Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- The August 28, 2026 xAI Imagine Image API update added auto to quality and changed the omitted-parameter default from medium to auto.
- Currently, auto uses low for image generation and medium for editing. The provider can adjust auto routing, so do not treat its current behavior as a permanent contract.
- Images are billed at the quality actually served. Pass low or medium explicitly when you need stable quality or a predictable budget.
- Editing requests now support up to 5 reference images, previously 3. The update also adds 21:9 and 5:2 aspect ratios.
- Replay batch generation and editing separately before deciding whether to pin quality. Do not rely solely on the subjective appearance of one image.
What changed: omitting the parameter no longer means medium
The xAI release notes explicitly state that quality for grok-imagine-image-2.0 accepts low, medium, and auto, and that omitting quality now defaults to auto instead of medium. Currently, auto selects low for generation alone and medium for editing, with billing based on the quality actually served.
A typical symptom is a change in batch-generation cost or image detail despite unchanged code. Instead of first blaming prompts, random seeds, or the network, check request logs to establish whether quality was omitted and record actual usage and cost fields in the response.
How to interpret auto, low, and medium
auto delegates quality selection to the server and suits workloads that can accept routing changes. low suits sketches, high-throughput previews, and cost-sensitive tasks; medium suits final outputs requiring more consistent detail. These are engineering choices, not quality guarantees, and still need validation with your own material.
If your budget or visual acceptance criteria depend on a fixed tier, pass low or medium explicitly. Do not hard-code the current choice of low for auto generation as a permanent assumption: the official wording is currently.
Run separate regressions for generation and editing
The same auto value currently makes different choices for generation and editing. Prepare separate text-to-image and reference-image editing samples, fix prompts, resolution, aspect ratio, and output count, then compare detail, text, subject consistency, elapsed time, failure rate, and actual cost.
Editing now accepts up to 5 reference images, previously 3. More reference images expand the input, failure, and cost boundaries. Do not send every asset at once simply because the limit increased, or leave the meaning of multi-image ordering to accidental results.
Check aspect ratio, resolution, and output count together
This update also adds 21:9 and 5:2 aspect ratios. The image-generation documentation says an omitted aspect_ratio defaults to auto, an omitted resolution to 1k, and n to 1. These are separate parameters from quality and should be recorded independently during troubleshooting.
Batch-generation costs also depend on output count, resolution, and actual quality. Do not extrapolate the cost of one low image to n=10, 2k, or editing with multiple reference images. Read usage and cost fields from each response and reconcile them with the bill.
Six steps from finding implicit defaults to a limited rollout
Proceed in order: find calls missing quality; separate generation and editing; establish a baseline of previous outputs and costs; test auto, low, and medium explicitly; configure each task type; then roll out to a small share of traffic and observe billing. Client SDKs or OpenAI-compatible layers may not directly expose extension fields, so verify the actual request body.
Configuration changes should support rollback by task. Do not apply a single global default to sketches, product images, and editing workflows together. If an intermediary strips quality, record and test this at the adapter layer instead of assuming that the explicit setting reached xAI.
Separate request errors, unusual costs, and network problems
For a 400, first check the model, endpoint, quality, resolution, aspect_ratio, and reference-image count. For a 429, check quota. For unusual costs, verify actual quality and output count. Investigate the network only when there is evidence of DNS, TLS, connection timeouts, or proxy authentication errors. Retain the request ID, request-body summary, and response usage.
Use the AI API configuration guide and AI API error troubleshooting guide to distinguish model fields, API semantics, and the connection layer.
Conditions for stopping and rolling back
Stop expanding the rollout if auto output falls below the acceptance threshold, costs fluctuate without explanation, an intermediary fails to forward parameters, multi-image editing becomes inconsistent, or batch failure rates rise. Roll back to explicit medium or a verified configuration, and retain auto comparison samples and billing evidence.
If your team needs a stable connection to xAI Console or developer documentation, visit the PuppyIP website to learn about fixed network exits. Network conditions cannot pin auto routing or change quality tiers, quotas, or billing rules.
Sources
Frequently Asked Questions
What is the default when quality is omitted for Grok Imagine 2.0?
It now defaults to auto, not medium. Official guidance currently says auto uses low for generation and medium for editing.
Will auto generation always remain at low quality?
Do not assume so. Official documentation describes the current routing with currently. Pass low or medium explicitly when stable behavior is required.
How is auto billed?
Images are billed at the quality actually served. Check response usage and cost fields and the bill instead of estimating solely from auto in the request.
How many reference images can an editing request include?
Up to 5 after this update, previously 3. Still validate ordering, input costs, and the actual editing result.
Which aspect ratios were added?
This update adds 21:9 and 5:2. Consult the latest image-generation documentation for the complete supported list and default behavior.
When should medium be pinned explicitly?
Pin it and repeat a limited rollout when final outputs require consistent detail, the budget model depends on a fixed tier, or auto variability exceeds acceptance limits.