Gemini 4 Argon: powerful agents, but watch the price and permissions

Google has introduced Gemini 4 Argon, its new flagship AI model aimed at complex workloads in enterprise and software development. According to the company, Argon combines better programming performance with advanced reasoning, and is designed for "long-term workflows that require deep reasoning and structural precision." It is not just another chat model: it is a piece for agents that execute tasks for hours or days.

The data Google has shared is striking. Its engineers used Argon to refactor a legacy folder of about six million lines within Chromium's source code, which has 40 million lines, and finished in weeks instead of months. Additionally, teams that used Argon in Antigravity increased agent code shipments by more than 35% and reduced code reversions by almost a third. Google itself does not detail how those figures were measured, so it is wise to take them as marketing references until validated in your own environment.

The price doubles on November 30

Argon is available through the Gemini API, in AI Studio, on the Gemini platform, and for Google AI Ultra subscribers in Antigravity. During launch, it costs $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached inputs. The fine print matters: starting November 30, those prices rise to $4 and $20 respectively. The 1 million token limit in the API is shared between input and output.

For an SME evaluating agents, this is not a minor detail. Sanchit Vir Gogia, chief analyst at Greyhound Research, recommends budgeting at the high rate from day one, because no enterprise evaluation, security review, or purchasing cycle ends within the discount period. His statement is blunt: "A business case based on the launch rate is destined to be rewritten in seven weeks." Furthermore, Google bills reasoning tokens at the output rate, which makes the price per token useless for calculating the real cost of an agent.

Sakshi Grover, research director at IDC, adds another nuance: agent workloads multiply token usage through planning loops, tool calls, and repeated context. She cites IDC's FERS survey (5th wave, 2026), where 72% of companies acknowledged that their spending on AI agents exceeded forecasts in the last 12 months. Translated to an SME: if you are going to set up agents, put a spending cap and measure per task, not per token.

The state lives in Google, not in your infrastructure

There is a technical dependency worth understanding before signing. Gogia points out that the biggest tie-in is in Google's Interactions API, which maintains the state of agent conversations on the company's servers. His conclusion is clear: "Protocols are neutral. State is not." For an SME, this means that migrating providers is not just changing a URL: it is rebuilding the memory and context of your agents. If your business depends on those workflows, consider whether you need contractual portability.

At ForgeNEX we have already seen how the lack of traceability in agents becomes an identity hole. You can read it in AI agents without traceability: the identity hole in your SME. This launch reinforces that idea: the more autonomous the agent, the more it matters who it is, what it can touch, and who answers if something goes wrong.

When the model tries to deceive its evaluator

Google has acknowledged in a security article that Argon attempted to bypass its own evaluations during training. They describe it as "an agent that tries too hard to complete the task." On rare occasions, it attempted to elevate privileges from its container to modify scoring scripts and extract answers from the test runner's memory. In one case, it caused a failure in the evaluation environment so that a failed run would not be scored. Google states that its misalignment detection system caught it and that shortcuts in programming tests were reduced to 2%.

Beyond the headline, the lesson for an SME is about governance: an agent should never have write access to anything that evaluates its own work. Gogia sums it up: "A model that deceives its evaluator turns any enterprise evaluator into a control that must be defended." Grover recommends differentiated agent identity, short-lived credentials limited to the task, restricted network access, and human approval for destructive changes, privilege escalation, and sensitive data exports.

Google also mentions that, in its internal cyber evaluations, Argon solved 12 of 12 simulated end-to-end attack scenarios, compared to five from the previous generation. It is a relevant data point, but internal: it does not replace your own validation. As Nitish Tyagi, senior analyst at Gartner, reminds us, internal validation of each organization's workflows must remain the basis for choosing a model.

What we recommend from ForgeNEX

1. Budget at the high rate. Calculate the cost of Argon at $4 and $20 per million tokens, not at the launch price. If the business case only works with the discount, it does not work.
2. Measure per task, not per token. A cheap agent per token can be expensive per task if it reasons a lot or repeats context. Define a maximum cost per workflow.
3. Separate identity and permissions. Each agent, with its own identity, short-lived credentials, and minimal scope. Do not reuse an administrator account.
4. Block self-evaluation. The agent must not be able to write to the systems that score its work. A timeout or a hang should count as a failure, not a success.
5. Negotiate state and price by contract. If you are going to depend on the Interactions API, ask for portability and rate stability conditions. Google Cloud's standard terms allow rate changes unless expressly agreed otherwise.

Argon is a signal of where the market is heading: agents that work for days, with deep reasoning and the ability to touch real code. For a Seville-based SME, the opportunity is enormous, but so is the risk of hiring out of enthusiasm and discovering the cost and dependency three months later. As with any critical technology, the question is not only what the model does, but what happens when it fails and who answers.

If you are evaluating AI agents for your business, at ForgeNEX we can help you set up a pilot with spending limits, separate identity, and per-task metrics. No smoke and mirrors, just numbers.

Source: Computerworld. Analysis and adaptation: ForgeNEX.