Gemini 4 Argon: a flagship reasoning model for complex Agents
Gemini 4 Argon is for hard problems, not every small task. Use it when changing one part of a project could affect many files, or when you need to read a lot of material before making an important decision. Hermes first puts the useful material in order. Then Argon helps find the cause and plan the next step. Give simple work to a faster, cheaper model.
First organize the material. Then let Argon handle the hard part.
It costs more than Flash, so save it for a small number of hard jobs.
MODEL ID
gemini-4-argon
CONTEXT WINDOW
1,048,576 tokens
OFFICIAL INPUT
$2 / 1M → $4
OFFICIAL OUTPUT
$10 / 1M → $20
DEFAULT THINKING
medium
A HERMES ROLE TO PLAN FOR
Use it as a helper for difficult work. Do not make it do everything.
When using Argon in Hermes, state the job in plain terms: what needs to be solved, what material it may use, and what it must not change. Let cheaper models collect material, sort it, and format it. Ask Argon to continue only when the simpler models cannot answer the question or when a mistake would cause a real problem.
HERMES MODEL ROUTING
Do the simple work first. Save Argon for the hard part.
The best workflow does not use the most expensive model every time. Hermes can first use lower-cost models to collect sources, remove duplicates, and list the important points. If the sources disagree, or if changing one file may affect other parts of a project, ask Argon to work through the problem step by step. This costs less and makes the answer easier to check.
Hermes assigns the work. Argon handles the hard part.
01
Handle lots of text first
Use DeepSeek V4.1 Flash to find key facts, group similar items, and clean up source material.
If the earlier steps still cannot answer the question, ask Argon to explain the cause, make a plan, and say how to check the next step.
TASK FIT
When should a Hermes workflow escalate to Argon?
Good fit
Fixing a hard bug, checking how a codebase is designed, reading a long set of materials, solving math or science questions, or jobs that need several tools in sequence.
Do not start here
Bulk extraction, simple grouping, routine rewriting, and the first pass of research. Start with a fast, low-cost model. Send only the difficult cases to Argon.
Always review
Payments, changes to a live site, public messages, legal conclusions, and security conclusions. Argon can prepare facts and options, but a person must make the final decision.
1M IS NOT UNLIMITED SPACE
Reading, thinking, and answering all use the same space.
Think of 1M tokens as one large backpack. The material you send, the model’s hidden thinking, tool results, and the final answer must all fit inside it. More source material leaves less room to think and answer. For a very long answer, use a background task. A normal request can return up to 65,536 tokens.
OFFICIAL API PRICING
Start with less thinking. Spend more only when the job is truly hard.
Through November 30, 2026, Google lists $2 for 1M input tokens and $10 for 1M output tokens. Output pricing also includes tokens used while the model thinks. A cached input is $0.10. From December 1, those rates become $4, $20, and $0.20.
REAL CASE: IMPROVING A LARGE CODEBASE
Argon should not write all the code alone. Its job is to help a team find the problem, plan a safe change, and prove that the change worked.
Google published an ADK Python example with about 205,000 lines of code, nearly 800 files, and roughly 17,000 unit tests. After an 11-hour run, preparing a long task took about 40% less time and the code started about 21% faster. But the first run did not get the project rules. It changed a public data format, called JSON, and missed some unusual test cases. The second run got the rules and the full test suite first. Then it produced a change an engineer could review. The lesson is simple: do not drop a whole project into a model. First say what it must not touch and how you will check the result.
Example: a site starts more slowly after an update. How can Hermes help an engineering team find the reason?
This is a practical Hermes example, not a measured result from Google or Hermes. It shows what Argon should do and what it must not do by itself.
The problem
Find why the site starts more slowly after an update, then write a repair plan an engineer can check. The model must not change the live site.
What Hermes gets first
Speed-test results, error logs, the list of recent changes, relevant folders, public interfaces that must stay the same, project rules, and test commands.
When to ask Argon
Fast models have found the suspicious files, but there are still several possible causes, or changing one file may affect another feature.
What a person must still do
Merge code, change a data format, change permissions, deploy to production, and explain the problem outside the team.
01
Collect the material first
Hermes first asks low-cost models to read the speed tests, error logs, and recent changes. They remove duplicates and keep only likely files, useful clues, places where sources disagree, and interfaces that must not change. Argon does not spend expensive time on basic searching.
02
Ask Argon to explain the cause
Give Argon one clear question: compare the possible causes, explain which files affect one another, and suggest the smallest safe change. It must say why, what could go wrong, which tests to run, and what it still does not know. It may draft a change, but it must not merge it.
03
Check with real tools
Hermes then runs unit tests, interface checks, startup-speed tests, and code-rule checks. A checking Agent looks only at results you can repeat: did speed improve, which tests failed, and did an interface or data format change? If the evidence is weak, ask Argon to keep working. Do not accept the first answer.
04
Let an engineer decide
An engineer reads three things before merging: what we found, what we plan to change, and what the tests showed. Google’s example shows why this matters. Even a strong model can change the wrong thing without project rules and human review.
REAL CASE: SECURITY REVIEW
A model can help find a problem. It must not handle risky actions alone.
Google says Wiz used Argon to find a serious hospital-software bug that earlier models missed. Google also used it to find issues in internal and open-source software. That makes it useful for finding clues and preparing a repair plan. It does not mean the model should fix a bug, deploy a change, or disclose a bug by itself. Hermes should separate scanning, reproducing, testing, and publishing. Give the model only the permissions it needs for the current step. A person must approve changes to a live system, payments, external messages, and security conclusions.
Do not give the model a pile of sources. First turn the pile into a clear set of questions.
First let the Hermes Agent that finds sources remove duplicates, mark stronger sources, and show where sources disagree. Then ask Argon to compare the claims. Ask for a fixed answer: what we can trust, where the proof is, what disagrees, what we still do not know, and who checks next. This helps you trace a conclusion back to its source instead of getting one long summary that only sounds convincing.
Give ArgonFiltered sources, conflicts, a task boundary, and fixed output fields.
Ask it to returnConclusions, evidence, counterevidence, unknowns, and the next verifier.
Never auto-approvePayments, production changes, public publishing, legal, or security conclusions.
COMMUNITY DISCUSSION: TEST IT YOURSELF BEFORE YOU SCALE
Early developer discussion often covers long context, hard code problems, thinking levels, and price. This can help you decide what to test. For example: is medium thinking enough, or does high thinking prevent enough rework to justify the cost? Other people’s reports are not a replacement for your own test. This page should label them as community observations with a date and link, not as a promise.
WHAT TO KNOW BEFORE USING THE OFFICIAL API
Moving to this model is more than changing one name.
The model ID is gemini-4-argon. The default thinking level is medium. Choose low, medium, or high based on the job; start with low or medium for a simple task.
You cannot save a fixed cache yourself. When a prompt is longer than 4,096 tokens, the system may use an automatic cache. Google says a cache hit can reduce input cost by 95%.
The common settings temperature, top_p, and top_k do not work here. For an answer longer than 64K, use a background task and allow more waiting time.
WHAT IT CAN AND CANNOT DO
It can read many kinds of material. It does not generate images or live voice.
It can read text, images, video, audio, and PDFs, then answer in text. It can also call functions, return a fixed format, run code, search, and read a web link. Computer use is still in preview. It cannot generate images, use the Live API, or use a manually set fixed context cache.
FAQ
Answer these questions before you start.
What does Gemini 4 Argon API cost?
Through November 30, 2026, 1M input tokens cost $2 and 1M output tokens cost $10. Tokens used while the model thinks count as output. The rates become $4 and $20 on December 1.
What does a 1M context window mean?
It is one shared space. The material you send, the model’s thinking, tool calls, and the final answer all use it. More material leaves less room to think and answer.
Should I use Gemini 4 Argon or Gemini 3.8 Flash?
Start with Gemini 3.8 Flash to read mixed material quickly and get a first understanding. If the job needs you to consider many files at once, make a hard plan, or carefully review important code, define the task clearly and send it to Argon.
SOURCES
This page uses official model and API documentation.
Capabilities, context limits, pricing, and migration requirements come from Google’s Gemini API documentation, checked on October 14, 2026.
