Transcription
[Music] OpenAI just unlocked full MCP tool support in Chat GPT's developer mode, which is the update people have been waiting for because it means Chat GPT can finally go beyond answering questions and start taking real actions.
Also, Bite Dance is challenging Google's nano banana with a model that's faster and cheaper. Deep Agent is letting anyone spin up apps with payments in minutes. Adobe dropped enterprisegrade AI agents into its ecosystem. and Claude can now edit Word, Excel, PowerPoint, and PDFs without you even opening them. We've got a stack of fresh AI updates to cover today. So, let's talk about it.
Open AI first because this one changes how you actually use Chat GPT. Developer mode now has full support for MCP tools. And what that really means is you're no longer stuck just pulling information from systems. You can now act on those systems directly from chat. We're talking about updating a Jiraa ticket, firing off a Zapier workflow, tweaking something in your CRM, or even linking multiple services together into a chain of actions without ever leaving the conversation window.
It's rolling out in beta for plus and prousers on the web. And setting it up is actually pretty straightforward. You head into settings, then connectors, then advanced, and flip on developer mode. After that, you can plug in your own MCP servers right inside the connectors tab, and they'll appear in ChatGpt's composer once you're in developer mode. From there, ChatGpt can talk to those servers through protocols like SSE and streaming HTTP. And you've got authentication options like OOTH if your tools require it. You can even toggle specific tools on or off per connector and refresh them whenever you update your server so everything stays in sync.
Now, here's the part that really changes the game. Chat GPT can use any tool your connector exposes. Not just readonly, fetch or search, but fullon write actions. That means the way you prompt matters more than ever. If you have multiple connectors that overlap, you'll want to tell chat GPT which one is the source of truth. If two tools look similar, you'll want to be explicit about which to use. You can even define the exact sequence, like telling it to first read a repo file, then write a modified version back, and to ignore everything else.
To make this all smoother, the descriptions you write for your tools on the MCP server really count. The clearer and more actionoriented those names and notes are like use this when you need to update a customer record, the better chat GPT will pick the right one. Because this is right access, open AI's put a big safety bar across the flow. That's why every tool call shows you the full JSON input and output before anything runs. And right actions always need your approval unless you choose to let it remember. OpenAI's clear prompt injections, model slip-ups, or shady connectors could mess things up fast. But if you handle it carefully, chat GPT stops being a passive dashboard and turns into an active control center that actually runs your workflows in real time.
All right. Now, Bite Dance, their Cream 4.0 image model just landed and the target is clear. Gemini 2.5 flash image, the one everyone nicknamed Nano Banana. Bite Dance says Cream 4.0 know beats it on their internal magic bench across prompt adherence, alignment, and aesthetics. They haven't shipped a formal technical report with those results yet, so treat that as a claim, not a settled fact. What is clear is the product move. Cedream 4.0 essentially merges the text to image chops of Cream 3.0 with the editing strength of Seeddit 3.0. And according to artificial analysis, that combo is the real evolution.
Pricing is aggressive. The domestic sticker stays the same as Cadream 3.0 at $30 per 1,000 generations. On Fall.AI for global hosting, it's around 0.03 cents per image, while Gemini 2.5 Flash image sits closer to 0.039. Performance-wise, Bite Dance is saying raw image inference is over 10 times faster versus earlier versions, and early user feedback online has been positive about editing accuracy quick. Faithful changes from text prompts without mangling composition. Availability is straightforward if you're in their ecosystem. Jimang and Dupau on the consumer side. Volcano engine for enterprise.
On public leaderboards, Gemini 2.5 Flash Image still holds the top slot for both generation and editing. Cream 4.0 hasn't been scored there yet. For context, Cream 3.0 sits around fifth for generation and sixth for editing on those rankings. broader scene in China keeps heating up. Quaisho and Tencent are in the mix. The government recognized copyright for AI generated content in late 2023. And now there's mandatory labeling. And if you're experimenting with multi-reference workflows, VidU rolled out a reference to image tool internationally that blends up to seven reference images at roughly 9 cents per output, while Gemini allows up to nine references. So, you've got options depending on how referenceheavy your pipelines are.
All right, now let's talk Deep Agent because this one hits creators, freelancers, and small teams right where it matters. Getting paid. The update is simple to say and big in impact. Deep agent can now generate a real app from a single prompt and wirestripe payments as a first class part of that flow. Not will give you a code step live payments. If you've ever done Stripe the hard way, you know why this matters. Keys, scopes, environment variables, web hooks, test sandboxes, and hours in docs.
Here, the agent asks you for the essentials: product names, prices, what's included, and it scaffolds the whole thing with checkout handled. You link your Stripe account in a couple of secure steps, and from that moment, sales settle straight to your account. No duct tape in the middle. The sites it spits out don't feel like prototypes, either. You get a clean build with three practical funnels ready on day one. A product page with visuals and Stripe checkout. A workshop page where people can book and pay and a consultation page that locks payment before someone gets on your calendar. The customer experience is endtoend pay instant confirmation email done.
The agent keeps the friction low on updates too. Want to change pricing? Add a product throw in a discount to say it and the Stripe logic and the UI update together. And it's not limited to a storefront. You can prompt for a mini CRM, a notion style workspace with permissions, a marketplace, even an Xstyle micro blog. Payments remain baked in. The team at abacus.ai is trying to spark an ecosystem around this with a weekly human AI build contest that pays $2,500 and the base tier starts around 10 bucks a month. Put that next to the old way. months of engineering, a handful of devs, and a six-figure budget. And it's obvious why someone reportedly got a working paid app Live in about half an hour. That's not a flex. It's a signal that the bottleneck moved from, "Can I build this to do I have something worth selling?"
Over in enterprise land, Adobe just made its AI agents generally available across the experience cloud. And this is less hype and more turn the crank and ship value. The backbone is the Adobe Experience platform agent orchestrator. It uses decision science plus language models to interpret intent then activates the right agents towards a business goal. The lineup covers audience building, customer journey orchestration, experimentation analysis, site optimization, data insights and product support. These agents sit inside journey optimizer, customer journey analytics, experience manager, and the real time CDP, which means they're plugged into the data you're already turning over inside Adobe's stack.
Adoption is not theoretical either. Adobe says over 70% of AE customers are already using the AI assistant, which is the conversational layer those agents sit behind. Big brands on the roster include the Hershey company, Lenovo, Merkel, and Wegman's. The customization story is getting stronger too. Agent composer, agent SDK, and an agent registry are on the way so teams can tune agents within brand rules and policy. On the ecosystem front, Adobe lined up partners like Cognizant, Google Cloud, Pavvice, Medallia, Omnicom, PWC, and VML to push multi- aent collaboration in verticals where you've got both compliance overhead and complex handoffs.
The interesting bit under the hood is their reasoning engine for dynamic adaptive reasoning. So the system doesn't just match intents to skills statically. It adjusts as the journey evolve. The pitch is consistent. Augment teams, lift return on investment, and personalize at scale without ripping out your existing workflows. If you're already deep in Adobe land, these agents aren't a sidecar. They live where your campaigns and customer data already exist.
Now to Claude because Anthropic just made a very practical move that most people need weekly. Claude can now create and edit office style files directly from natural language and large uploads. You give it some data or just a plain request and it'll spit out an Excel file, a PowerPoint deck, a Word document, or even a PDF. It can handle pretty big files, too. Up to around 30 megabytes on upload or download, so you're not stuck with tiny limits.
It can convert CSV or TSV into structured reports, generate charts, and the key feature, bulk contextaware edits without you opening the file. You can tell it to replace every ABC with XYSD, convert all USD prices to EU at a current or specified rate, and change a RO label from manager to executive. Then it applies those updates while preserving the original layout. One shot, consistent formatting, no manual cleanup. Google's Gemini can generate docs, too. For example, exporting deep research into a Google doc. So, the gap here isn't creation, it's the speed and control of edits on existing files across formats.
There's also a bigger signal. Microsoft reportedly has a deal to bring Claude into the Office 365 suite. If that lands the way it sounds, you'll have Anthropic's editing and creation flow sitting natively inside Word, Excel, PowerPoint, and the rest, which turns Claude into an actual operator inside your day-to-day instead of a separate tab. For anyone wrangling reports, decks or financials, that means offloading the tedious parts, find replace with context, currency changes, role, title normalization, table fixes without the risk of breaking templates.
Before I go, just remember, Faceless Empire is up and running, but only for the first 200 people. Hit the link in the description and secure your spot. That's it for today's updates. Tell me what you think in the comments. Make sure to subscribe and drop a like. Thanks for watching and catch you in the next one.