, Java, TypeScript/JavaScript, Python, and Rust repositories | Frontier agent evaluation | [CWE-bench v1 leaderboard and methodology](https://cwe-bench.com/) |
| MazeBench (MazeBench: visual spatial reasoning in a 3D open world) | 14 | 100 hidden gems across 256 rooms in one open world | Long-horizon spatial planning, with single puzzles that can take over one hundred moves | [MazeBench Results](https://mazebench.com/blog?post=maze-bench-results) |
| RoboDojo Sim | 34.2 | Simulated robot-control tasks (count not published) | Embodied agent control | [Introducing Claude Fable 5.2 and Claude Mythos 5.2](https://www.anthropic.com/claude-fable-and-mythos-5-2) |
| CWE-bench (CWE-bench v2) | 65.0 | Audit and patch C/C++, Go, Java, TypeScript/JavaScript, Python, and Rust repositories; 100 public and 150 held-out tasks | Frontier agent evaluation | [CWE-bench v2 leaderboard and methodology](https://cwe-bench.com/) |
| Terminal-Bench 4.1 | 55.50 | 66 professional computer-work tasks | Frontier autonomous knowledge work | [Terminal-Bench 4.1](https://www.tbench.ai/news/terminal-bench-4-1) |
| OSWorld 2.1 Offline (OSWorld 2.1 offline subset (82 tasks)) | 74.7 | 82 offline computer-use workflows | Long-horizon professional workflows | [OSWorld 2.1 release (xlang-ai/OSWorld-V2, tag osworld-v2.1)](https://github.com/xlang-ai/OSWorld-V2/releases) |
| MKA (Multimodal Knowledge Agent Benchmark) | 51.3 | 200 multimodal professional tasks across 10 industry verticals | Expert multimodal knowledge work | [MKA Benchmark Leaderboard](https://artificialanalysis.ai/evaluations/mka) |

### Multimodal & Grounded

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| Chartography (no tools) (Chartography without tools) | 71.6 | 100 specialized chart types | Professional chart reasoning | [Chartography](https://surgehq.ai/blog/chartography) |
| Design Arena Website (Design Arena Website Elo) | 1312 | Website generation comparisons | Design and website generation | [OpenRouter Grok 4.3 benchmarks](https://openrouter.ai/x-ai/grok-4.3/benchmarks) |
| LVBench | 91.7 | Long-form video question answering | Extended temporal reasoning | [Qwen3.8-Max: A New Bar for Coding and Cowork](https://qwen.ai/blog?id=qwen3.8) |
| Blueprint-Bench 2 | 54.4 | Spatial reasoning from blueprints | Agentic spatial reasoning | [Gemini 3.5 Flash launch screenshots](https://x.com/GoogleDeepMind) |
| ExtractBench Pro (ExtractBench Pro: long-tail document extraction challenges (LlamaIndex)) | 30 | 100 synthetic documents, 2,622 pages, 86 document types, six challenge groups | The best system completes 58 of 100 documents perfectly; several complete none | [ExtractBench Pro leaderboard and challenge descriptions](https://www.extractbench.ai/pro) |
| ParseBench-2 (ParseBench-2: severity-aware document parsing (LlamaIndex leaderboard)) | 85.32 | 4,317 page-level test cases from 2,645 documents in 13 datasets | Charts and visual grounding separate the field; several systems that lead on text and tables score near zero on charts | [ParseBench-2: Severity-Aware Evaluation of Document Parsing](https://www.parsebench.ai/parsebench-2.pdf) |

### external

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| SEC-Bench Pro | 83.4 | Security engineering tasks | Advanced cybersecurity | [Introducing GPT-5.6](https://openai.com/index/gpt-5-6/) |
| Chess Puzzles (Chess Puzzles (Epoch AI)) | 59.0 | 100 programmatically generated chess positions with a single best move | Novel chess positions, single best move | [Chess Puzzles (Epoch AI)](https://epoch.ai/benchmarks/chess-puzzles) |
| Mystery Game Puzzles (Mystery Game Puzzles (Epoch AI)) | 64.0 | 100 single-best-move positions from an undisclosed game | Unfamiliar game positions, single best move | [Mystery Game Puzzles (Epoch AI)](https://epoch.ai/benchmarks/mystery-game-puzzles) |
| Furniture Assembly (Furniture Assembly Benchmark (Epoch AI)) | 72.5 | 60 photos of three IKEA builds, some with deliberate assembly mistakes | Spatial reasoning from technical diagrams to real photos | [Furniture Assembly Benchmark (Epoch AI)](https://epoch.ai/benchmarks/furniture-assembly) |
| EBR-bench (Earthborne Rangers (EBR-bench)) | 64.8 | Ten playthroughs of an Earthborne Rangers campaign segment per sample (eight learning, two scored) | Learning an unfamiliar game over multi-hour playthroughs | [Earthborne Rangers (EBR-bench), Epoch AI](https://epoch.ai/benchmarks/ebr-bench) |
| MirrorCode (MirrorCode (Epoch AI and METR)) | 58.9 | 30 reimplementation tasks (15 programs in two languages), three attempts each | Days-long software engineering without source code | [MirrorCode (Epoch AI)](https://epoch.ai/benchmarks/mirrorcode) |
| FACTS Multimodal | 54.2 | About 1,500 image questions: 711 public and 811 held out | Image-grounded factual answering | [BenchLM](https://benchlm.ai/benchmarks/facts-multimodal) |
| DAYJOB | 15.7 | 130 long-horizon tasks: 80 finance and 50 healthcare | Long-horizon professional work | [BenchLM](https://benchlm.ai/benchmarks/dayjob) |
| DAYJOB Healthcare | 11.6 | 50 long-horizon healthcare tasks across five subfields | Long-horizon clinical and healthcare operations work | [BenchLM](https://benchlm.ai/benchmarks/dayjob-healthcare) |
| Code2Video | 1539 | Creative briefs rendered as motion-graphics video | Code-driven motion design | [BenchLM](https://benchlm.ai/benchmarks/code2video) |
| Landing Page Arena | 1409 | 12 client briefs for small-business landing pages, rated by human votes | Marketing page design and copy | [BenchLM](https://benchlm.ai/benchmarks/landing-page-arena) |
| Chart Arena | 1522 | 13 chart briefs with small datasets, rated by human votes | Data visualisation | [BenchLM](https://benchlm.ai/benchmarks/chart-arena) |
| Joke Arena | 1479 | 100 joke prompts in 10 categories, rated by human votes | Humour writing | [BenchLM](https://benchlm.ai/benchmarks/joke-arena) |
| Pelican Arena | 1676 | One prompt: an animated SVG of a pelican riding a bicycle, rated by human votes | Vector drawing and animation in code | [BenchLM](https://benchlm.ai/benchmarks/pelican-arena) |
| Landmark Arena | 1596 | 20 famous landmarks built as three.js web pages, rated by human votes | 3D scenes in browser code | [BenchLM](https://benchlm.ai/benchmarks/landmark-arena) |
| ASCII Arena | 1967 | ASCII drawings of subjects from Google's Quick, Draw! dataset, rated by human votes | Text-character drawing | [BenchLM](https://benchlm.ai/benchmarks/ascii-arena) |
| Game Arena Go | 1686 | Head-to-head Go games on a 13 x 13 board | Strategic planning and rule-following | [BenchLM](https://benchlm.ai/benchmarks/game-arena-go) |
| NYT Connections (NYT Connections LLM Benchmark (extended version)) | 98.6 | 940 NYT Connections puzzles, each with up to four added trick words | Wordplay, word knowledge and decoys that fit almost-but-not-quite | [NYT Connections LLM Benchmark (Lech Mazur, GitHub)](https://github.com/lechmazur/nyt-connections) |
| StealthMark (StealthMark Intelligence Benchmarks (overall score)) | 67.4 | 8 original tasks: Newton's Mirror, Ground Truth, Latch, Foresight, Mona Lisa, Selfie, Beach and Raccoon | Chained physical reasoning, planning without tools and precise visual construction | [StealthMark: The AI Benchmark Explained (Stealth Models)](https://stealthmodels.com/stealthmark) |
| FrontierPhysics (End-to-End Physics Research Benchmark for AI Agents) | 8.3 | 56 research projects from 42 physicists across nine fields, three trials each (168 rollouts per setup) | Open-ended research projects that took their authors about 270 days on average | [FrontierPhysics: Evaluating Agents for End-to-End Physics Research (arXiv, November 4, 2026)](https://arxiv.org/abs/2611.06910) |
| HoneyBench (HoneyBench v0.2 (Goodhart Labs)) | 63.6 | 20 private honeypot tasks across 7 domains with exploitable evaluators | Agentic alignment and specification gaming | [HoneyBench v0.2](https://goodhartlabs.com/honeybench) |

### Instruction Following

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| Gray Swan IPI (15 attempts) (Gray Swan Indirect Prompt Injection, attack success at 15 attempts) | 0.7 | Indirect prompt-injection attacks | Adversarial prompt injection | [Gemini 4 Argon evaluation methodology](https://deepmind.google/models/evals-methodology/gemini-4-argon) |

## Independent scores (Epoch AI)

A second, independent score source: Epoch's own figures in Epoch's units, one line per model version Epoch evaluated (effort setting as Epoch labels it); higher is better. Not part of this site's capability scores or rank.

| Benchmark | Variant | Score | Unit | Epoch date | Evidence |
| --- | --- | --- | --- | --- | --- |
| GPQA Diamond | high | 0.964 | fraction (0–1) | 2026-10-13 | Observed 2026-11-21 · source epoch:benchmarks |
| FrontierMath Tiers 1-3 (v2) | high | 0.8596 | fraction (0–1) | 2026-10-13 | Observed 2026-11-21 · source epoch:benchmarks |
| FrontierMath Tier 4 (v2) | high | 0.7805 | fraction (0–1) | 2026-10-13 | Observed 2026-11-21 · source epoch:benchmarks |
| SimpleQA Verified | high | 0.76 | fraction (0–1) | 2026-10-13 | Observed 2026-11-21 · source epoch:benchmarks |

Source: [Epoch AI, CC BY 4.0](https://epoch.ai/benchmarks) ([licence](https://creativecommons.org/licenses/by/4.0/)). Epoch AI, ‘Capabilities & benchmarking’. Published online at epoch.ai. Retrieved from https://epoch.ai/benchmarks.

## Lifecycle and limitations log

Lifecycle events the source associates with this model.

| Event | Type | Confirmation | Announced | Effective | Replacement | Source |
| --- | --- | --- | --- | --- | --- | --- |
| Gemini 4 Argon | Release | confirmed | Unavailable | 2026-09-30 | — | [Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/) |

## What this profile does not claim

Values are reproduced exactly as their sources published them, in the units those sources declared; none are converted, interpolated, or averaged across providers. Any field marked unavailable was attempted and not returned. Last attempted fetch for this model's score: 2026-11-21 05:17 UTC.
