AI Inflection Point May 2026: What I'd Bet On Now

The AI Inflection Point of May 2026: A Field Report Before Google I/O

I had the article half-written when Tuesday happened.

It was supposed to be a calm preview piece. The Gemini 3.x AB-test variants leaking through the iOS app. A few rumors about an Omni video model. Standard pre-keynote anticipation. I had my coffee, my notes, my outline. And then on May 5, 2026 — the same Tuesday — three companies fired three loaded weapons in roughly six hours.

A Miami startup called Subquadratic walked out of stealth with a 12-million-token context window and a claim that their architecture uses less than 5% of the compute Claude Opus burns. OpenAI quietly swapped ChatGPT's default brain to a new model that hallucinates 52.5% less on medical, legal, and financial questions. Anthropic shipped ten production-ready Claude finance agents and a full Microsoft 365 integration. Perplexity launched a competing financial agent on the same day with 35 pre-built workflows and live data feeds from Morningstar, PitchBook, Daloopa, and Carbon Arc.

I scrapped the outline.

What you are about to read is what May 2026 actually looks like fourteen days before Google I/O — not as a press release roundup, but as a field report from someone who has been deploying production AI for actual paying clients while all of this was landing. Some of these announcements will reshape the next twelve months of how I build. Some of them are noise dressed up as news. And one of them — the one nobody on Twitter was screaming about — is, in my read, the most important AI moment of 2026 so far. It is not the one you think.

Let me walk you through what I am betting on, what I am holding back from, and what every developer reading this should do this week before I/O reshuffles the deck again.

Why This Particular Tuesday Mattered

I have been writing about AI weekly roundups for two years. Most weeks blur together. A new model. A pricing change. A feature drop. They land, they get a headline, you keep working.

This Tuesday was different in a way that took me a day to fully process.

What landed on May 5 was not three product launches. It was three architectural bets converging on the same week, four days after Google sunset Project Mariner — its long-running browser-agent research project — and folded the technology into the Gemini Agent personal assistant inside the Gemini app. That sunset was not a footnote. It signaled that Google is repositioning ahead of I/O, away from "experimental browser agents" and toward "the 24/7 agent that lives where you live." Two weeks before the keynote.

So now zoom out. In one week:

• The compute floor moved. SubQ's sub-quadratic sparse attention architecture posted benchmark numbers that — if they hold up under independent review — collapse the assumption that frontier intelligence requires frontier compute.
• The default ChatGPT model got smarter on the things that matter most. GPT-5.5 Instant's 52.5% hallucination reduction on high-stakes domains is the kind of release note OpenAI used to save for full point-releases.
• The financial-services AI war went hot. Anthropic and Perplexity dropped competing finance-analyst agent suites on the same day, both targeting the exact junior-analyst workflows that have employed armies of MBAs for decades.
• Google cleared the runway. Project Mariner discontinued. Omni model leaked in the Gemini UI. I/O 2026 keynote on May 19 with a model reveal almost universally expected.

I have rebuilt my mental model of the field three times in fourteen days. If you are running production AI right now, you should too. Let me start with the announcement that I think matters most — and that almost nobody is treating as the headline.

Subquadratic and the 12-Million-Token Question

The Subquadratic launch was buried under the GPT-5.5 news cycle. That is a mistake.

Here is the short version. A Miami-based startup called Subquadratic exited stealth on May 5 with $29 million in seed funding, a frontier model called SubQ, and a context window of 12 million tokens built on what they call Subquadratic Sparse Attention (SSA). According to their technical blog, SSA achieves a 7.2x prefill speedup over dense attention at 128,000 tokens, rising to 52.2x at 1 million tokens, and at the full 12-million-token context the model uses less than 5% of the compute of comparable frontier systems — what they describe as nearly a 1,000x reduction.

Read those numbers slowly. Then re-read them.

The dominant assumption since GPT-3 has been that scale costs compute and compute costs money and money gates intelligence. Every frontier model release of the last three years has reinforced that wall. Opus 4.6 is excellent and expensive. Gemini 3 Pro is excellent and expensive. GPT-5 is excellent and expensive. The pricing tiers we have been arguing about are all scoped to that compute floor.

If SSA's claims survive third-party verification, that floor moved.

The benchmarks they posted are not modest. On RULER at 128K, SubQ scores 97.1 against Opus 4.6's 94.8. On SWE-Bench Verified, SubQ reports 82.4% versus Opus 4.6's 81.4% and Gemini 3.1 Pro's 80.6%. At long-context evaluations specifically, the kinds of tasks where most models fall apart past 200K tokens, SubQ apparently holds together to 12M.

I want to be careful. The honest read is more cautious than the headline.

The skepticism camp is not wrong. Subquadratic does not yet have a public technical paper detailing the architecture in enough depth to reproduce. The benchmark numbers are self-reported. The complexity claims have not been independently verified. We have all seen this pattern before — a lab posts magic numbers, the community runs the eval suite, the magic shrinks.

So why am I leading the article with this instead of GPT-5.5 or the finance agents? Because the direction of the bet matters more than the exact accuracy of the launch numbers.

The financial sector watchers are not the only ones who should be paying attention here. If sub-quadratic attention works at frontier scale — even at half the efficiency they claim — it changes what's possible to put in a context window for normal applications. A 12M context is not a slightly bigger 1M context. It is the entire codebase of a mid-sized SaaS product, in one prompt, at compute costs that look closer to a current Flash model than a current Opus model. That is a different category of tool.

I am running my first production test of SubQ this week. I will not commit to anything until I have my own numbers on my own data. But I am also not betting against architectural innovation that posts results this aggressive on a benchmark suite this competitive. I have been wrong about that bet too many times before.

If you are deploying production AI in May 2026, here is the practical move: don't migrate yet, but architect for a world where context-window pricing collapses. Stop optimizing for 200K-token chunking strategies that assume the cap will hold. Build retrieval pipelines that can elastically scale up if the next twelve months turn the 1M-context tier into the new Flash tier. (For the practical playbook on managing 1M-token sessions today, see my Claude Code 1M context management notes — the same patterns scale.) The tooling decisions you make now will look very different if the SSA bet pays off.

GPT-5.5 Instant and the Quiet Default Switch

While SubQ was getting researcher Twitter into a fight, OpenAI was making a different kind of move — a quieter, more enterprise-shaped one.

On May 5, OpenAI rolled out GPT-5.5 Instant as the new default model for ChatGPT, replacing the GPT-5.3 Instant that had been default since earlier this year. The headline numbers in the company's release post:

• 52.5% fewer hallucinations on high-stakes medical, legal, and financial prompts in internal evaluations
• 37.3% fewer inaccurate claims on a separate set of prompts users had previously flagged for factual errors
• HealthBench score of 51.4 out of 100, up from 49.6 (GPT-5.3 Instant)
• HealthBench Professional (clinical) at 38.4, up from 32.9
• AIME 2025 at 81.2, versus 65.4 for GPT-5.3
• MMMU-Pro at 76.0, versus 69.2

If you skim past those numbers, you will miss the actual story.

The story is not that the model got better. Models get better. The story is which axes it got better on. OpenAI optimized GPT-5.5 Instant explicitly for the things that legally and financially matter: medical, legal, financial. The model that millions of people will hit by default when they open ChatGPT is now significantly more reliable on the questions where being wrong has real consequences.

That is a strategic choice, not a technical accident. And it tracks with the broader May 5 pattern. Both OpenAI and Anthropic — on the exact same Tuesday — pointed their highest-leverage releases at high-stakes professional domains.

Here is what that means in practice for me.

I have been testing GPT-5.5 Instant on the kinds of tasks I typically route to Opus for safety reasons — legal contract review for client work, financial analysis for SaaS pricing audits, basic medical-adjacent research where I am explicitly trying to avoid the model making something up. The early signal is real. It is not Opus-on-research-mode quality. But for fast, default-tier responses on those domains, the hallucination rate drop is noticeable in a way GPT-5.3 was not.

Paid users keep access to GPT-5.3 Instant for the next three months in case the new default behaves differently for their specific workflows. That detail matters. OpenAI is signaling they expect some users to feel the change as regression — likely because GPT-5.5 Instant trades certain stylistic behaviors for accuracy gains. If you have prompt scaffolding tuned to GPT-5.3 quirks, audit it before the three-month window closes.

The under-discussed implication: this is OpenAI quietly conceding that the default model matters more than the flagship. Most ChatGPT users will never opt into the most expensive tier. The model that gets most of the world's AI questions answered is the default. Optimizing it for high-stakes accuracy is a much bigger societal-impact lever than another tenth of a percent on AIME.

I keep my Opus subscription because of the long-context reasoning and the agent integrations I have built around Claude Code. But for a meaningful share of my one-shot questions, especially the kind where I would have previously double-checked the answer in a second tool, GPT-5.5 Instant is now the call I make first. That has not been true since GPT-4.

The Anthropic Finance Agent Drop — And Why Microsoft 365 Is the Real Story

Anthropic's May 5 announcement was the densest of the week, and the part of it that got the most coverage — the ten finance agent templates — was not the most important part.

Let me cover the templates first, because they are real. Anthropic released ten ready-to-run agent templates for financial services, split into two categories:

Research and Client Coverage (5 agents):

• Pitch builder
• Meeting preparer
• Earnings reviewer
• Model builder
• Market researcher

Finance and Operations (5 agents):

• Valuation reviewer
• General ledger reconciler
• Month-end closer
• Statement auditor
• KYC (Know Your Customer) screener

Each agent is what Anthropic calls a "reference architecture" — a packaged combination of skills (instructions and domain knowledge for the task), connectors (governed access to the data the task runs on), and subagents (additional Claude models for sub-tasks). They can run as plugins inside Claude Cowork and Claude Code alongside human analysts, or they can be deployed as Anthropic-managed agents where Anthropic handles the production infrastructure.

That is the kind of release that deserves a serious paragraph from anyone covering financial AI. But here is what got buried.

Same announcement. Same day. Anthropic shipped full Microsoft 365 integration — Claude functioning as a single agent across Excel, PowerPoint, Word, and Outlook, carrying context across all four applications simultaneously.

If you do not work in finance, that sentence might not register. If you do, it should land like a falling piano.

The standard junior-analyst workflow looks like this: pull data into Excel, model it, build a deck in PowerPoint, draft the cover memo in Word, send it through Outlook with three follow-up emails. Each tool break used to mean a context break — a place where information had to be manually carried between applications, where errors crept in, where junior analysts spent the unglamorous hours that justified their entry-level salaries.

A single agent that holds context across all four Microsoft 365 apps is not an "AI productivity tool." It is the structural disappearance of an entry-level job category. Combined with the Moody's data partnership Anthropic announced the same day, the message is unambiguous: Anthropic is not building chat companions for analysts. They are building the digital workforce that used to be the analysts.

For the strategy parallel, my field notes on Anthropic's managed agent rollout cover the "secure production infrastructure" model in more depth — that's the same plumbing now powering these finance templates.

This is also where the Perplexity story enters.

The Perplexity Counterpunch — And Who Actually Wins

On the exact same Tuesday, Perplexity launched Computer for Professional Finance.

The structural similarity is not subtle:

• 35 dedicated finance workflows automating the work analysts repeat every week
• Licensed data integrations with Morningstar, PitchBook, Daloopa, and Carbon Arc
• A PitchBook Essential MCP server integration that gives Perplexity native access to PitchBook's firmographic intelligence
• Output formats that include tearsheets, annotated stock charts, and equity research comparisons with every figure linked back to its source

If Anthropic's pitch is "an AI workforce that operates inside your existing Microsoft 365 stack," Perplexity's pitch is "the financial operating system itself" — a destination tool, not an integration. Where Anthropic is asking enterprises to plug Claude into their existing toolchain, Perplexity is asking them to migrate to a new working surface where the data lives natively.

Both bets can win. They probably won't both win in the same accounts.

My honest read: Anthropic has the upper hand right now, for a reason that has nothing to do with model quality. The Microsoft 365 integration is the moat. Most large financial-services firms run their work on Excel and PowerPoint. Asking them to migrate analyst workflows into a new destination tool is friction. Asking them to add Claude as a layer over the tools they already use is closer to free. That is structural advantage that does not depend on which model writes a slightly better earnings summary.

But Perplexity has something Anthropic does not: native data partnerships built into the product surface itself. The PitchBook MCP integration in particular is a different shape of advantage. When the question is "find me every Series B SaaS deal in the last 18 months that closed at over 12x ARR," the model that has PitchBook data already wired in has a structural edge over the model that has to be told where to look.

The honest forecast is that this is going to be a workflow-by-workflow split. KYC screening and month-end close go to Anthropic because of the operational integration. Market research and deal sourcing go to Perplexity because of the data layer. Pitchbook building and earnings review get fought over for the next eighteen months.

If you are deploying AI in a financial-services context this quarter, do not pick one. Run both, scoped to specific workflows. The competitive pressure between the two of them will pull pricing and capability faster than either would have moved alone.

Gemini 3.2 Flash, AB Tests, and the Pre-I/O Scramble

Now to the part I had originally planned to lead with — and which has been demoted by everything above.

Google has been AB testing multiple Gemini 3.x variants for weeks ahead of I/O. The names spotted in iOS Gemini app traffic logs include Gemini 3.2 Flash, Ajax, Hercules, Hector, and Orpheus. The variants appear to be cycling — one Reddit user reported their iOS Gemini app shifting from Gemini 3 Flash to 3.1 to 3.2 over a 24-hour period.

The leaked pricing for Gemini 3.2 Flash, based on AI Studio API logs, is $0.25 per 1 million input tokens and $2 per 1 million output tokens. If those numbers hold at I/O launch, Gemini 3.2 Flash hits flash-tier pricing with capability close to Gemini 3.1 Pro — which would extend Google's pricing-vs-quality lead at the mid-tier.

One important correction worth flagging, since I have seen this go around in roundups this week. The knowledge cutoff for Gemini 3 models is January 2025, not January 2026. I saw the 2026 number cited in a few summary threads. It is not what Google's model documentation says. Worth getting right before you architect retrieval logic around an assumption that does not match.

The bigger Google story is the Omni model leak. A UI string spotted in the Gemini video generation interface this week shows the line "Start with an idea or try a template. Powered by Omni" sitting next to "Toucan" — the internal name for the existing Veo-3.1-powered video pathway. The placement of "Omni" inside the consumer UI, not just in code logs, is what makes observers think this is bigger than a rename.

There are three plausible interpretations:

1. Omni is a public name for the same Veo pathway. Possible but unexciting.
2. Omni is a new Gemini-trained video model alongside Veo. Possible.
3. Omni is a unified Gemini omni-model handling both image and video natively in one system. The most architecturally significant possibility — and the one that would land hardest at I/O.

If interpretation three holds, Google ships the first top-tier omni-model that handles video and images in a single unified system. Combined with the Project Mariner sunset on May 4 and folding into the Gemini Agent personal assistant, the I/O narrative is being staged carefully: a flagship model reveal, a unified multimodal generation system, and a 24/7 agent that lives inside the Gemini app and replaces the experimental browser-agent work Mariner was doing.

Three plausible model reveals at I/O 2026 (Monday, May 19 — Tuesday, May 20):

• Gemini 3.5 Pro / 3.5 Flash — most likely shape of the headline launch
• Gemini 4.0 — Polymarket traders are at 94.5% on "no" for 4.0 release by June 30, but I/O has surprised before
• Omni as the multimodal generation flagship paired with whatever the new Gemini headline is

What I am watching for specifically: pricing on the new Flash tier, whether the agent inside the Gemini app gets a separate name and pricing model from the chat experience, and whether Google announces anything that addresses the agentic coding gap with Codex and Claude Code — because that is the place Google has been losing ground fastest.

For the broader race context, I covered the AI super agent race in May 2026 last week — the side-by-side test of Codex, Cowork, and Gemini that ended with only one finishing my morning task cleanly. Spoiler: it was not Gemini. I/O is Google's chance to change that.

Gemma 4 MTP Drafters — The Most Useful Release Nobody Talked About

While SubQ was eating the headlines, Google's open-source team shipped something that almost every developer reading this should care about more than they currently do.

Quick clarification first, because this got muddled in the source notes I was working from. The multi-token prediction drafting release was for Gemma 4 — Google's open-source model family — not Gemini 4. Two different products, two different release tracks. Gemma 4 is the one you can actually run.

Here is what shipped. MTP (Multi-Token Prediction) drafters for the Gemma 4 family using a specialized speculative decoding architecture. The drafter pairs with a heavy target model — say, Gemma 4 31B — and uses idle compute to predict several future tokens at once with the lightweight drafter, in less time than the target model takes to process one token. The target model then verifies all the draft tokens in parallel.

The result: up to 3x speedup without any output quality degradation.

The MTP drafters are released under the same Apache 2.0 license as Gemma 4, with model weights available on Hugging Face and Kaggle, and support for Transformers, MLX, vLLM, SGLang, and Ollama out of the box.

For developers running local Gemma 4 models on consumer GPUs or Apple Silicon, this is a serious latency upgrade for free. If you have a real-time chat application, an agentic workflow, or a voice product where user-perceived latency matters, MTP drafters are a one-evening integration that drops response times noticeably without changing the model itself.

This is the kind of release that does not generate cycles of discourse but quietly improves the production experience of everyone running open models. Worth ten minutes of your week to evaluate.

Pomelli Catalog and the AI Marketing Tool Quietly Eating SMB Workflows

One more Google release that fits the pattern of "quiet ship, real impact."

Pomelli — Google Labs and DeepMind's AI marketing tool for small and medium businesses — added a feature called Pomelli Catalog. The flow is: you upload your products or services, Pomelli stores them in your catalog, and the tool generates personalized marketing campaigns and AI-created product photos on demand. Free, globally available where Pomelli is launched (US, Canada, Australia, New Zealand, with Europe expanding).

Pomelli works by analyzing your website to create a Business DNA profile — your tone of voice, custom fonts, images, color palette — and then generates campaigns that match. With the Catalog addition, the loop closes: products go in, branded campaign creative comes out, downloadable for Instagram, TikTok, Facebook, YouTube, and LinkedIn.

The January 2026 addition of Pomelli Animate, powered by Veo 3.1, lets the tool transform static marketing content into on-brand video animations. Combined with Catalog's Photoshoot feature, which uses Nano Banana 2 to turn any product photo into professional studio-quality images, you have a full SMB marketing workflow — branded photo, branded video, branded campaign — in one free tool.

For solo operators and SMBs running e-commerce, this is the version of the AI marketing automation story that I keep telling friends about and they keep underestimating. It is not as flashy as a finance-agent armada. It is more useful for more people. If you run a Shopify store with under fifty SKUs, you should have tested Pomelli Catalog by Friday.

The Boston Dynamics Sidebar Worth Filing Away

A note that does not fit the AI software story but belongs in the May 2026 picture.

Boston Dynamics' Atlas humanoid robot is going to production. At CES 2026 in January, the company unveiled the production-ready version. As of May 2026, all 2026 Atlas deployments are fully committed. Fleets are scheduled to ship to Hyundai's Robotics Metaplant Application Center and — significantly — to Google DeepMind, which is integrating its Gemini Robotics AI foundation models into the Boston Dynamics system.

The relevant detail is not the dancing videos. It is the partnership with DeepMind. The same company shipping Gemini 3.x variants and an Omni multimodal model is the one putting frontier AI inside humanoid robots. The convergence of language models, multimodal generation, and embodied AI is happening in May 2026, on Google's roadmap, with Boston Dynamics' chassis. File this for the post-I/O conversation. We are going to be reading a lot more about Gemini Robotics in the back half of 2026.

What I Would Actually Bet On If I Were Deploying Production AI This Month

Eight thousand words in, here is the field-report distillation. If you are deploying production AI workflows in May 2026, this is what I would actually do this week.

Architect for a context-window collapse. Don't migrate to SubQ yet — wait for independent verification — but stop building chunking strategies that assume 200K is the ceiling. The next twelve months will likely turn 1M-context into table stakes and 10M+ into a real possibility. Build retrieval pipelines that scale elastically.

Use GPT-5.5 Instant as the new default for one-shot factual questions in high-stakes domains. Keep your Opus subscription for long-context reasoning and agent work. But for fast medical, legal, or financial lookups, GPT-5.5 Instant is now the call I make first.

Run both Anthropic Claude finance agents and Perplexity Computer side by side, scoped to different workflows. Anthropic for everything that lives inside Microsoft 365. Perplexity for anything that needs PitchBook, Morningstar, Daloopa, or Carbon Arc data natively. Don't pick one until the fight has run for ninety days.

Wait until I/O before committing to a Gemini integration. Gemini 3.2 Flash pricing is extremely competitive on paper, but launching production work on a model two weeks before its successor announces is a recipe for a migration you didn't plan for. Watch the keynote on May 19, then commit.

Integrate Gemma 4 MTP drafters into any local-model workflow you are running. It is a free latency win.

If you run an SMB or e-commerce business under fifty SKUs, test Pomelli Catalog this week. It is the version of the AI marketing automation story that consistently overdelivers relative to its publicity.

Watch for the agentic-coding response from Google at I/O. That is the gap that Google needs to close, and the one that will most directly affect every developer reading this. If they ship something that competes with Claude Code or Codex on the kind of long-running agentic coding workflows we covered in the May super agent race breakdown, your tool stack changes.

The One Thing I Almost Missed

I have been writing AI roundups long enough to know that the announcements that feel biggest in week one are often not the ones that matter in month six. Looking back at the announcements I wrote breathlessly about a year ago, half of them are footnotes now. The same caution applied to the April 2026 industry shakeup — half of those panic stories normalized within thirty days, and the durable signal was buried in the quieter releases.

So I have been forcing myself to ask, on every Tuesday like this one: which of these will I still be talking about in November?

GPT-5.5 Instant is a quiet, durable release. The hallucination drop on high-stakes domains is the kind of improvement that matters every week, forever, for billions of users. That is durable.

The finance agent fight is durable. Whether it is Anthropic or Perplexity that wins more workflows, the disappearance of junior-analyst entry points is now in motion. By 2027 we will be talking about how this changed financial-services hiring.

Gemma 4 MTP drafters are durable in the boring, useful way. Faster local inference is not glamorous, but it ships a real improvement to anyone running open models locally. That stays in my stack.

The Gemini 3.2 Flash AB-test variants — Ajax, Hercules, Hector, Orpheus — are not durable. They are pre-launch noise. By June, all of this gets replaced by whatever Google actually announces at I/O. If you are spending mental cycles on the variants today, redirect those cycles to the I/O keynote on May 19.

And SubQ. SubQ is the wild card. If the architectural claims survive, it is the most significant release of 2026 — bigger than anything I expect Google to announce at I/O. If they do not survive, it joins the long graveyard of "magic numbers in launch posts that did not reproduce." I am watching for the third-party benchmark replication threads to start landing in the next two weeks. If they line up with the company's claims, we are in a new compute regime by autumn. If they don't, we keep building on the floor we have.

I/O is in two weeks. The picture today, on May 6, 2026, is going to look different by May 21. But the direction of the bets — toward higher-context, lower-cost models, professional-domain accuracy, financial-services automation, and embodied AI partnerships — is not going to reverse. The next twelve months are going to be defined by which of those bets cash out and how fast.

The article I sat down to write would have been a calm preview of Google I/O 2026. It is not that anymore. It is a snapshot of the moment the field genuinely shifted under everyone's feet — and a working theory of which footing to take first.

If you only do one thing after closing this tab: watch the I/O keynote on May 19 with the framework above in your head. Look for which gaps Google closes, which ones they punt, and which announcements they make that nobody saw coming. The gap between what they ship and what the rest of this week shipped will tell you exactly where the next twelve months are going.

I'll be live-noting the keynote. See you on the other side of it.

Let's Work Together

Looking to build AI systems, automate workflows, or scale your tech infrastructure? I'd love to help.

• Fiverr (custom builds & integrations): fiverr.com/s/EgxYmWD
• Portfolio: mejba.me
• Ramlit Limited (enterprise solutions): ramlit.com
• ColorPark (design & branding): colorpark.io
• xCyberSecurity (security services): xcybersecurity.io