Gemini 4 Argon benchmarks
Google published 19 benchmark scores for Gemini 4 Argon in a single comparison table against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, plus three cybersecurity results shown only as charts. Argon has the top score in 13 rows, ties in one and trails in five. This page lists every number, shows who actually ran each test, and puts the losses next to the wins.
Short answer: Argon's headline results are 77.9% on DeepSWE v1.1, 91.7% on LVBench, 51.3% on AutomationBench and 68.9% on the Vals Index, all first place. It ties GPT-6 Astra at 68.0% on CWE-bench v1, and it trails on FrontierSWE v2, Terminal-bench 4.0, PostTrainBench, Terminal-Bench Science 0.1 and OSWorld-2.0.
What "verified" means here. Argon is only rolling out to trusted cyber defenders through Google's Fairwind Program, so developers, enterprises and consumers cannot rerun any of these tests yet. According to Google DeepMind's methodology, 9 of the 19 scores were taken from third-party leaderboards and 10 were run by Google itself. Competitor scores come from a mix of public leaderboards, the providers' own reports and Google's own runs. The table has no MMLU or GPQA results.
All 19 scores in Google's table
Rows and figures as published in Google DeepMind's evaluation document; the best score in each row is highlighted. The last column shows who produced Argon's number, per the same document: green means it was taken from a third-party leaderboard, amber means Google ran it itself.
| Benchmark | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 | Claude Opus 5.5 | Argon's score from |
|---|---|---|---|---|---|
| Knowledge work | |||||
| Vals Index | 68.9% | 63.1% | 65.8% | 67.0% | Vals AI |
| AutomationBench (score) | 51.3% | 41.4% | 31.4% | 42.5% | Zapier |
| Vals Finance Agent v2 | 65.4% | 53.5% | 58.9% | 58.6% | Vals AI |
| Harvey's Legal Agent Benchmark | 19.6% | 5.4% | 6.7% | 3.8% | Vals AI |
| Agentic coding | |||||
| DeepSWE v1.1 | 77.9% | 74.1% | 67.4% | 74.2% | |
| FrontierSWE v2 | 55.0% | 65.5% | 56.3% | 62.3% | Proximal |
| Vibe Code Bench | 91.9% | 89.6% | 90.3% | 90.3% | Vals AI |
| Terminal-bench 4.0 | 57.4% | 58.2% | 57.9% | 66.4% | |
| ML engineering | |||||
| PostTrainBench | 45.3% | 44.3% | 40.2% | 49.3% | |
| Science and math | |||||
| Terminal-Bench Science 0.1 | 57.6% | 68.1% | 52.6% | 63.3% | |
| LABBench 2 | 88.8% | 85.4% | 68.6% | 73.1% | |
| RiemannBench | 76.0% | 72.0% | 65.6% | 69.6% | Surge |
| Long context | |||||
| GraphWalks, up to 128k (BFS F1) | 99.7% | 98.7% | 91.4% | 90.6% | |
| GraphWalks, 256k to 1M (BFS F1) | 84.2% | 71.8% | 65.0% | 66.8% | |
| Computer use | |||||
| Agent's Last Exam (pass rate) | 39.5% | 34.2% | — | 38.2% | |
| OSWorld-2.0 (offline subset, partial score) | 69.2% | 72.6% | — | — | |
| Multimodal understanding | |||||
| Chartography | 71.6% | 71.0% | 46.2% | 66.3% | Surge |
| LVBench | 91.7% | 87.5% | 79.7% | 83.7% | |
| Cybersecurity | |||||
| CWE-bench v1 | 68.0% | 68.0% | 58.0% | 67.0% | CWE-bench |
"—" means Google reports no comparable result: Claude Fable 5.1 is missing from the Agent's Last Exam public leaderboard, and Anthropic reports OSWorld-2.0 only as a combined online and offline score.
Where Argon loses
Five rows of Google's own table go to another model. Two of them are agentic-coding benchmarks, the use case Google leads with, and on both Argon has the lowest score of the four models.
| Benchmark | Argon | Top score | Argon's place | Gap |
|---|---|---|---|---|
| FrontierSWE v2 | 55.0% | 65.5% · GPT-6 Astra | 4th of 4 | −10.5 pts |
| Terminal-bench 4.0 | 57.4% | 66.4% · Claude Opus 5.5 | 4th of 4 | −9.0 pts |
| Terminal-Bench Science 0.1 | 57.6% | 68.1% · GPT-6 Astra | 3rd of 4 | −10.5 pts |
| PostTrainBench | 45.3% | 49.3% · Claude Opus 5.5 | 2nd of 4 | −4.0 pts |
| OSWorld-2.0 (offline subset) | 69.2% | 72.6% · GPT-6 Astra | 2nd of 2 | −3.4 pts |
Two details are worth reading closely. On Terminal-bench 4.0 the bottom three models sit within a point of each other (57.4% to 58.2%), so the gap that matters is the one to Claude Opus 5.5. On Terminal-Bench Science 0.1, Google ran Argon with a 6× verifier timeout to address timeout issues with verification, while the other three scores come from the public leaderboard, and Argon still finished third.
The balance cuts the other way too. Argon's wins are not concentrated in the tests Google ran itself: seven of its 13 outright wins come from third-party leaderboards.
The headline numbers, with the fine print
DeepSWE v1.1: 77.9%
Google's lead coding result, which it calls a new state of the art on real-world, long-horizon software engineering tasks. Argon is 3.7 points ahead of Claude Opus 5.5 (74.2%) and 3.8 ahead of GPT-6 Astra (74.1%). The sources are mixed: Google computed Argon's score itself with a mini-swe agent harness, GPT-6 Astra's comes from the public leaderboard, and both Claude scores come from Anthropic's system cards.
LVBench: 91.7%
The long-video understanding result, 4.2 points ahead of GPT-6 Astra (87.5%). Google ran every model on this one itself, without tools, but not under the same conditions: Gemini was sampled at one frame per second, while GPT-6 Astra got 800 frames, Claude Opus 5.5 600 and Claude Fable 5.1 300, which Google attributes to API limitations. At one frame per second, any video longer than about 13 minutes gives Argon more frames to look at than GPT-6 Astra received.
AutomationBench: 51.3%
Zapier's benchmark for end-to-end execution across core business functions. Google's figure uses the private set and is taken from Zapier's public leaderboard, and Argon leads Claude Opus 5.5 (42.5%) by 8.8 points. Read the absolute number as well as the rank: first place is a competitive statement, while 51.3% is the capability statement.
CWE-bench v1: 68.0%, a three-way tie
Google says Argon "ties for first place", and its table shows the tie with GPT-6 Astra. The CWE-bench leaderboard itself lists three models at 68% pass@1: Grok 4.7, Gemini 4 Argon and GPT-6 Astra. Ties there are ordered by pass@4, the same rule Google's methodology cites, which puts Argon second at 75%, behind Grok 4.7 at 81%. Grok 4.7 is not in Google's comparison set. The same leaderboard puts Argon's average cost per rollout at $6.63, against $2.75 for Grok 4.7, $2.85 for GPT-6 Astra and $0.79 for Claude Opus 5.5.
Vals Index: 68.9%
Vals AI's index of economic impact across finance, coding, legal and tax work, with every sector weighted by its contribution to U.S. GDP. This one can be checked: Vals AI's own leaderboard lists Argon first at 68.90%, ahead of Claude Sonnet 5.5 at 67.04%, a model that is not in Google's table, and Claude Opus 5.5 at 66.97%.
Harvey's Legal Agent Benchmark: 19.6%
Legal research and drafting, with results sourced from Vals AI. Argon's score is nearly three times the next-best model's (Claude Fable 5.1, 6.7%), but in absolute terms it is still under 20%.
Cybersecurity: the charts outside the table
Three more results appear only as charts in Google's announcement. Argon is going to Fairwind participants without cyber guardrails so they can use its full cyber-defence capabilities, and Google says it can autonomously find, validate and patch critical software vulnerabilities. Two of the three results compare it only with Google's previous cyber model, Gemini 3.8 Flash Cyber.
Real-world vulnerability discovery
Scanning source code for vulnerabilities across 20 languages, pass@1, against 71.0% for Gemini 3.8 Flash Cyber. Google's internal dataset tests recall over recent, confirmed historical vulnerabilities in popular open-source projects, run in an internal harness that is not cyber-specialised.
Wiz Penetration Test Benchmark
Wiz's internal benchmark for writing exploits against real web-application vulnerabilities without access to the source code, pass@1, against 58.2% for Gemini 3.8 Flash Cyber.
Gray Swan Indirect Prompt Injection: 0.7% (lower is better)
Results sourced from Gray Swan. The figure is the attack success rate at 15 attempts, so lower is better. Argon has the lowest rate of the 13 models in Google's chart, 0.3 points below Claude Opus 5.5 and Claude Fable 5.1 at 1.0%.
| Model | Attack success rate (k=15) |
|---|---|
| Gemini 4 Argon | 0.7% |
| Claude Opus 5.5 | 1.0% |
| Claude Fable 5.1 | 1.0% |
| Claude Opus 5 | 4.6% |
| Gemini 3.8 Flash | 5.5% |
| Gemini 3.8 Flash Cyber | 6.0% |
| GPT-6 Astra | 8.5% |
| GPT-6 Sol | 10.1% |
| Muse Spark 1.3 | 15.9% |
| GPT-5.6 Sol | 27.0% |
| GLM 5.3 | 31.5% |
| Grok 4.6 | 51.8% |
| Kimi K3 | 52.7% |
Output token limit: 1M, up from 64K
Not a benchmark, but it changes how the model can be used. Google has raised Argon's output token limit to what it calls an industry-leading 1M tokens, up from the previous 64K, so the model has the headroom to think deeply and generate hundreds of thousands of tokens in a single trajectory and solve hard problems in one go.
The cost follows directly. Argon launches at an introductory $2 per million input tokens and $10 per million output tokens, so one response that used the full 1M output allowance would cost $10 in output alone. After the introductory period the price rises to $4 input and $20 output per million tokens, or $20 for the same response.
Results from Google's own engineering work
Google also described four results from using Argon internally. None of them can be compared across models, because there is no shared task and no baseline anyone else can run, but they are concrete.
Quantum subroutine optimisation
Argon is helping Google's quantum computing researchers optimise the spacetime resources (qubits × gates) of subroutines that bottleneck important applications. In one example it beat the published baseline by 40% in a matter of minutes.
Fleet-wide memory savings
A team of Argon agents analysed fleet-wide profiling telemetry and autonomously applied memory optimisations across Google's data centres, freeing over 300 TiB once rolled out, with an estimated 500 TiB to 1 PiB in total savings.
libgav1 in safe Rust
In Google's open-source video decoder, Argon agents took an existing Rust port and replaced 32K lines of SIMD code with safe Rust the compiler could vectorise. The result runs 2.7× faster than the Rust port, with identical video output.
C/C++ to Rust migrations
Argon agents are migrating C/C++ codebases to Rust across Google, from tens of thousands of lines in core libraries like re2 and libgav1 up to more than 800K lines for the Fuchsia Zircon kernel.
How to read these numbers
- Who ran what. Ten of the 19 scores, including DeepSWE v1.1 and LVBench, were run by Google. Nine, including AutomationBench, the Vals Index and CWE-bench v1, come from third-party leaderboards. We checked two of those nine on October 1, 2026, the Vals Index and CWE-bench v1, and both match Google's figures.
- A three-model comparison set. Google's table compares Argon only with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Grok 4.7, which ties on CWE-bench v1, and Claude Sonnet 5.5, second on the Vals Index, are not in it.
- Settings are not identical. LVBench frame counts differ by model, Terminal-Bench Science 0.1 used a 6× verifier timeout for Argon only, Argon's OSWorld-2.0 score is the best of three runs, and the DeepSWE v1.1 scores come from mixed sources. Argon's scores are pass@1 at Google's highest thinking setting unless noted otherwise.
- No general-knowledge tests. There is no MMLU or GPQA score anywhere in Google's materials.
- Nobody else can test it yet. Google says it will make Argon available to developers, enterprises and consumers as soon as possible, starting with paid API customers and Google AI Ultra subscribers. Until then these are the only numbers. When independent results appear they will be added here, and every change will be logged below with its date rather than quietly edited in.
Frequently asked
How many benchmarks does Gemini 4 Argon lead?
In the comparison table in Google's September 30, 2026 announcement, Gemini 4 Argon has the top score in 13 of 19 rows, ties for the top score in one (CWE-bench v1, with GPT-6 Astra) and trails in five: FrontierSWE v2, Terminal-bench 4.0, PostTrainBench, Terminal-Bench Science 0.1 and OSWorld-2.0. The table compares it with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5.
What does Gemini 4 Argon score on DeepSWE v1.1?
77.9%, which Google calls a new state of the art for real-world, long-horizon software engineering. In Google's table Claude Opus 5.5 scores 74.2%, GPT-6 Astra 74.1% and Claude Fable 5.1 67.4%. Google computed Argon's score itself with a mini-swe agent harness; the competitor scores come from the public leaderboard and Anthropic's system cards.
Where does Gemini 4 Argon score lower than other models?
On five rows of Google's own table: FrontierSWE v2 (55.0% vs 65.5% for GPT-6 Astra), Terminal-bench 4.0 (57.4% vs 66.4% for Claude Opus 5.5), Terminal-Bench Science 0.1 (57.6% vs 68.1% for GPT-6 Astra), PostTrainBench (45.3% vs 49.3% for Claude Opus 5.5) and the OSWorld-2.0 offline subset (69.2% vs 72.6% for GPT-6 Astra). On FrontierSWE v2 and Terminal-bench 4.0 it has the lowest score of the four models.
Have Gemini 4 Argon's benchmark scores been independently verified?
Partly. According to Google DeepMind's methodology, 9 of the 19 scores in Google's table were taken from third-party leaderboards, and the two we checked on October 1, 2026 — Vals Index and CWE-bench v1 — match Google's figures. The other 10, including DeepSWE v1.1 and LVBench, were run by Google itself. Argon is not yet available to developers, enterprises or consumers, so no one else can rerun the tests.
What does Gemini 4 Argon score on CWE-bench v1?
68%, tied for first place. The CWE-bench leaderboard lists a three-way tie at 68% pass@1 between Grok 4.7, Gemini 4 Argon and GPT-6 Astra. Ties are ordered by pass@4, which puts Argon second (75%) behind Grok 4.7 (81%). Grok 4.7 is not in Google's comparison table.
What is Gemini 4 Argon's output token limit?
1M output tokens, up from the previous 64K. At the introductory price of $10 per million output tokens, a single response that used the full limit would cost $10 in output tokens; after the introductory period the output price rises to $20 per million.
Sources
- Google, Gemini 4 Argon: our next era of frontier intelligence (September 30, 2026): the announcement, including the cybersecurity charts, availability and pricing.
- Google DeepMind, Gemini 4 Argon model evaluation: approach, methodology & results: the 19-row results table and who ran each benchmark.
- Vals AI, Vals Index leaderboard (version 2.1, checked October 1, 2026).
- CWE-bench leaderboard (v1, checked October 1, 2026).
- Google DeepMind, Fairwind Program: the programme Argon is currently rolling out through.