Gemini 4 Argon opened the gate. Now the million-token model needs ordinary receipts

THE SIGNAL IN ONE SENTENCE

A locked gate is easy to understand. An open one needs rules for traffic. Google announced Gemini 4 Argon on September 30, but initially limited the model to selected cyber defenders and trusted testers. On October 13, that changed. Google made Argon generally available to developers and enterprises through the Gemini API and AI Studio under the stable model name gemini-4-argon. It also brought the model into consumer and agent products through several paid plans. That is a material new event. The model has moved from a capability announcement into something ordinary developers can put on an expense report and point at production. The plain signal is that availability turns impressive claims into testable obligations. Argon's headline feature is a unified window of 1,048,576 tokens. Input, internal thinking and output share that window. The default synchronous output limit is 65,536 tokens, while a long-decode mode can continue in the background toward the unused portion of the full window. One million is a ceiling, not a recommendation. A team could use the extra room to inspect a large codebase, compare a stack of contracts, trace a scientific question through many sources or keep an agent working through a complicated migration. It could also pay for a very long wrong answer. The API guide makes several constraints clearer than the launch spectacle. Thinking tokens count toward the output bill. Introductory pricing through November 30 is two dollars per million input tokens and ten dollars per million output tokens, including thinking. On December 1, Google says those prices double. Long requests may take several minutes. Google recommends background execution for outputs above 65,536 tokens or client timeouts of at least fifteen minutes. This is not chat with a longer scroll bar. It is a small asynchronous job system. The sensible first deployment is not, run the model until it feels finished. It is, give the model one bounded task with a budget, checkpoints and a definition of done. Start with the receipt. For every long-running job, record the model version, input sources, user goal, thinking level, token budget, tool permissions, start time, checkpoints, external actions, human approvals, final artifacts, total cost and repair work. Preserve the failed attempt too. If the result is useful only after an engineer spends three hours untangling it, that labour belongs in the evaluation. The receipt should distinguish model output from system evidence. Google reports that Argon performs strongly on coding, automation, legal, finance, video and cybersecurity evaluations. It describes internal agents that found memory optimizations, helped migrate C and C++ code to Rust and improved a video decoder. Those examples justified attention in September. They remain Google reporting about Google's model inside Google's workflows. General availability gives outside teams a chance to test narrower questions: Did the migration preserve behavior? Which tests failed? How often did the model change direction? What did it cost? Which edits survived review? Did a shorter or cheaper model complete the same work? That final comparison matters. A million-token model should not receive a million-token task merely because the dial exists. Begin with the smallest useful budget. Run a baseline with the current model or human process. Test Argon at low, medium and high thinking levels. Measure completed work, not generated volume. Longer reasoning may improve a hard result. It may also add latency and expense without changing the decision. Use gates inside the run. A long job should periodically restate its objective, list the evidence used, name unresolved assumptions, summarize files or systems touched, report current spending and describe the next intended action. A person or separate policy service can compare that checkpoint with the authorized plan. If the model has drifted, stop there. This is especially important because Argon can call functions, execute code, use search grounding, read URLs and, in preview, use a computer. More reasoning headroom plus more tools creates a larger action surface. A mistaken summary is annoying. A mistaken summary that triggers a database change, deploys code or sends a message is an incident. Tool permissions should shrink as capability grows. Give read-only access by default. Use temporary credentials. Separate development and production. Require human approval before a destructive change, financial commitment, external publication, privilege escalation or message to another person. Log the exact tool request and response, not just the model's later description of what happened. The same discipline applies to data. Argon accepts text, images, video, audio and PDFs, but a large context window does not make every source trustworthy. Documents can be stale, duplicated, contradictory or malicious. Web pages can contain prompt injections. Private files may include data that should never leave a defined project boundary. Build a source manifest before the run. Name each input, owner, date, sensitivity and permitted use. Mark material retrieved from the web separately from approved internal records. When the model produces a conclusion, require links back to the evidence that supports it. Long context can hide a weak source under a mountain of competent prose. Caching deserves attention too. Google says explicit context caching is not supported for Argon. Implicit caching is enabled for sufficiently large prompts and can reduce the input price when a cache hit occurs. Teams cannot assume that every repeated request will hit. Cost estimates should show both cached and uncached cases, and data-governance reviews should explain what the service retains for operation, safety and billing. The migration guide contains other small traps. Sampling controls such as temperature, top-p and top-k are ignored. The default thinking level is medium. A developer who sends an old configuration may believe it still controls behavior when it does not. A team that sets a low output cap may cut off the answer because thinking and visible output share the limit. Treat migration as a new system review, not a model-string edit. Pin the model name. Test the SDK version. Remove unsupported caching calls and ignored parameters. Set hard budgets. Verify retry behavior and cancellation. Confirm how background jobs are resumed, audited and deleted. Test what happens when a tool times out, a source disappears, a person rejects an action or the model reaches its token ceiling. Then run ugly tests. Give the model conflicting instructions, a poisoned document, an unavailable API, a stale file, an ambiguous deadline and a task that cannot be completed with its permissions. Ask it to stop halfway. Revoke a credential during execution. Put a false fact in a long appendix. See whether the system notices, asks for help or confidently builds a cathedral on wet cardboard. Evaluation should include abstention and recovery. Count unsupported claims, unsafe actions, unnecessary tool calls, repeated work, cost overruns, failures caught by humans and failures that reached a downstream system. Measure time to detect and repair. Compare those numbers with the existing workflow. The consumer rollout needs similar humility. Google says paid users can reach Argon in the Gemini app and AI Mode in Search, with Antigravity access reserved to its higher tier. Product access does not give consumers the same logs, controls or evaluation harness available to an API team. People should not assume that a longer-thinking answer is verified merely because it took longer to arrive. For consequential work, ask for sources. Check the original record. Keep personal, legal, health and financial decisions under qualified human authority. Do not upload a sensitive archive simply because the context window can swallow it. Google deserves credit for publishing concrete migration guidance, token accounting and pricing. Those details let teams reason about the product instead of admiring a benchmark mural. But general availability is not a safety certificate. It is the beginning of outside measurement. Argon may prove unusually useful for work that is too large for one short response and too structured for vague conversation. The winning deployments will probably look less dramatic than the launch demos. One bounded task. One source manifest. One checkpoint at a time. One human who can still say stop. The gate is open. Bring a receipt.