Model guides

Gemini 4 Argon on SJolt: API Access, Pricing, and Setup

Gemini 4 Argon is now available on SJolt. Start in the Playground or connect through Chat Completions, with $0.60 input and $3.00 output pricing per million tokens.

SJolt Editorial6 min readUpdated
Gemini 4 Argon title surrounded by glowing purple and cyan glass tubes
SJolt’s Gemini 4 Argon model artwork. Conceptual illustration, not an image generated by Argon.

Gemini 4 Argon is now available on SJolt for coding, source-based analysis, and visual understanding. Use google/gemini-4-argon through the OpenAI-compatible Chat Completions API, or open its model page to try a text prompt in the Playground.

As of October 15, 2026, SJolt charges $0.60 per million uncached input tokens, $3.00 per million output tokens including reasoning, and $0.03 per million cache-read tokens. This guide covers the request format, pricing, and capabilities available through SJolt, including how its output limit relates to Google’s long-decode feature.

Gemini 4 Argon is live on SJolt

To get started, sign in to SJolt and open the Gemini 4 Argon Playground. For application access, create an SJolt API key, ensure your account has balance, and send requests to the LLM base URL below. The same model page provides its API documentation and current pricing.

SettingSJolt integration
Public model IDgoogle/gemini-4-argon
API base URLhttps://llm.sjolt.ai/v1
EndpointPOST /chat/completions
Inputs and outputAPI: text and base64 images. Playground: text. Output: text.
Shared context window1,048,576 tokens for input, thinking, and output
Maximum output65,536 tokens, including thinking
Reasoning effortlow, medium, high; default medium
StreamingAvailable through the API; Playground uses non-streaming responses

Gemini 4 Argon pricing: SJolt and Google Standard

The following USD rates are per million tokens, checked on October 15, 2026. SJolt’s three token rates are 70% lower than Google’s current Standard introductory rates. This compares token prices; it does not imply that both integrations expose the same features.

Token categorySJoltGoogle Standard introductory
Uncached input$0.60$2.00
Output, including reasoning$3.00$10.00
Cache reads$0.03$0.10

For example, 100,000 uncached input tokens plus 20,000 billed output tokens cost $0.12 on SJolt: $0.06 for input and $0.06 for output. The same token quantities at Google’s Standard introductory rates total $0.40. This is a planning calculation, not a measured request; actual reasoning, cache hits, and retries change usage.

Cache-read pricing applies only to reported cache hits. SJolt’s current Argon rates apply across context lengths, with no additional long-context surcharge. In the current SJolt route, completion_tokens already includes thinking and is the complete output count used for billing. Any completion_tokens_details.reasoning_tokens breakdown is part of that count; do not add it again. Use the returned usage and the settled account record when reconciling a request.

Google lists its introductory rates through November 30, 2026, followed by $4 input, $20 output, and $0.20 cache reads per million tokens from December 1. That schedule is separate from SJolt’s pricing. Check the SJolt model page for current charges before budgeting a production workload.

Send your first Gemini 4 Argon API request

Set SJOLT_API_KEY in your server environment, then use this Chat Completions request. It asks for a small, checkable deliverable and reserves up to 32,768 output tokens for thinking and the final answer. The limit is a ceiling, not a target length.

bash
curl --fail-with-body 'https://llm.sjolt.ai/v1/chat/completions' \
  --header "Authorization: Bearer $SJOLT_API_KEY" \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "google/gemini-4-argon",
    "messages": [
      {
        "role": "system",
        "content": "Use only supplied product facts. Separate confirmed facts from creative suggestions and flag missing information."
      },
      {
        "role": "user",
        "content": "Product facts: a blue ceramic mug with a curved handle. Draft three short product-shot ideas and list facts to verify before publication."
      }
    ],
    "reasoning_effort": "medium",
    "max_completion_tokens": 32768,
    "stream": false
  }'

Read the final answer from choices[0].message.content and inspect choices[0].finish_reason. A length finish reason means the result reached a limit and needs review. For streaming, set stream to true, read the Chat Completions SSE chunks, and assemble choices[0].delta.content until the stream ends; receiving a partial fragment does not mean the task is complete.

An OpenAI-compatible client can use the same base URL, SJolt key, and model ID with its Chat Completions method. This model is exposed through Chat Completions on SJolt; select that method rather than a client’s Responses or native Gemini method.

Set reasoning, context, and image inputs correctly

Start with reasoning_effort set to medium. Compare low for simpler tasks and high for difficult analysis using the same source material and acceptance criteria. Reasoning stays enabled at every level. Omit temperature, top_p, and top_k: SJolt rejects those sampling fields for this model.

Input, thinking, and output share the 1,048,576-token window. Leave room for the response when supplying a large archive. Either max_completion_tokens or max_tokens can set the output ceiling, up to 65,536; choose one. A small ceiling can exhaust the budget during thinking before a useful final answer appears.

For images, use a user message with text and image_url content blocks. Set image_url.url to a complete base64 data URL, such as data:image/png;base64, followed by the encoded image bytes. Remote HTTP or HTTPS image URLs are rejected on this route. The Playground currently accepts text only, so use the API for image analysis.

SJolt’s Argon integration accepts text and images and returns text. Audio, video, and PDF file inputs are not exposed here. For a document workflow, extract the relevant text first and label the source sections. Image and video generation use separate SJolt models.

Does SJolt support Gemini 4 long decode?

SJolt’s current Chat Completions integration caps output at 65,536 tokens, including reasoning, for both streaming and non-streaming requests. Turning on streaming does not raise that ceiling, and Google’s background-execution controls do not enable long decode on this endpoint.

Google separately documents long decode through its native APIs. Its Interactions API requires background: true and max_output_tokens above 65,536; clients then assemble text across the completed interaction’s model_output steps. Use Google’s documentation if that execution model is a requirement. For SJolt workflows, split large deliverables into reviewable sections and preserve the source context needed for each request.

Choose a task with an answer you can verify

  • Code review: provide the relevant implementation, expected behavior, and failing case. Ask for a diagnosis, a focused change, and checks that would verify the fix.
  • Document analysis: label each source and revision, then ask for a synthesis that cites those labels and identifies contradictions or missing facts.
  • Visual analysis: send a screenshot or diagram through the API and ask for observable details before conclusions.
  • Creative planning: turn verified product facts into shot ideas, copy options, or a production brief, with proposed details clearly marked.

Keep the prompt and source packet fixed while comparing reasoning levels. Score factual corrections, usable output, completion time, and billed tokens. These are suggested evaluation tasks; this guide does not report a benchmark or claim that one reasoning level always performs best.

Turn an approved Argon brief into images or video

Argon can plan a scene or explain a reference image. To render an asset, review its brief and send the approved prompt to an SJolt image or video model. For example, use Nano Banana 2.1 for a product image or MiniMax H3 for a video shot, following each model’s own request format.

Carry the approved product facts, aspect ratio, visual references, and camera instructions into the generation step. Review the rendered result against those facts before publishing it.

Start with one prompt, then connect your application

Open the Gemini 4 Argon Playground for a text evaluation, or use the API example above for application integration. Begin with medium reasoning, record the response and usage, and adjust the prompt or effort after reviewing the result.

Sources & further reading

Take the next idea into production.

Explore the models, test a workflow in the playground, and use the same request in your application.

Keep exploring

← All stories