Gemini 4 Argon
Google DeepMind · released
Google's newest and largest Gemini, built to keep reasoning through very long agentic jobs such as codebase migrations, financial and legal research, and finding and patching software vulnerabilities. It is not generally available yet.
Best for: Very long agentic jobs · Finance and legal research · Long-context reasoning
- Intelligence Index
- 52.6
- #5 of 38 · at High effort
- Price per 1M tokens
- $2 in $10 out
- Blended $4 at 3:1
- Context window
- 1M
- Up to 1M output
- Output speed
- —
- $1.99 measured cost per task
Capability profile
One 0–100 score per area, placing this model between the weakest and strongest tracked model across that area's benchmarks.
- Overall95 · #2 of 38
Based on 3 of 3 benchmarks
- Coding90 · #4 of 39
Based on 5 of 5 benchmarks
- Agents & knowledge work90 · #3 of 40
Based on 6 of 6 benchmarks
- Reasoning & science89 · #6 of 39
Based on 2 of 2 benchmarks
- Knowledge & reliability90 · #5 of 39
Based on 2 of 2 benchmarks
- VisionNot enough data
Based on 0 of 2 benchmarks
Benchmark results
All third-party results on record. Rank is among the tracked models that have been tested on that benchmark.
Overall
#5 of 38· High effort· Artificial Analysis
- Vals Index68.9%
#1 of 26· High effort· $15.7 per test· Vals AI
- Arena · Text1525 ±9
#1 of 33· High effort· 4,932 votes· Arena
Coding
- Terminal-Bench 4.057.1%
#4 of 38· High effort· Artificial Analysis
- SciCode61.8%
#3 of 38· High effort· Artificial Analysis
- Arena · WebDev1678 ±15
#6 of 34· High effort· 1,772 votes· Arena
- Vibe Code Bench91.9%
#2 of 26· High effort· Vals AI
- Code Migration69.5%
#2 of 26· High effort· Vals AI
Agents & knowledge work
- GDPval-AA56.3
#11 of 39· High effort· Artificial Analysis
- Arena · Agent+7.6 ±1.9
#6 of 29· High effort· Arena
- Finance Agent v265.4%
#1 of 26· High effort· Vals AI
#4 of 26· High effort· Vals AI
- Legal Research Bench54.8%
#3 of 26· High effort· Vals AI
- Tax Agent Bench76.2%
#2 of 26· High effort· Vals AI
Reasoning & science
- Humanity's Last Exam57.1%
#3 of 39· High effort· Artificial Analysis
- CritPt27.1%
#7 of 39· High effort· Artificial Analysis
- ARC-AGI-2saturatingNot tested
- ARC-AGI-3Not tested
- GPQA DiamondsaturatingNot tested
Knowledge & reliability
#4 of 39· High effort· Artificial Analysis
- AA-LCR79.7%
#23 of 39· High effort· Artificial Analysis
Vision
- MMMU-ProNot tested
- Arena · VisionNot tested
Reported by Google DeepMind
Numbers the lab published at launch, run on its own harness and settings. Useful context, but not independently verified and never used in the rankings on this site.
- DeepSWE v1.1 · highest thinking77.9%
- Terminal-Bench 4.0 · highest thinking57.4%
- AutomationBench · private set51.3%
- OSWorld 2.0 · offline subset, partial score69.2%
- GraphWalks 256K–1M · BFS F184.2%
Specifications
- Released
- Sep 30, 2026
- Availability
- Limited access
- Context window
- 1,000,000 tokens
- Max output
- 1,000,000 tokens
- Accepts
- Text, Images, Video
- Knowledge cutoff
- Not published
- Weights
- Closed
- Reasoning control
- Thinking levels (not yet documented)
Announced 30 September and rolling out first to vetted cyber-defence partners through Google's Fairwind Program. Google says paid API customers and AI Ultra subscribers come next, with no date.
Pricing
- Input / 1M tokens
- $2
- Cached input / 1M
- $0.10
- Output / 1M tokens
- $10
- Blended (3:1)
- $4
- Measured cost / task
- $1.99
Introductory price. Google has said it rises to $4 in / $20 out afterwards, without giving a date.
Worth knowing
- Evaluators tested a pre-release build; scores may shift when the model reaches general availability.
Official sources
Closest alternatives
The models nearest to this one on overall score.
23 benchmarks tracked. See methodology and sources.