Skip to main content

​
Key Takeaways

  • Available from day one: Google announced Gemini 4 Argon on September 30 and opened it to the paid tier of the Gemini API on October 13. APIYI is among the first platforms with access. Model name gemini-4-argon, in the default and svip groups, served from official Google capacity
  • Limited-time promotional rate: $2 input / $10 output per 1M tokens through November 30, 2026. From December 1, Google’s price goes to $4 / $20, and APIYI will adjust at the same time
  • Same accuracy, an order of magnitude less thinking: on 12 auto-graded tasks run 3 times each, Argon and 3.1 Pro both scored 36/36, but Argon used about 1/12 of the thinking tokens, and median latency dropped from 35.0s to 15.4s
  • 1M context, verified: a 1.01M-token input returned normally; the limit is 1,048,576
  • One required migration change: Argon does not support thinkingBudget and passing it may return a 400; use thinkingLevel instead

​
Background

Argon is the first model in the Gemini 4 family. Google is not pitching it as a better chat model but as a model for long-horizon work: large codebase refactors, cross-language migrations, finding and patching vulnerabilities, and multi-step knowledge work in fields like law and finance. The rollout came in stages. On September 30 it went to trusted cyber defenders through the Fairwind Program; from October 13 it opened to the Gemini API paid tier, Google AI Ultra subscribers and enterprise customers. The internal case studies in Google’s posts are ambitious: helping the Chrome team split a roughly 6-million-line legacy folder out of the 40-million-line Chromium codebase, speeding up Gemma inference on vLLM by 2.85x, and migrating 800K lines of the Fuchsia kernel from C/C++ to Rust. The biggest API change is a 1M-token window shared by input and output. Earlier Gemini models capped output at 64K; Argon can in principle generate up to “1M minus your input” in one go.

​
Detailed Analysis

​
Benchmarks

The figures below are published by Google and have not been independently reproduced:
One outlet, citing Bloomberg, reported that some Google engineers with early access found Argon underwhelming on real front-end work. We could not verify this and include it for context only: high benchmark scores do not guarantee results on your workload, so test with your own tasks before switching.

​
APIYI Test Results

We ran 400+ calls on October 13–14 (UTC+8), using the production gemini-3.1-pro-preview and gemini-3.8-flash as references. Quality: 12 auto-graded tasks (math, date arithmetic, counting, strictly formatted Chinese writing, JSON extraction, and two Python coding tasks graded by running unit tests locally), 3 runs each: None of these tasks were hard for any of the three models, so accuracy does not separate them. The real difference is how many tokens each model spent to get there: Argon used less than a fifth of 3.1 Pro’s output plus thinking tokens. Compatibility: across 150 paired cases on both protocols, Argon scored 14/14 on the native endpoint and 12/13 on the OpenAI-compatible endpoint, with its one miss matching 3.1 Pro. 3.1 Pro actually lost points in two places: parallel function calling on the OpenAI endpoint (0/2), and a long-context task with an 8.6K-token prefix and a 128-token output cap (0/2, because thinking used up the output budget). Speed: writing a roughly 1,300-token article in Chinese, Argon’s first token arrived in 2 to 7 seconds versus 27 to 54 seconds for 3.1 Pro (which also timed out once at 10 minutes). All three models generated at 70 to 97 tokens per second; Argon’s fast first token comes from doing almost no thinking on writing tasks.

​
Key Features and Caveats

Adaptive thinking

Defaults to medium. Low and medium adapt to difficulty and answer simple questions directly; for reliable deep reasoning, pass high explicitly (on the same proof, thinking went from 0 to 1,097 tokens).

No thinkingBudget support

Google’s docs say it is deprecated with a warning; on APIYI we have seen hard errors (400). Remove thinkingBudget when migrating from 3.x. The minimal level is not supported either.

Sampling parameters are ignored

Google says Argon ignores custom temperature, topP and topK. Passing them does not error in our tests, but do not count on them taking effect.

Low cache hit rate

Google says implicit caching is on by default. In our tests most repeated same-prefix requests did not hit, and only some did. Budget at the full input price and treat any hit as extra savings.

​
Specifications

Long decode beyond 64K needs the Interactions API, which the APIYI gateway does not support yet, so design around the 65,536-token output limit for now.

​
Practical Use

  • Refactoring, migrating and reviewing large codebases: Argon’s headline use case; pair it with thinkingLevel: "high"
  • Long document and long video analysis: cross-checking contracts, financial reports and legal filings in bulk, all within a 1M context
  • Agent orchestration: Google suggests Argon as the lead planner and reviewer, with Gemini 3.8 Flash running parallel subagents, to balance quality and cost
  • Security audits: locating vulnerabilities and proposing fixes

​
Code Example

With the OpenAI SDK, set base_url to https://api.apiyi.com/v1 and use the model name gemini-4-argon. More examples are on the Gemini 4 Argon Overview.

​
Best Practices

  1. Set timeouts to 15 to 20 minutes: this is Google’s recommendation, and during the launch window we also saw occasional queuing of several minutes
  2. Pass high explicitly for deep reasoning: the default level adapts, so do not rely on it for complex tasks
  3. The OpenAI-compatible endpoint is not the place for fine-grained thinking control: reasoning_effort has little visible effect, so use the native endpoint when you need controllable reasoning
  4. Connect agent clients over the chat protocol: Argon does not support /v1/responses, so configure clients such as Codex to use /v1/chat/completions. For multi-turn tool calls on the native endpoint, send the previous turn back unchanged, including thoughtSignature

​
Pricing and Availability

​
Pricing

Model prices are aligned with the official site and may change with it; the table above is for reference only, and the Model Pricing tab in the top navigation is authoritative: Model Pricing.
Billing was checked against real charges: across three calls, the amount actually deducted was within 1.5％ of what the unit prices give. Beyond the unit price, usage matters more: to get the same task right, Argon spends far fewer output and thinking tokens than 3.1 Pro. During the promotional period its input price is slightly above 3.1 Pro and its output price slightly below, so most tasks end up costing less. When prices double on December 1, re-run your cost estimates.

​
Top-up Promotions

Model billing matches the official price item by item; discounts come from top-up bonuses. See Top-up Promotions.

​
Summary and Recommendations

Gemini 4 Argon showed no capability gaps in our tests. Its standout trait is reaching the same accuracy with much less thinking, which makes it both faster and cheaper. Teams using 3.1 Pro for complex reasoning, code and long documents should shift some traffic to Argon during the promotional period and compare. When migrating, remove thinkingBudget, raise your timeouts, and leave caching out of your cost model.
Sources (retrieved 2026-10-14): Google blog posts blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/ (2026-09-30) and blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/ (2026-10-13); Google AI Studio developer guide aistudio.google.com/learn/gemini-4-argon-developer-guide; the engineer feedback comes from Startup Fortune’s summary of a Bloomberg report. Test data comes from APIYI’s pre-launch testing on October 13–14, 2026 (UTC+8).