GOOGLE / GOOGLE DEEPMIND · MODEL RELEASE TRACKER
gemini-4-argon (Gemini API)
Gemini 4 Argon is a Google / DeepMind Gemini 4 frontier model described as excelling at deep, long‑horizon reasoning and agentic workflows. Google positions Argon as the recommended choice for most software‑engineering and knowledge‑work tasks. It was launched prior to Gemini 4 Carbon and is listed with an introductory token price parity referenced alongside Carbon.CURRENT SNAPSHOT4/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Token pricing (introductory)Referenced with an introductory price parity of $2 per million input tokens and $10 per million output tokens (same introductory price cited for Gemini 4 Carbon and Argon).
Not established from the available sources.
ModalityMultimodal: includes vision/video and action-oriented (vision-language-action) capabilities used for robot video labeling and video-to-sim-to-real workflows.
Benchmark result (ERQA)Argon scores 87% on the Embodied Reasoning Question Answering (ERQA) benchmark using pure reasoning (without helper tools); Google reports this result leads other frontier models on that test.
Empirical result (Real2Sim fine-tuning)Google reports synthesizing domain-randomized MuJoCo episodes from one RGB photo to fine-tune a vision-language-action policy, citing a rise in success on the corresponding real setup from 56% to 75%.
Benchmark result (ERQA)Reported to score 87% on the Embodied Reasoning Question Answering (ERQA) benchmark using pure reasoning, leading other frontier models on that test in the report.
Artificial Analysis Intelligence Index scoreReported score of 50 for Argon at the default 'high' thinking level on the Artificial Analysis Intelligence Index (Hydrogen reported 51 at the same level).
AvailabilityAvailable to Google AI Ultra subscribers and through the paid Gemini API; listed with model id gemini-4-argon in the Gemini API models documentation.
RELEASE TIMELINE
Published, source-backed release events only.
VERIFIABLE FACTS
Every value stays attached to a source and date.
RELEASE · Release statusDEVELOPER CLAIM
Described as having been launched prior to Gemini 4 Carbon (referred to as a recent launch).
CAPABILITIES · Capabilities (robotics and simulation)DEVELOPER CLAIM
Framed as a general-purpose physical-reasoning model for robot control, Real2Sim scene construction, multimodal labeling of robot video, and iterative policy improvement. Demonstrated tasks include directing a Boston Dynamics Spot through a foot-pedal task after practice in a MuJoCo twin, steering a robotic arm with visual coordinates and HTTP motion primitives, and closing loops with the Gemini Omni video world model.
MODALITIES · ModalityDEVELOPER CLAIM
Multimodal: includes vision/video and action-oriented (vision-language-action) capabilities used for robot video labeling and video-to-sim-to-real workflows.
BENCHMARKS · Benchmark result (ERQA)DEVELOPER CLAIM
Argon scores 87% on the Embodied Reasoning Question Answering (ERQA) benchmark using pure reasoning (without helper tools); Google reports this result leads other frontier models on that test.
BENCHMARKS · Empirical result (Real2Sim fine-tuning)DEVELOPER CLAIM
Google reports synthesizing domain-randomized MuJoCo episodes from one RGB photo to fine-tune a vision-language-action policy, citing a rise in success on the corresponding real setup from 56% to 75%.
AVAILABILITY · AvailabilityDEVELOPER CLAIM
Available to Google AI Ultra subscribers and through the paid Gemini API; listed with model id gemini-4-argon in the Gemini API models documentation.
LIMITATIONS · Limitations (reporting)DEVELOPER CLAIM
Google uses long, task-specific prompts and scaffolding in demonstrations and does not publish per-demo failure rates beyond a single VLA case.
CAPABILITIES · Primary capabilitiesDEVELOPER CLAIM
Framed as a general-purpose physical-reasoning model for robot control, Real2Sim scene construction, multimodal labeling of robot video, and iterative policy improvement.
BENCHMARKS · Benchmark result (ERQA)DEVELOPER CLAIM
Reported to score 87% on the Embodied Reasoning Question Answering (ERQA) benchmark using pure reasoning, leading other frontier models on that test in the report.
CAPABILITIES · Positioning relative to Gemini 4 CarbonDEVELOPER CLAIM
Google describes Argon as the ideal choice for the majority of software engineering and knowledge work, positioning it as complementary to Gemini 4 Carbon (which is targeted at more demanding, long-horizon tasks).
LIMITATIONS · Reported evaluation / transparency notesDEVELOPER CLAIM
Published demos and results used long, task-specific prompts and system-level scaffolding; Google did not publish per-demo failure rates beyond a single VLA case, so reported outcomes depend on the presented scaffolding and evaluation choices.
CAPABILITIES · Model classDEVELOPER CLAIM
Described by Google as a 'frontier' Gemini 4 model.
CAPABILITIES · Reasoning strengthsDEVELOPER CLAIM
Said to excel at deep, long‑horizon reasoning and agentic workflows.
PRICING · Token pricing (introductory)DEVELOPER CLAIM
Referenced with an introductory price parity of $2 per million input tokens and $10 per million output tokens (same introductory price cited for Gemini 4 Carbon and Argon).
CAPABILITIES · Recommended use casesDEVELOPER CLAIM
Google recommends Argon as the ideal choice for the majority of software‑engineering and knowledge‑work tasks.
LIMITATIONS · Relative token consumption vs. CarbonDEVELOPER CLAIM
Google states that Gemini 4 Carbon uses more tokens than Argon for extended reasoning and verification loops (implying Argon consumes fewer tokens for similar tasks).
BENCHMARKS · Artificial Analysis Intelligence Index scoreDEVELOPER CLAIM
Reported score of 50 for Argon at the default 'high' thinking level on the Artificial Analysis Intelligence Index (Hydrogen reported 51 at the same level).
CAPABILITIES · Available thinking-level settingsDEVELOPER CLAIM
Reported to offer medium and high thinking-level settings (contrast: Hydrogen introduced low, medium, high, and xhigh).
WHAT CHANGED
Stored passport versions, without reconstructed history.
BENCHMARKS · Artificial Analysis Intelligence Index score+ Reported score of 50 for Argon at the default 'high' thinking level on the Artificial Analysis Intelligence Index (Hydrogen reported 51 at the same level).
CAPABILITIES · Model class+ Described by Google as a 'frontier' Gemini 4 model.
RELEASE · Release status+ Described as having been launched prior to Gemini 4 Carbon (referred to as a recent launch).
LIMITATIONS · Relative token consumption vs. Carbon+ Google states that Gemini 4 Carbon uses more tokens than Argon for extended reasoning and verification loops (implying Argon consumes fewer tokens for similar tasks).
CAPABILITIES · Reasoning strengths+ Said to excel at deep, long‑horizon reasoning and agentic workflows.
CAPABILITIES · Recommended use cases+ Google recommends Argon as the ideal choice for the majority of software‑engineering and knowledge‑work tasks.
RELEASE · Release− Google has released Gemini 4 Argon (model id gemini-4-argon).
CAPABILITIES · Available thinking-level settings+ Reported to offer medium and high thinking-level settings (contrast: Hydrogen introduced low, medium, high, and xhigh).
PRICING · Token pricing (introductory)+ Referenced with an introductory price parity of $2 per million input tokens and $10 per million output tokens (same introductory price cited for Gemini 4 Carbon and Argon).
AVAILABILITY · Availability− Argon is available to Google AI Ultra subscribers and through the paid Gemini API; the Gemini API models documentation lists the model id gemini-4-argon.
AVAILABILITY · Availability+ Available to Google AI Ultra subscribers and through the paid Gemini API; listed with model id gemini-4-argon in the Gemini API models documentation.
BENCHMARKS · Benchmark result (ERQA)+ Reported to score 87% on the Embodied Reasoning Question Answering (ERQA) benchmark using pure reasoning, leading other frontier models on that test in the report.
LIMITATIONS · Reported evaluation / transparency notes+ Published demos and results used long, task-specific prompts and system-level scaffolding; Google did not publish per-demo failure rates beyond a single VLA case, so reported outcomes depend on the presented scaffolding and evaluation choices.
MODALITIES · Modality+ Multimodal: includes vision/video and action-oriented (vision-language-action) capabilities used for robot video labeling and video-to-sim-to-real workflows.
MODALITIES · Modality− Multimodal — application to video/vision, language (prompts), and action/robotics control (vision-language-action workflows).
CAPABILITIES · Positioning relative to Gemini 4 Carbon+ Google describes Argon as the ideal choice for the majority of software engineering and knowledge work, positioning it as complementary to Gemini 4 Carbon (which is targeted at more demanding, long-horizon tasks).
RELEASE · Release− Google has released Gemini 4 Argon.
RELEASE · Release+ Google has released Gemini 4 Argon (model id gemini-4-argon).
CAPABILITIES · Primary capabilities+ Framed as a general-purpose physical-reasoning model for robot control, Real2Sim scene construction, multimodal labeling of robot video, and iterative policy improvement.