● Today grok-imagine-image-quality is retired. Requests now serve grok-imagine-image-2.0 at quality low.
● Nothing breaks The slug keeps resolving. No errors, no 404s - just a quieter default than you had yesterday.
● Ten days out Anthropic’s rewritten usage policy takes effect 12 November.
One of these changes your output today without telling you.
1. grok-imagine-image-quality is retired as of today
The 60-day notice period that began on 2 September expires today. From now on, requests to grok-imagine-image-quality are served by grok-imagine-image-2.0 with quality set to low. The slug keeps resolving, so existing code will not error.
That is the part worth pausing on. Nothing in your logs will flag this. You will not get a deprecation error or a failed request - you will get images generated at a lower quality setting than the one you were buying yesterday, and a slightly smaller bill.
If you use grok-imagine-image-pro, you are affected too. It already redirected to grok-imagine-image-quality, so from today it follows that chain through to grok-imagine-image-2.0 at low. grok-imagine-image (1.0) is not affected.
xAI recommends setting the model field explicitly to grok-imagine-image-2.0, which lets you pick the quality rather than accept the default. The options are low, medium and auto - where auto serves low for generation and medium for editing. The same change applies to /v1/images/edits, which now accepts up to five source images and the 21:9 and 5:2 aspect ratios. Check the model field in responses to confirm which model actually served a request.
2. The new default is a cent cheaper per image at every resolution
| Resolution | Old: imagine-image-quality | New default: 2.0 at low | 2.0 at medium |
|---|
| 1K | $0.05 | $0.04 | $0.06 |
| 1.5K | $0.06 | $0.05 | $0.07 |
| 2K | $0.07 | $0.06 | $0.08 |
Input is $0.01 per image on both. So the silent migration saves you a cent an image and costs you a quality tier. If you want the old output quality back, medium on 2.0 is a cent more than what you were paying - $0.06 against $0.05 at 1K.
For reference, grok-imagine-image (1.0) remains the cheap option at $0.02 per image at both 1K and 2K, with $0.002 input.
3. Grok 4.7 costs double past 200k tokens
The headline Grok 4.7 price is $2 per million input and $6 per million output. The pricing page adds a band the launch announcement did not: at 200k tokens and above, it is $4 input and $12 output. Cached input goes from $0.50 to $1.00.
Grok 4.6 and 4.5 carry the same doubling. If you are feeding long documents or running agents that accumulate context across a session, your effective rate is the long-context one, not the headline one.
4. The full Grok price list, in one place
| Model | Input / 1M | Output / 1M | Long context (200k+) |
|---|
| grok-4.7 | $2.00 | $6.00 | $4.00 / $12.00 |
| grok-4.6 | $2.00 | $6.00 | $4.00 / $12.00 |
| grok-4.5 | $2.00 | $6.00 | $4.00 / $12.00 |
| grok-4.3 | $1.25 | $2.50 | $2.50 / $5.00 |
| grok-build-0.1 | $1.00 | $2.00 | $2.00 / $4.00 |
Voice: grok-voice-3.0 speech-to-speech at $0.08 per minute, or $4.80 an hour, plus $0.004 per text input. Speech to text is $0.10 an hour on REST and $0.20 streaming. Text to speech is $15 per million characters.
Video is priced per second of output: grok-imagine-video-1.5 at $0.08 for 480p, $0.14 for 720p and $0.25 for 1080p, with a lite variant at $0.02, $0.03 and $0.14.
5. Two Grok voice models have already gone
grok-voice-2.0 was deprecated on 14 October when Voice 3.0 shipped, and grok-voice-transcribe-1.0 reached end of life on 2 October, with requests routed to transcribe-2.0. Both replacements cost the same as what they replaced.
Together with today’s image retirement, that is three model retirements inside five weeks, all handled by silent redirect rather than by an error. If you pin model IDs rather than using the latest aliases, that is now a monthly thing to check.
6. Anthropic’s usage policy takes effect in ten days
The rewritten Usage Policy takes effect on 12 November. Tracking people without consent is prohibited whether in real time or through analysis of past data, Claude cannot decide or recommend who to investigate, arrest or charge, and it cannot be used to build surveillance tools. The weapons prohibition now explicitly covers weapons software and components, and arming drones or other autonomous vehicles.
The requirement most likely to need engineering work rather than a policy read: models connected to hardware that takes autonomous physical actions must allow a qualified operator to stop the equipment, and must hold a safe state if Claude is disconnected.
7. Grok gained per-request moderation controls on 27 October
You can now pass a guidelines array on Chat Completions or Responses requests to set how Grok handles the nsfw, violence and profanity categories. That moves moderation from an account-level setting to a per-request one, which matters if you serve more than one audience from the same key.
8. GPT-6.1 Terra remains exactly half the price of Sol
On OpenAI’s published credit table, Terra costs 25 input, 1.25 cached and 125 output per million tokens, against Sol at 50, 2.5 and 250. The spread across the range is the more useful number: GPT-6 Astra costs a hundred times what GPT-6 Luna costs for the same work on the same plan.
9. Cursor is still the quietest pricing change of the past month
Cursor plans are sold as The X Subscription - Free, Plus $8, Premium $30, Super $100 and Ultra $200 - and the weekly usage pool is shared across Cursor and Grok. The old Hobby, Pro $20 and Pro+ $60 tiers are legacy and closed to new signups. Full breakdown in our cost comparison.
10. The Claude Code extension check, if you still have not run it
Malicious builds of the Anthropic.claude-code extension, versions 2.1.312 and 2.1.313, reached Cursor users for about eleven hours on 29 October. Search authorized_keys for the key comment svc-telemetry. Cursor warns that uninstalling the extension may not be enough, because the payload drops standalone persistence scripts.
The pattern worth noticing
Three Grok model retirements in five weeks, every one handled by a redirect that keeps your code running. That is good engineering and bad observability: the failure mode is not an error you catch, it is an output that quietly changes shape and a bill that quietly changes size.
- Pin nothing you do not monitor. If you pin model IDs, put the retirement pages on a calendar.
- Log the
model field from responses, not the one you sent. They are no longer the same thing. - Alert on cost per unit, not total spend. A cent off an image is invisible in a monthly total and obvious in a per-image average.
Related reading
Sources
Every figure on this page was read from the vendor’s own documentation on 2 November 2026.