Evidence at a glance
Broader Access Changes Where the Competition Plays Out
On October 13, 2026, Google DeepMind made Gemini 4 Argon globally available to developers, businesses, and consumers. Designed for coding, knowledge work, enterprise workflows, and cyber defense, it can be accessed through the Gemini API, Gemini app, AI Mode in Search, and Antigravity. When Google first introduced it on September 30, access was largely limited to trusted cybersecurity defenders through programs such as Fairwind.
This is more than a shift from application-based access to general availability. Argon is also available in developer tools including Cursor, Devin, and GitHub Copilot, so competition plays out not just in labs and leaderboards but inside existing workflows. For technical leaders, the question is no longer only whether the model is capable. It is also whether it fits the workflow, performs reliably, and remains affordable.
Long-Task Capacity Does Not Mean Every Call Can Produce a Million Tokens
Google's model card describes Argon as a sparse mixture-of-experts Transformer with native text, image, and audio input. Its training data spans web content, code, and multiple media types, while post-training includes instruction tuning, reinforcement learning, and human preference data. Public materials do not disclose the number of experts, model size, or detailed training recipe. The disclosures indicate a design direction, but do not let outside teams reconstruct where its capabilities come from.
Claims about long tasks also need to be read against the API limits. The documentation lists a combined input-and-output ceiling of 1,048,576 tokens, but ordinary synchronous calls are capped at 65,536 output tokens. Longer generations require long decode, while Vals records a maximum of 262,144 output tokens in its test setup. In practice, “a million tokens” should not be treated as a promise that every call can complete in one pass. Task decomposition, latency, and recovery from failures remain part of the system design.
Two Leaderboards Answer Different Questions
In results current through October 13, Vals gives Argon a Vals Index score of 68.90%, placing it first among 48 models. Claude Sonnet 5.5 follows at 67.04% and Claude Opus 5.5 at 66.97%. The index combines tasks across fields such as finance, coding, law, and tax, weighted by each industry's share of US GDP. It measures performance across that particular task mix, not coding alone and not workplace productivity directly.
CursorBench offers a narrower view, and does not show Argon in the lead. On Cursor's current public CursorBench 4.0 leaderboard, Argon High scores 52.9% and ranks sixth, below Opus 5.5 Max at 57.8% and Sonnet 5.5 Extra High at 53.1%. The same listing records a $4.57 cost, an average of 109,500 tokens, and 118 steps for Argon High. AINews also cites Cursor's 52.8% result at default effort as level with Opus 5.5. That is a different setting from Argon High on the 4.0 board, so the two results should not be collapsed into a claim of a tied lead.
Internal Case Studies Show Potential, Not a Productivity Measurement
Google says Argon helped its team separate core web-page handling from a legacy Chromium directory containing six million lines of code, compressing work that had taken months into weeks and eliminating a category of browser crashes. Google also reports 35% more agentic code submissions and nearly one-third fewer rollbacks. These examples are relevant to engineering teams because they concern real development work, not just question-answering scores.
Their evidentiary limits matter just as much. These are internal deployment claims from Google, not controlled productivity studies. The public account does not provide a complete measurement method that would establish how widely the results apply, and it cannot guarantee the same gains in other codebases. A separate Apollo robot demonstration has a similar caveat: researchers describe building a cornhole simulation from human video, making attempts on the physical robot, and refining the throws before achieving seven consecutive successes. It shows a model engaging with a physical task, not execution without testing or adjustment.
Today's Price Is an Entry Point, Not a Long-Term Budget
Argon's introductory API rates are $2 per million input tokens and $10 per million output tokens, with a 95% discount for cached input. Artificial Analysis reports that the rates double after November 30, to $4 per million input tokens and $20 per million output tokens. Vals lists a $15.68 cost for one test run, but that is an estimate under its test conditions, not a fixed charge for every real-world task.
Google is also offering 10,000 scientists the opportunity to apply for six months of free access to Antigravity's highest usage tier. That lowers the cost of experimentation, but does not establish that the same workflow will remain economical after the trial. Teams considering production use should measure token consumption, retries, and human review on their own tasks, then budget against the post-November 30 rates rather than treating a temporary discount as durable unit economics.