google has made gemini 4 argon publicly available. it is google's first public release above the flash tier in more than seven months.
how powerful is argon
the main results fall into four areas:
- competitive performance: gemini 4 argon scores 53 on the artificial analysis intelligence index, matching gpt-6 astra (max). it ranks fifth on arena's agent arena leaderboard, with a +13.2% net improvement.
- lower costs: until november 30, the model costs $2 per million input tokens and $10 per million output tokens. prices will then rise to $4 and $20, respectively. artificial analysis estimates a cost of $1.99 per task at the discounted rates, around 40% less than gpt-6 astra (max) for comparable intelligence.
- improved agentic capabilities: argon performs better on aa-briefcase, automationbench-aa, and terminal-bench 4.0, areas where google's models have historically struggled to keep up with leading competitors.
- low hallucination rate: argon has a 15% hallucination rate on aa-omniscience, the lowest among models scoring 45 or higher on the intelligence index.
artificial analysis says the release puts google back among the top ai labs by intelligence.
where argon stands out
argon's strongest results include long-context tasks, where it scores 84.2% on graphwalks with context lengths of 256k–1m tokens. it also performs well on legal-agent benchmarks and office automation, as well as financial tasks and multimodal evaluations.
robotics is another area google is working on. according to the company's developer blog, argon can control robots, generate simulations from real-world data, and improve its performance through repeated attempts. training on more than 4,500 simulated episodes increased the task success rate from 56% to 75% and cut completion time from 35 to 24 seconds. in another experiment, a so-101 robotic arm reduced the time needed to complete a pick-and-place task from 58.9 to 17.4 seconds over three iterations. that is 3.4 times faster.
where gemini 4 argon falls short
argon still trails leading models on complex agentic coding tasks. it scores 57.4% on terminal-bench 4.0, compared with 66.4% for claude opus 5.5. on frontierswe v2, it scores 55%, behind gpt-6 astra at 65.5%. competitors also lead on individual benchmarks for ml engineering and computer use.
overall, argon looks like a versatile model for information-heavy tasks, workplace automation, and robotics. it has yet to catch up with the leaders in complex agentic software development. after a long gap without a major frontier release its launch marks a significant return for google to the frontier ai race. meta made a similar return with its muse wonder models.
sources
- gemini 4 argon is now available for everyone, google blog – oct 13, 2026; availability, pricing, benchmarks
- gemini 4 argon for robotics, google developers blog – real2sim, so-101 self-improvement
- gemini 4 argon analysis, artificial analysis – intelligence index, cost per task, hallucination rate
