Claude Opus 4.5, OpenAI vs. Google, Nano Banana Pro

Another news cycle is met with another major frontier model release: this time Claude Opus 4.5. While it's conceptually somewhat incremental on Opus 4.1 capabilities-wise, what's notable about this release is how cost plays out. Not only is it 66% cheaper than Opus 4.1 on a per-token basis, but early feedback is also that it tends to complete tasks with fewer tokens used than Sonnet 4.5.

In fact, Claude Code users on the Claude Max plan will be happy to learn: 4.5 Opus and Sonnet usage limits are now combined. They were previously decoupled, with Opus usage hitting limits painfully quickly - it's now much less-so the case and makes Opus 4.5 a more practical daily driver model.

Amp, the coding agent, which had moved its default model off of Anthropic's models for the first time onto Gemini 3 a week before, shifted back to Opus 4.5 as the default on release with a thoughtfully-evaluated approach.

With Anthropic handily staying at the top of the coding and enterprise workflow markets, OpenAI and Google are duking it out for consumer-facing dominance. The Information reported last night that OpenAI is in a "code red" to marshal more resources to improve its current product offerings in response to Google's rise in AI market share. The Financial Times reported numbers that back up this narrative a few days before.

Notably, The Information's report acknowledged that OpenAI is working on introducing ads to ChatGPT. This confirms recent sleuthing (and adjacent mis-reporting) from ChatGPT users that ads were imminent in the platform. But the "code red" effort means a delayed ads launch.

Moving in lockstep, Google is testing ads in AI Mode - its most direct ChatGPT-competitor and AI product with its biggest userbase. We see these developments as thoroughly unsurprising, given that advertising is what drives the massive internet economy, especially on the consumer side. What we're eager to learn is how much better ad targeting will be through these AI mediums compared to traditional search and social media. If advertiser cost-per-conversion metrics are significantly cheaper than Google Search ads or Meta's social media ads, expect a massive shift (and increase!) in where advertiser dollars flow. OpenAI's chunk of that as they approach 1 billion weekly users (consider: there are ~6 billion worldwide internet users) should be considerable, and might give us some compelling data to quell (or amplify) "AI bubble" fears.

Google, for its part, appears to be making a big bet on multi-modality being key to continued Search dominance. While OpenAI, Anthropic, and others are mostly focusing on text-based models with a side of image and video capabilities, Google has consistently been at or near the top of the conversation regarding the latter. And they're actively combining all those experiences into one in products like AI Mode.

While the future is probably that combined multi-modality, Google's recent release of Nano Banana Pro, the state of the art image reading/manipulation/generation model, stands well on its own. It's being lauded (and also criticized) for its ability to create actually-usable infographics, sans the usual typographic typos that tend to plague image-generation models. The photorealistic image generation capability has very much crossed from "uncanny valley" to "indiscernible from a real photo".

Have questions or feedback for the Pulse team? Requests for changes on pulsemcp.com? Ideas for content/collaboration? Join our Discord community.

What's upcoming for MCP?

→ The big news in the MCP community this week is the announcement of MCP Apps: a collaboration between the MCP-UI open source project, Anthropic, and OpenAI. This is the standardized form of OpenAI's Apps SDK we've been waiting for, allowing for the embedding of interactive UI widgets inside AI chat and agentic workflows. With Anthropic seemingly on track to adopt it in their products, there seems to be no doubt that "UI in MCP flows" will become a permanent staple in the ecosystem. The need is obvious and broad-reaching: not only is there the consumer-facing need of "I'm a brand and I want to retain control over my branding and in-chat experience" common to most of OpenAI's launch partners, but we also see an acute need in many enterprise workflows: for example, easy-to-navigate interfaces to "review AI bulk-processed work before finalizing it". That kind of experience just can't live in a text-only interface, so either you build an app around it - or now, an app within it.

If you're ready to start dabbling in MCP Apps, jump in to participate in Alpic's virtual hackathon, accepting submissions through December 7th. Ship an MCP App, deploy it on Alpic, and win up to $1k in cash prizes.

→ The 2025-11-25 MCP specification release happened as planned, with a finalized blog post and changelog entry. We reported on the highlights last week. Core Maintainer Den Delimarsky has since published a personal thorough breakdown of what this means for the auth parts of the updates.

We expect the next news cycle around MCP to be quiet for the holidays - expect more progress on the roadmap in the new year!

Featured MCP Work

Blueprint MCP Server (#10 this week) by Arcade.dev
→ A simple but useful concept: generate a diagram to help understand a codebase or system architecture. Most interestingly, the happy path demonstrates Arcade's gateway platform to great effect: from secrets management on through centralized MCP control, complete with a beautiful diagram rendered by Nano Banana Pro.

Antigravity IDE by Google
→ Google launched the Antigravity code editor a couple weeks ago. This is the product of the controversial acquisition of Windsurf, with the ex-Windsurf founders continuing to take a beating in the court of public opinion with this launch. The launch itself felt incremental on other IDE competition, though deep integration with the aforementioned multi-modality that Google is betting on in Search could make for interesting developer UX enhancements. We don't think it's likely to beat the MCP ecosystem, but it's good to see Antigravity supporting MCP out of the gate as well: Google is playing it on all fronts.

Mercury Official MCP Server
→ You can always count on Mercury to be at the forefront of digitizing banking capabilities: here now with an official MCP server. Finding transactions, graphing cash flows, and anything else you might want to analyze from your suite of Mercury services opens the door to interesting personal or business finance applications. Alas, we need more of the banking industry to follow Mercury's lead before we can build a full financial picture of a business or individual in an agentic workflow.

Horizon MCP Gateway by Prefect
→ Another week, another MCP gateway platform announcement. This one comes from the creators of the wildly popular open source FastMCP framework. While the feature list does not come off as novel, the experience backing it is certainly unique: FastMCP Cloud has long been one of the more popular ways for the broader public to get their MCP servers up and running in a hosted environment.

Browse all 400+ clients we've cataloged. See the most popular servers this week and other recently released servers.

A Few Good Links

→ OpenAI launched a ChatGPT Shopping Research experience, and we think it's a great move on their part. The demonstrated UX - follow on questions and feedback loops like "why didn't you like this" and "show me more like this" - feels properly in line with how people actually shop online, and gives ChatGPT the benefit of excellent user feedback data they can use to improve the experience (and their moat) over time. It's also in line with their challenge to Google's Search empire, where Google Shopping + Amazon effectively control the entire market for online shopping entrypoints today. OpenAI even "trained a version of GPT-5 mini specifically for shopping tasks" - they clearly think shopping is a significant frontier in their quest to win the consumer AI market.

→ Anthropic published an engineering blog post on effective harnesses for long-running agents - an authoritative source given Claude Code's clear excellence in the domain. Many takeaways with parallels to Claude Code features; but one in particular that's sticking with us is the notion that your agents should "make incremental progress in every session, while leaving clear artifacts for the next session." This is true not only for cross-session tasks, but even within a single long-running session: "[when the model runs] out of context in the middle of its implementation, [it leaves] the next session to start with a feature half-implemented and undocumented. The agent would then have to guess at what had happened, and spend substantial time trying to get the basic app working again." Using "artifacts" (usually markdown files or the like) as sort of "checkpoints" that condense insights and progress hard-won over multiple conversation turns and tool calls are key to stringing together many cohesive sessions - or one very-long-running session - to success.

→ This was a great take on "ask me anything" chatbots: they're doomed to fail. "When you promise "anything," you can't test it properly, evaluate success, or even know which failures are costing you customers." The writeup focuses on voice agents, but it's true for agents more broadly. Build an agent or product that "does anything," and the reality will be that it does a whole lot of things poorly, and you'll get overwhelmed trying to address its infinitely long tail of failures. Scope it properly, and it becomes a tractable problem where you can iteratively build a moat of data and evals. Of course, take it too far, and your hyper-specific workflow will quickly get subsumed by a quality more-general-purpose agent… So the trick is to make sure you're threading the needle somewhere in the middle, the sweet spot of which is specific to the complexity of the domain you're tackling.

→ For our weekly check in on, "are we in an AI bubble?", we turn to the opinions of leading AI researchers: this week, Ilya Sutskever made his position clear that "scaling the current thing will keep leading to improvement … but something important will continue to be missing." This aggregated take from other leaders is a helpful summary, in that experts may disagree on time horizons (from Amodei's ~2 years to LeCun's 10+ years), but generally agree massive transformation is coming. They're splitting hairs, and seemingly all would agree "the current paradigm is likely sufficient for massive economic and societal impact, even without further research breakthroughs." That's all quite the far cry from the narrative spun by AI-fear-mongerers.

→ Looking for your next MCP-related speaking opportunity? MCP Connect Day, coming up on February 5th in Paris, just opened their call for speakers - get those proposals in!

Cheers,
Mike and Tadas