This week brought two notable generative model releases, a wave of voice and phone features for AI agents, and a cluster of changes to how much of these services subscribers can actually use. OpenAI shipped Images 2.6 with native animated GIF creation and persistent reference photos. Google released Gemini Omni 1.2 Flash, which pushes single video clips to 15 seconds. Meta Muse turned on a voice mode, and Grok Bot learned to place phone calls. At the same time, Grok Bot users saw their allowances cut sharply, Anthropic restructured its Claude plans, and OpenAI merged separate usage pools for ChatGPT and Codex. For paying users, the practical question is less about which model is newest and more about how far a monthly subscription now goes.
OpenAI Images 2.6: GIFs, speed controls and saved references
OpenAI announced Images 2.6 across ChatGPT, Codex and its API. The headline feature is animated GIF generation inside a normal chat. A prompt such as asking for the words "Happy Birthday" spelled out in colorful marshmallows returns a looping animation, and the interface adds a playback speed selector with 0.25x, 0.5x, 1x, 1.25x, 1.5x, 1.75x and 2x options. Generation and edits each completed in roughly 20 seconds in testing.
Edits are conversational and multi-turn. After the marshmallow GIF was generated on a wooden desk, a follow-up instruction to change the desk to grass swapped the surface while keeping the lettering intact. Uploading a portrait and asking for a GIF of that person as a dancing medieval knight also worked, as did a request for a poker-playing pig with a meme-style caption. One caveat: the caption text in the meme example was somewhat hard to read against certain backgrounds.
The second major addition is persistent references. A user can upload a photo once, tell ChatGPT to save it as a reference under a short name tag, and then open a brand-new chat and simply ask for a GIF of that name surfing. The model pulls the saved subject without a fresh upload. For creators who reuse the same character, product or face across many sessions, this appears to be a substantial quality-of-life improvement.
For developers, OpenAI introduced two API image models, Flare and Sunburst, which the company says generate images up to 50% faster than previous versions.
Prompting tips for Images 2.6
- Phrase requests as "Create a gif of [subject] [action]" to reliably trigger animated output.
- Use the on-canvas speed selector if the motion looks rushed or sluggish.
- Type "Save this as a reference" with a simple name to store frequently used subjects.
Google Gemini Omni 1.2 Flash: longer clips and tighter reference control
Google's Gemini Omni 1.2 Flash arrives in Google Flow and Gemini for Business. The maximum length of a single generated shot rises from 10 seconds to 15 seconds, and Google claims better character consistency, closer prompt adherence and more cost-effective generation.
The showcase material covers several use cases:
| Capability | What the demonstration showed |
|---|---|
| Virtual try-on | A single portrait dressed in multiple outfits across a 360-degree turn |
| Reference fidelity | A 3D text scene reading "GLOW LUMEN HALO" where 1.2 Flash held typography and lighting better than 1.1 Flash |
| Natural-language editing | A forest behind a knight and queen replaced with a medieval stable and a white horse |
| Multi-character audio | Lip-synced dialogue among animals in an orchid conservatory |
| Camera control | A drone-style flythrough rising through the oculus of the Pantheon |
| Macro detail | A 15-second beehive explainer with motion-tracked callouts for egg, larva and pupa |
The rollout targets Google Flow Ultra and Pro subscribers, but access is not instant for everyone; at least one Ultra/Pro account still showed only Omni 1.1 Flash and the Veo 3.1 variants in the model menu at the time of testing. The comparison with 1.1 Flash and the try-on sequences come from Google's announcement material rather than independent tests.
Grok Bot cuts allowances, and how to live within them
Grok Bot subscribers reported usage gauges jumping from 50% straight to 100% while their bots sat idle. Elon Musk confirmed on X that this was not a bug: the cost of serving Grok Bot had been unsustainably high relative to revenue. No revised numeric caps or reset intervals have been published, which leaves active users guessing about how much they have left.
Even with the cuts, Grok Bot still looks like one of the most capable platforms for deep, autonomous agent work. The question is how to keep a working setup from burning through its allowance. An audit of a heavily used account surfaced two main drains.

▲ Trimming routines and using direct app connectors
Drain one: background routines
The audited account was running 24 scheduled routines. A breaking-news monitor ran 85 times a week, and two digest bots each checked their sources 14 times a week, whether or not anything new turned up. Remedies that preserved most of the value:
- Pause duplicate bots that scan the same feeds.
- Drop non-critical maintenance, such as a spam-folder checker, from daily to weekly.
- Restrict news alerts to business hours instead of every 30 minutes around the clock.
- Route scheduling through a single coordinating bot rather than many scattered loops.
Some community members suggest deleting long-running bots after about four months to reset their state. That advice seems unnecessary; trimming routines achieves sustainable usage without wiping a bot's memory.
Drain two: bot-to-bot chatter and browser automation
Group chats where several bots talk to each other, or cascade subtasks down a chain, consume tokens and session limits quickly. Under the new limits, this kind of multi-bot coordination may not be feasible for heavy workloads. The safer pattern is a dedicated one-on-one chat with each specialized bot for its specific purpose.
Computer-use actions are the other heavy cost. These run in a cloud virtual machine where a bot visually navigates a website, for example typing credentials into a Salesforce login page. Grok Bot's Connect Apps panel offers direct integrations for Slack, Outlook, Google Sheets, Datadog, Vercel, 1Password, Notion and Atlassian, and these connectors are far cheaper in usage terms than having a bot drive a browser.
Meta Muse gains voice and becomes the budget alternative
Several Grok Bot users on X have documented moving their primary workflows to Meta Muse, citing substantially more throughput on its $20 monthly plan than on higher-tier competitors priced around $300; one post referenced a figure of 1 billion tokens. Meta also published a teaser featuring a plush scout character wearing merit badges that appear to represent features: an envelope for dedicated inbound addresses called Muse Mail, a voice-mode badge, Ray-Ban Meta smart glasses integration, and the Muse Charm wearable that Mark Zuckerberg previously said would reach consumers before the holiday season.
The voice mode is now available, and in testing it worked from the mobile app after a desktop attempt failed. A customized Muse assistant answered as a small 3D plush avatar whose mouth and expressions animated in sync with its speech. Notable behaviors:
- The assistant said it could hear the user clearly and handle normal chat actions, but needed an uploaded image for anything visual.
- Asking to change the voice surfaced a picker with styles such as warm, smooth, Playful Whistle, Aria, Boomy Mountain and Engaging Plum.
- The voice used natural fillers like "um" and stopped instantly when interrupted.
- A sharing feature exports looping clips of the avatar dancing to trending tracks.
The judgment here is that Muse's voice is more responsive and emotionally expressive than Grok Bot's or ChatGPT Voice, even though Muse has less raw reasoning power than GPT-6 or Claude. It suits informal tasks, hands-free brainstorming and high-volume casual use, but it does not appear ready to replace Grok Bot for complex autonomous execution.
Grok Bot can now make phone calls for you
Grok Bot's new calling feature lets a bot dial a business or person on the user's behalf. Unlike Muse's voice mode, the user does not hear the audio; instead the chat shows a running text transcript. Before dialing, the bot asks for context, states the call's goal, and warns that it may ping the user in chat if it needs more detail. A confirmation card previews the number, objective and context with a green Call button.

▲ AI phone call with a human-in-the-loop prompt
In a real test, the bot called a business partner to decide whether an X-timeline digest routine that ran twice a day should stay paused. It introduced itself as a virtual assistant calling on a recorded line, correctly explained what the routine did, and then hit a question it could not answer: whether the user was keeping up with the timeline another way. The bot asked the recipient to hold, muted the call, and posted a card in the chat with a roughly 60-second countdown. A typed reply, "I am checking manually," flowed straight into the spoken conversation, the partner agreed to keep the routine paused, and the bot closed the call politely.
Response latency was fast and the cadence natural, though the voice carried a slightly muffled, call-center-headset quality. Two practical rules follow: always review the goal, number and context card before confirming, and keep the chat open during the call so a timed question does not expire.
Plan changes at Anthropic and OpenAI
Anthropic moved in the opposite direction from Grok Bot. The Max 5x plan ($100 per month) is now simply "Max," and Max 20x ($200 per month) is now "Ultra." Existing subscribers transition automatically. The substantive change is that Ultra subscribers no longer face the 5-hour rolling session cap; only the overarching weekly ceiling remains. Heavy users running Claude 3.5 Opus all day can spread intense workloads flexibly on the $200 tier without hitting a short-term wall.
OpenAI emailed Pro subscribers that ChatGPT and Codex will draw from one pooled quota. Previously, ChatGPT separated conversational "Chat," heavier "Work" tasks and "Codex" coding into distinct pools, and power users stretched subscriptions by exhausting chat first. The Chat and Work tabs are merging into a single input that switches modes based on intent: light brainstorming stays in chat, while actionable creation prompts escalate to Work. The loss of siloed pools may reduce flexibility for power users, but OpenAI doubled the baseline allowance for $200 Pro subscribers. Plus subscribers at $20 received no increase, which may feel like an imbalance for that tier.
Quick updates
- Atmee v2: a real-time video avatar model demonstrated by its developer. In a side-by-side benchmark, a clip cost under $0.01 and took 0.03 seconds versus $0.50 and 233.6 seconds for Seedance 2.5, roughly 10x to 100x cheaper. Demos included an unscripted two-avatar talk show and a two-way video call on a tablet. The avatars still read a little like stiff actors, but they look far more human than earlier generations. Access is invite-only via API.
- Project Suncatcher: Google launched a prototype satellite on a SpaceX mission carrying a solar-powered TPU. Sundar Pichai reported that ground control booted the TPU, ran a prompt on a Gemma model and received a haiku back: "Blue marble below, Data streams in silent grace, Starlight feeds my core." This is a single-satellite proof of concept, though Google's research outlines scalable constellation designs.
- Personal finance: ChatGPT Voice now connects to bank and credit accounts for US users through the Fina connector, strictly read-only. Anthropic launched Personal Finance in Claude, powered by Plaid, for budgeting questions, post-payday cash flow and spotting duplicate or recurring charges; it is in beta for US Pro, Max and Ultra subscribers, and Anthropic says these conversations will not train future models.
- New OpenAI models: an OpenAI developer announced that ChatGPT dots will incorporate GPT-6 Astra and GPT-6.1 Terra for faster speeds, improved agent personalities and wider availability, potentially including Plus members.
- Perplexity Teammates: persistent background agents, such as an On-call Engineer, Legal Research Assistant and Storefront Manager, that share Perplexity memory, skills and plugins and hand off tasks to one another. Available on the web for Perplexity Max subscribers, temporarily free, with desktop and mobile to come. Along with Manus Cue, it closely follows the multi-agent architecture Grok Bot popularized, while ChatGPT dots and Claude Duet appear to have drawn slower community enthusiasm.
What to take from this week
The creative releases are easy to act on: try Images 2.6 for quick GIFs and save your recurring subjects as references, and check whether Gemini Omni 1.2 Flash has appeared in your Google Flow model menu. The usage-limit story requires more care. If you depend on Grok Bot, audit every scheduled routine, cut bot-to-bot chats, and swap browser automation for direct connectors before assuming you need a different platform. If you want high-volume conversational use at low cost, Meta Muse's $20 plan and its new voice mode are worth a trial, with the understanding that it is not yet a substitute for a heavy-duty agent. Claude Ultra subscribers can now work in longer bursts, while ChatGPT Pro users gain capacity and Plus users should watch how the pooled quota affects their day. Across every service, caps are moving quickly, so it is worth rechecking your plan's allowances rather than relying on what they were last month.


