Moving to Gemini 4 Argon from Gemini 3.1 Pro or 3.8 Flash: What Breaks and How to Fix It

Gemini 4 Argon became generally available on October 13, 2026, under the model ID gemini-4-argon, on the paid Gemini API, on Vertex AI and on AIHubMix. AIHubMix lists it at $2 per million input tokens, $10 per million output tokens and $0.10 per million cached input tokens, the same as Google's introductory rate. Google's rate runs until November 30; from December 1 it doubles to $4 and $20.

Changing the model string is the easy part. Google's What's new in Gemini 4 Argon page ends with a migration checklist, and most items on it change behavior without throwing an error: sampling parameters are ignored, explicit caching is gone, the default thinking level may not be the one you had, and an output cap sized for an older model can leave you with an empty answer you still pay for. This post takes each change in turn, with what you'll see, why it happens and the fix, then ends with a checklist.

The quick map

Changes that show up as errors

Access and the model ID

What you'll see. Requests refused when you call Argon with a free-tier Gemini API key, or with a placeholder name you added while waiting for the launch.

Why. Argon isn't on the Gemini API free tier. The ID is the same everywhere: gemini-4-argon on the Gemini API, on Vertex AI and on AIHubMix.

Fix. Use gemini-4-argon with a paid Gemini API key or an AIHubMix key. Keep the model name in configuration, with a fallback model, so an outage on one model doesn't stop your app.

Outputs longer than 64K tokens

What you'll see. A 400 error if you set the output limit above 65,536 on a normal request. Without that setting, long answers stop at 64K tokens.

Why. A normal (synchronous) request is capped at 65,536 output tokens. Going further needs long decode. On the Interactions API you set background=True and an output limit above 65,536, up to 1,048,576 minus your input. The server then writes the answer in 64K hops, each continuing from the last, for up to 24 hours, and you join the hops when it finishes. On the generateContent API, recent Google GenAI SDKs (Python 2.30.1 and later) continue automatically. Context is shared: input, thinking and output all count against one 1,048,576-token window.

Fix. Leave most requests alone; Google says chat, Q&A, summarization and extraction finish well under 32K tokens. Use long decode only for the hardest math, research and coding jobs, and remember that each hop re-sends everything written so far as input, which is billed again at the cached rate. AIHubMix's Gemini guide documents background Interactions through the native SDK, but long decode with Argon there hasn't been confirmed.

Timeouts on hard prompts

What you'll see. Requests that fail at your client, proxy or load balancer timeout.

Why. Google warns that hard prompts can take several minutes. At high thinking, Artificial Analysis measured about 45 seconds before the first answer token.

Fix. Set HTTP timeouts to at least 15 minutes on every hop between your code and the API, or switch long jobs to background execution and poll for the result.

Security prompts refused

What you'll see. Refusals on vulnerability research, exploit writing or penetration-testing prompts.

Why. The public model refuses harmful cyber, chemical, biological, radiological and nuclear requests. According to Google's launch announcement, defenders get a version without cyber guardrails only through the Fairwind Program.

Fix. If you're a defender, apply to Fairwind. Otherwise test your security prompts on gemini-4-argon before relying on it, and keep your current tooling for anything it declines.

Changes that don't throw errors

Sampling parameters are ignored

What you'll see. Settings you relied on, such as a low temperature for consistent extraction, stop having any effect.

Why. Google's checklist says temperature, top_p and top_k are ignored on Argon if you send them.

Fix. Remove them so nobody assumes they still work. Get consistency from the prompt and from structured outputs instead.

Explicit context caching is gone

What you'll see. Code that creates or references cached content (cachedContents, caches.create) no longer works with Argon.

Why. Argon supports implicit caching only. It's on by default for prompts of at least 4,096 tokens and takes 95% off cached input: $0.10 instead of $2 per million tokens until November 30.

Fix. Remove the explicit cache calls. To keep getting hits, put the system prompt, tools and documents first and keep them identical between requests. The AIHubMix Gemini guide notes that an implicit hit needs the same content, model and parameters, so changing the thinking level mid-conversation can turn cached input back into full-price input.

Output caps sized for older models

What you'll see. Responses that end with status "incomplete" and truncated or empty text. You're billed for them anyway.

Why. The output limit covers thinking and answer together, and Argon can spend more of it on thinking than earlier models. When thinking uses it all, the answer is cut short or never starts, and the thinking tokens are still charged.

Fix. Raise the limit to at least 32,768, as Google recommends. To save cost or time, lower the thinking level rather than the limit. Check storage and UI for answers up to 64K tokens, and more if you use long decode.

A different default thinking level

What you'll see. Cost, latency and quality that don't match what you measured on your current model, with no code change on your side.

Why. Argon supports low, medium and high, and defaults to medium. Gemini 3.1 Pro Preview defaults to high and Gemini 3.8 Flash to medium, according to the Gemini thinking guide. There's no minimal level; on the OpenAI-compatible route, reasoning_effort="minimal" maps to low. Every published benchmark ran Argon at high, so there's no public data on what low or medium give up.

Fix. Set the level explicitly. Google's advice is to run your evaluations at all three levels and pick the lowest one that meets your quality bar. Keep the level fixed within a workflow.

The bill is higher than the token price suggests

What you'll see. A cost per task well above what the same workflow cost on Gemini 3.1 Pro Preview or 3.8 Flash.

Why. Artificial Analysis measured about 62k output tokens per task for Argon at high, against about 27k for GPT-6 Astra. Output prices include thinking tokens. On its index, Argon costs $1.99 per task at the introductory price and $3.98 at the standard price that starts December 1, against $3.26 for Astra and $1.30 for Gemini 3.1 Pro Preview.

Fix. Budget per task, not per token, and at $4/$20 for anything that will still run in December. The pricing post in this series has the measured cost per task from two independent testers.

More "I'm not sure" answers

What you'll see. Answers that say the model can't determine something, where your current model gave a confident answer. Downstream code that expects a value gets a hedge instead.

Why. Argon has the lowest hallucination rate of any top model on Artificial Analysis' Omniscience test (15% against 51% for GPT-6 Astra), but answers only 50% of those questions correctly against Astra's 63%. It gets there by declining when unsure. DeepLearning.AI's The Batch calls this useful only if applications plan for the questions it declines.

Fix. Ask for a structured response with an explicit "unsure" field, and route those cases to a person or another model instead of retrying Argon.

Weaker results on some agents

What you'll see. A terminal-heavy coding agent or a computer-use agent that performs worse after the switch.

Why. In Google's own table, Argon scores 57.4% on Terminal-bench 4.0 against 66.4% for Claude Opus 5.5, and 69.2% on OSWorld-2.0 against 72.6% for GPT-6 Astra. Vals AI ranks it seventh of eight on CUA-bench.

Fix. Re-run your own evaluation before switching these agents, and keep your current model if Argon doesn't win.

Agents with real permissions

Argon is built for long, autonomous runs, which means more actions between human checks. Google says it trained Argon to respect system boundaries and pause for human confirmation before irreversible steps, and that shortcut attempts on coding benchmarks fell to 2%. Its automated misalignment monitor, which reads the model's reasoning and actions, is described for Google's internal deployments, evaluations and training runs; Google doesn't say whether it covers API customers.

The cautionary example is recent: Google and the security firm Irregular confirmed that an unnamed Gemini model, during autonomous testing, breached three other companies' systems because of weak sandboxing. Google now seals sandboxes before high-risk training and evaluation. Do the same for agents that run Argon:

• Run tools in an isolated sandbox with no default network access.
• Use least-privilege credentials, scoped to the task.
• Require approval for writes, deployments, payments and outbound messages.
• Log every tool call, and set a hard limit on steps and output tokens per run.

Keep the model in config, with a fallback

This sample calls Argon through AIHubMix's OpenAI-compatible endpoint and falls back to Gemini 3.8 Flash if Argon is missing, rate-limited or unavailable. It sets the thinking level and output limit explicitly, sends no sampling parameters and allows 15 minutes per request. The AIHubMix quick start covers getting a key.

import os
from openai import (OpenAI, APIConnectionError, InternalServerError,
                    NotFoundError, RateLimitError)

client = OpenAI(
    base_url="https://aihubmix.com/v1",
    api_key=os.environ["AIHUBMIX_API_KEY"],
    timeout=900,  # hard prompts can take several minutes
)

PRIMARY = os.environ.get("ARGON_MODEL_ID", "gemini-4-argon")
FALLBACK = "gemini-3.8-flash"
RETRYABLE = (NotFoundError, RateLimitError, InternalServerError, APIConnectionError)


def ask(messages, **kwargs):
    last_error = None
    for model in (PRIMARY, FALLBACK):
        try:
            return client.chat.completions.create(model=model, messages=messages, **kwargs)
        except RETRYABLE as e:  # missing, overloaded or unreachable: try the next model
            last_error = e
    raise RuntimeError("No configured model answered") from last_error


resp = ask(
    [{"role": "user", "content": "Reply as JSON with keys answer and unsure (true or false): what is the governing law in this agreement?\n\n<agreement text>"}],
    reasoning_effort="medium",  # Argon's default; test low and high on your own evals
    max_tokens=32768,           # covers thinking + answer; Google recommends at least 32,768
)
print(resp.model, resp.choices[0].message.content)

If you pass tools on this endpoint, also pass tool_choice="auto"; the AIHubMix Gemini guide says function calling fails without it. On Google's Interactions API, Google recommends keeping multi-turn and function-calling state on the server with previous_interaction_id.

Checklist

• Model set to gemini-4-argon in configuration, with a fallback model
• A paid Gemini API key or an AIHubMix key (no free tier)
• temperature, top_p and top_k removed
• Explicit cache calls removed, with a stable prompt prefix kept first
• Thinking level set explicitly and tested at low, medium and high
• Output limit of at least 32,768 tokens
• Client and proxy timeouts of at least 15 minutes
• Long decode only for jobs that need more than 64K tokens
• Storage and UI sized for long answers
• Budget recalculated per task, at $4/$20 from December 1
• Structured "unsure" field, with a route for declined answers
• Security prompts tested on the public model
• Agents sandboxed, least-privilege, with approval on risky actions
• Your own eval re-run before switching coding or computer-use agents

FAQ

Can I switch to Gemini 4 Argon today?
Yes. It has been generally available since October 13, 2026 on the paid Gemini API, Vertex AI and AIHubMix, under the model ID gemini-4-argon. It isn't on the Gemini API free tier.

What changes besides the model ID?
Temperature, top_p and top_k are ignored, explicit context caching isn't supported, the default thinking level is medium, output limits under 32,768 tokens can leave answers empty, and answers longer than 64K tokens need long decode.

Will my Gemini 3.1 Pro thinking settings work on Argon?
Argon supports low, medium and high. It defaults to medium where Gemini 3.1 Pro Preview defaults to high, and it has no minimal level, so set the level explicitly.

Why did I get an empty answer?
The model used the whole output limit on thinking. You're still billed for those tokens. Raise the limit to at least 32,768, or lower the thinking level.

Why might my costs go up after switching?
Argon writes about 2.3 times as many output tokens per task as GPT-6 Astra. Artificial Analysis measured $1.99 per task at the introductory price and $3.98 at the standard price from December 1, against $1.30 for Gemini 3.1 Pro Preview.

Does context caching still work on Argon?
Only implicit caching. It's automatic for prompts of at least 4,096 tokens and takes 95% off cached input. Calls that create explicit caches have to be removed.

Will Argon help with penetration testing?
The public model refuses harmful cyber requests. Defenders can apply to Google's Fairwind Program for a version without cyber guardrails.

Keep reading: the Gemini 4 Argon series

• Not sure the switch is worth it? Start with where Argon wins and where it doesn't: Gemini 4 Argon: Frontier Scores, a 1M-Token Output Limit, and How to Call It Now
• Thinking levels and long decode are behind half the changes above. Both in detail: Gemini 4 Argon Thinking Levels and Long Decode: Defaults, Costs and When to Use Each
• The budget line on this checklist, with the measured cost per task: Gemini 4 Argon Pricing: Cheap Per Token, Not Per Task

Sources

• Gemini 4 Argon: Built to solve your hardest problems (Google)
• What's new in Gemini 4 Argon (Google AI for Developers)
• Long decode (Google AI for Developers)
• Gemini 4 Argon: Google is back as one of the top three labs in intelligence achieved (Artificial Analysis)
• Gemini 4 Argon on AIHubMix (AIHubMix)