52.5% fewer hallucinated claims. That one number from GPT 5.5 Instant buried the bigger story — which is that "chatbot that answers" is being quietly replaced by something that acts first.
Three things landed in the same week. Google leaked an internal agent called Remy. Anthropic pushed Orbit — a proactive briefing layer — into Claude's web and mobile builds as a settings toggle. And OpenAI replaced its default ChatGPT model with GPT 5.5 Instant, focused almost entirely on accuracy over cleverness. Separately, each is interesting. Together, they describe a very specific turn the industry is taking.
What Is Google Remy — And Why "Agent" Is the Right Word
Google employees are currently testing Remy internally — what the company calls a "dogfooding" phase, where staff uses the product before public release. The internal description positions it as something that "elevates the Gemini app into a true assistant that can take actions on your behalf, not just answer questions or generate content."
That last clause is the tell. Every AI assistant since 2022 has been built around answering questions. Remy is built around handling tasks — monitoring what matters, running workflows in the background, learning preferences over time. The integrations are deep: Gmail, Docs, Calendar, Drive, Search. All first-party Google services, all controlled by the same company. That is a real structural advantage over third-party agents trying to stitch together permissions from different platforms.
The comparison being made internally is to OpenClaw, a viral autonomous AI that gained attention earlier this year for responding to messages and conducting research without being prompted each time. OpenAI hired OpenClaw's creator in February. Google's answer appears to be: we already control the inbox, the calendar, and the document layer — we just need the agent layer on top.
No confirmed release date yet. Google I/O 2026 runs May 19–29 at Shoreline Amphitheater in Mountain View. If Remy is anywhere close to production-ready, that is the obvious venue.
Gemini 3.2 Flash: What Changed Under the Hood
A version labeled Gemini 3.2 Flash appeared on the Aluther AI Arena — an external benchmarking environment where models compete against each other under real-world conditions. Google did not announce it. It surfaced through evaluation results.
Three practical upgrades stood out. SVG generation is noticeably sharper — the model can produce detailed vector graphics with higher precision than current Gemini 3 Flash. Coding for interactive 3D environments improved, specifically voxel-based simulations and dynamic systems. And animation processing handles smoother transitions in interactive content, which matters for UI design work more than most people realize.
Using Aluther AI Arena rather than internal benchmarks is a deliberate choice. External evaluation platforms expose weaknesses faster — the model gets pushed across task types it was not specifically tuned for, and direct comparison with competitors happens in public. It signals confidence, or at minimum, urgency. This is not a quiet internal improvement cycle.
Gemma 4 MTP Drafters: How Google Is Squeezing 3x Speed Out of the Same Hardware
This one gets less attention and probably should not.
Google released multi-token prediction (MTP) drafters for the Gemma 4 model family. The target is inference speed — one of the least glamorous and most consequential bottlenecks in production AI. Here is the actual mechanism: standard language models generate one token at a time. Every token requires moving large amounts of data from memory into compute units. The system is memory-bandwidth limited, not compute-limited — meaning it spends more time shuffling data than calculating.
MTP changes this with speculative decoding. A smaller, faster model — the "drafter" — predicts multiple tokens ahead. The main model then verifies that entire predicted sequence in a single pass. If the prediction is correct, the system accepts the whole batch and adds one more token in the same step. You get multiple tokens generated in the time it normally takes to produce one. The drafter and main model share the same KV cache, so attention states do not need to be recomputed — that saves additional time.
Output quality does not degrade. The main model still does the final verification. Google's claimed ceiling is 3x faster inference. On Apple Silicon, batch size increases alone yield around 2.2x gains. Similar numbers on Nvidia A100 GPUs. For edge devices, Google added embedding layer clustering to speed up the final probability step — historically one of the slowest parts on limited hardware.
3x faster inference at no quality cost is a larger improvement than most new model releases. It is just harder to put in a headline.
GPT 5.5 Instant: The Accuracy Numbers That Actually Matter
OpenAI replaced GPT 5.3 Instant as the default ChatGPT model with GPT 5.5 Instant — which means this change touches the highest-volume deployment in the industry. The focus was not capability expansion. It was accuracy reduction.
GPT 5.5 Instant produces 52.5% fewer hallucinated claims compared to the previous version. On difficult conversations — the kind involving medicine, law, and finance — inaccurate claims dropped by 37.3%. These are not benchmark numbers optimized for a press release. They describe behavior in adversarial conditions, which is where hallucination actually damages users.
The model also improves on visual reasoning, math, coding, and image analysis. But the personalization layer is worth noting separately. GPT 5.5 Instant can draw on past conversations, uploaded files, and connected Gmail accounts to shape responses. It introduces memory transparency — users can see which past interactions influenced a specific answer and manage that data directly. That combination of accuracy improvement and visible memory is a specific trust-building move, not a feature expansion.
For context on earlier GPT-5.5 capability analysis, see the GPT-5.5 vs Claude Mythos breakdown published earlier this month.
Anthropic Orbit: What "Proactive" Means in Practice
Orbit is not released. It is showing up as a settings toggle inside newer Claude web and mobile builds — that staging pattern typically means the feature is being prepared for a phased rollout rather than a full launch announcement.
What it is designed to do: pull from Gmail, Slack, GitHub, Figma, Calendar, and Drive, and deliver a personalized briefing based on your time zone and connected apps. Without you asking. That connector list is very specific — it targets the daily workflow of developers, product managers, and designers. Not general office users. Someone who needs to know what changed in a GitHub repository overnight, what was discussed in Slack while they were offline, which Figma frames got updated, and what meetings are coming up — Orbit is built for that person.
Anthropic's Code with Claude conference runs May 6 in San Francisco, May 19 in London, June 10 in Tokyo. That calendar makes a quiet Orbit rollout or a formal reveal in the coming weeks both plausible. The OpenClaw cost comparison from earlier this year has relevant background if you want context on how Anthropic has been positioning Claude against autonomous agents.
My Take
The framing of "Remy vs Orbit" is a bit misleading — they are not competing for the same user. Remy targets everyone who uses Google's consumer suite. Orbit targets builders who live in GitHub and Figma. Different products solving the same underlying problem: reduce the number of tabs you have to open before you can start thinking.
What is interesting is how much the MTP story got buried under the agent announcements. A lossless 3x speed improvement in inference is the kind of infrastructure change that enables everything else on this list. Faster models make always-on agents feasible. Without it, a 24/7 background agent is a billing problem.
GPT 5.5 Instant's accuracy numbers are the move that nobody copied yet. 52.5% hallucination reduction is not a research paper result — it is a default model change at scale. That matters more than any agent that is still in internal testing.
FAQ
Is Google Remy available to the public?
No. As of May 2026, Remy is in internal employee testing at Google. No public release date has been confirmed. Google I/O 2026 (May 19–29) is the earliest likely reveal window, but availability depends on how internal testing goes.
What is Anthropic Orbit and when does it launch?
Orbit is a proactive briefing tool for Claude. It connects to Gmail, Slack, GitHub, Figma, Calendar, and Drive and prepares context-aware updates without requiring prompts. It has appeared as a settings toggle in Claude's web and mobile builds as of May 2026, suggesting a staged rollout is close but not yet active.
How is GPT 5.5 Instant different from GPT 5.3 Instant?
GPT 5.5 Instant is now the default ChatGPT model, replacing 5.3 Instant. The key improvements are a 52.5% reduction in hallucinated claims and a 37.3% reduction in inaccurate claims on difficult conversations in areas like medicine, law, and finance. It also adds memory transparency — users can see which past interactions influenced a given response.
What is multi-token prediction (MTP) and does it affect output quality?
MTP is a speculative decoding technique where a smaller "drafter" model predicts several tokens at once, and the main model verifies them in a single pass. If the prediction is correct, the system accepts the whole batch. Output quality is unchanged because the main model still does final verification — only the speed changes. Google's implementation for Gemma 4 claims up to 3x faster inference.
Will Remy replace Google Assistant?
That is not confirmed, but the positioning suggests Remy operates at a different layer entirely. Google Assistant handles voice commands and quick device tasks. Remy is described as handling complex, multi-step workflows across Google's productivity suite — closer to a background executive assistant than a voice interface. Whether Google consolidates them under one product is an open question.
The pattern across all three announcements is the same: less asking, more acting. Whether that resolves into useful tools or ambient noise depends entirely on how well these systems learn what actually needs attention versus what can wait. Remy and Orbit have to earn that judgment. GPT 5.5 Instant is at least building the reliability floor first.
0 Comments