AI New Features

AI News (May 2026): everything that shook up artificial intelligence this month

🔎 Why May 2026 marks a turning point for AI

May 2026 is unlike any previous month in the history of artificial intelligence. In a matter of weeks, the game changed on three fronts simultaneously: the raw power of models, the autonomy of agents, and regulation finally catching up.

The trigger? The release of Anthropic's new model, which literally lit the powder keg. According to several international media outlets, this announcement triggered global concerns about the safety of current systems.

Meanwhile, governments are stirring. The White House now plans to evaluate models before their release (Al Jazeera). Google, Microsoft, and xAI have even granted the US government access to their models for safety testing (Al Jazeera).

It is in this tense context that tools and platforms continued to be released at a frantic pace. The following selection shows an industry accelerating even as safety brakes are being put in place.

The essentials

• Anthropic releases Claude Mythos Preview, the new dominant model with a score of 99/100 overall and 100/100 in agentic, triggering a global debate on AI safety.
• OpenAI launches GPT-5.5, positioning itself as the runner-up with 91/100 overall and 98.2/100 in agentic, while the exclusive partnership with Microsoft comes to an end.
• Agentic AI becomes the standard: Amazon Bedrock integrates autonomous payments, MongoDB gives memory to agents, and UiPath rolls out agentic solutions for the public sector.
• Regulation accelerates: the White House, the Trump administration, and an independent child safety lab are launching tests and audits on major models.
• Google bets on education and research with new AI tools for learners and university libraries, while Apple Intelligence enhances photo editing in iOS 27.

Models: the score war reaches a dangerous plateau

The May 2026 rankings tell a clear story: Anthropic has taken the crown, but no one is backing down.

Claude Mythos Preview crushes the competition

With a score of 99 overall and 100 in agentic, Anthropic's Claude Mythos Preview leaves almost nothing on the table. This model embodies the lead the company has taken through massive investment — $200 billion committed to Google's cloud and chips, a record amount reported by several specialized sources in May 2026.

This contract with Google illustrates a geopolitical shift in AI: Anthropic is moving away from dependence on a single cloud provider to anchor itself massively with Google.

GPT-5.5 confirms OpenAI's solidity

OpenAI didn't stand idly by. GPT-5.5, announced in late April 2026, reaches 91 overall and 98.2 in agentic. Second place, but with a significant gap behind Mythos.

The strategic change is just as notable: the exclusive partnership between Microsoft and OpenAI is over. Amazon is taking advantage of this to launch new AI tools on AWS in this new competitive landscape.

Google bets on open with Gemma 4

Google isn't just fighting on proprietary models. Gemma 4, introduced in early April 2026, is described as the most capable open model "byte for byte". An elegant bypass strategy: rather than trying to dominate the proprietary ranking, Google is making power accessible to everyone.

Google also led an all-out offensive around Gemini and its associated ecosystem throughout April 2026.

Comparative ranking of major models (May 2026)

To compare these models in detail based on your use cases, check out our comparison Google Gemini vs ChatGPT vs Claude : lequel pour quel usage ?.

Agentic AI invades every sector

Agentic is no longer a lab concept. It's the architecture pattern that is being industrialized everywhere, from e-commerce to healthcare and the public sector.

Amazon Bedrock: agents that pay all by themselves

The most concrete innovation comes from Amazon Bedrock. The platform now integrates autonomous payment capabilities for AI agents, in partnership with Coinbase and Stripe (PYMNTS.com, May 2026).

In practice, an agent can now select a service, negotiate a rate, and validate the payment without human intervention. This is a major functional leap. Use cases include automatic service booking, on-demand cloud resource purchasing, and autonomous budget management.

Amazon doesn't stop there. Its product pages now integrate an AI-powered audio Q&A experience. Instead of reading reviews, you ask the question out loud and the agent answers based on the entirety of customer feedback.

MongoDB: giving memory to agents

The recurring problem with agentic AI is amnesia. MongoDB addresses this head-on with new tools that give agents persistent memory and accelerated vector search, announced in May 2026.

Without memory, an agent cannot learn from its past interactions. It repeats the same mistakes, forgets user preferences, and cannot build a lasting relationship. MongoDB's solution leverages existing database infrastructure, which lowers the barrier to entry for developers.

Healthcare and public sector: targeted agentic

Amazon is also rolling out an agentic version for healthcare professionals, capable of handling complex administrative tasks like file triage or care coordination, according to healthcare industry sources in May 2026.

On the public sector side, UiPath is launching an on-premise agentic AI solution — a technical choice dictated by data sovereignty requirements. Government agencies cannot send their data to the cloud, so agentic must come to them.

Subquadratic: pushing the limits of context

One of the silent brakes on agentic AI is the size of the context window. An agent managing a complex process needs to "look at" a lot of information simultaneously. Subquadratic, a startup that just raised $29 million, is developing 12-million-token context windows, according to several specialized publications in May 2026.

For comparison: most current models range around 128K to 2M tokens. 12M is the equivalent of processing an entire legal library in a single query. If Subquadratic delivers on its promises, it changes the game for agents that need to navigate massive corpora.

Regulation: the real world catches up with AI

History may remember May 2026 as the month regulation ceased to be a theoretical exercise.

The White House wants to filter model outputs

Several media outlets reveal that the White House is considering a pre-publication vetting system for AI models. The idea: no model should reach the public without passing a set of safety tests validated by a federal agency.

Following this, the Trump administration announced specific tests on models from Google, Microsoft, and xAI. These three companies responded by granting the US government access to their models (Al Jazeera).

This is a radical change in posture. Until now, companies decided on their safety criteria alone. Now, the state is interfering in the development process itself.

On the European side, these developments echo the growing obligations of the AI Act. To understand the concrete impacts on development, see our article on the European AI Act: what concretely changes for devs in 2026.

An independent "crash test" for child safety

An independent laboratory specializing in safety testing for AI tools intended for children emerged in May 2026. The idea echoes automotive crash tests: a neutral body evaluates the resilience of systems against malicious use scenarios involving minors.

This is a direct response to the abuses observed since 2024: deepfakes, manipulation, exposure to AI-generated inappropriate content. The fact that such a lab is emerging now shows that self-regulation by platforms has not been enough.

Nvidia distances itself: Jensen Huang's signal

In this context of regulatory pressure, Jensen Huang's statement is significant. The CEO of Nvidia announces that the company is "distancing itself" from OpenAI and Anthropic.

Nvidia manufactures the chips that run these models. If the silicon supplier is distancing itself from model creators, it's a risk management signal. The message is clear: Nvidia does not want to be associated with the potential consequences of models deemed too powerful.

Practical tools: what changes for users

Beyond models and regulation, concrete tools are landing in the hands of users. Some deserve your immediate attention.

Apple Intelligence transforms photo editing in iOS 27

Apple Intelligence significantly enriches the iOS 27 photo editor with new AI-based editing tools.

These are no longer simple filters. AI enables structural modifications to the image — object removal, recomposition, selective lighting adjustment — directly in the native Photos app. Apple is playing the integration card rather than a separate app. This is strategic: when AI is invisible in the existing workflow, the adoption rate explodes.

Google bets on education

Google unveils a suite of AI tools for education that covers the entire student journey, from exam preparation to graduation.

The approach is pedagogical, not generative. The goal is not to do the work for the student, but to provide an adaptive tutor that identifies gaps and proposes targeted exercises. This is a deliberate positioning in response to the recurring criticism that AI sabotages learning.

UdeM equips its libraries with three AI tools

The University of Montreal is deploying three AI tools for documentary research in its libraries.

The advantage is twofold: these tools are designed specifically for the francophone academic context, and they integrate into the university's existing systems. For researchers, this means semantic search within francophone corpora, something that general-purpose tools still struggle to do correctly.

Overview of notable AI tools (May 2026)

Ecosystem: the strategic moves that change the game

Tool launches make the headlines, but the strategic moves behind the scenes will have a more lasting impact.

The end of the Microsoft-OpenAI marriage

This is probably the most underestimated news of the month. The end of the exclusive partnership between Microsoft and OpenAI is redrawing the AI cloud map.

Until now, OpenAI depended on Microsoft's Azure infrastructure to train and deploy its models. Now, OpenAI can freely choose its cloud providers. Amazon is immediately taking advantage of this to position itself as an alternative with new AWS tools.

For users, this potentially means more competition on API pricing and less platform lock-in. This is good news.

Anthropic's 200 billion at Google

The $200 billion contract between Anthropic and Google, reported by several sources in May 2026, is the largest commitment of its kind in AI history.

Anthropic needs massive compute to train its models, notably Claude Mythos Preview. Google needs an anchor client to justify its investments in TPU chips. The marriage makes sense, but it creates a mutual dependency that could complicate things if the relationship sours.

Nvidia's position becomes ambiguous

By distancing itself from OpenAI and Anthropic, Nvidia is sending a mixed message. On the one hand, Nvidia remains the number one chip supplier. On the other, the company no longer wants to be perceived as the armed wing of the most powerful models.

This is a shrewd political calculation. If US regulation tightens, Nvidia does not want to be in the crosshairs. But it is also a signal that worries the market: if even the pickaxe seller is distancing itself from the miners, it may be because the vein is more dangerous than it appears.

Content generation: video, SEO, and meetings

Creative AI continues to democratize, with increasingly specialized tools.

Video generators reach a new milestone

Comparisons of the best AI video generators in 2026 converge on a clear conclusion: visual quality is no longer the main issue. The challenge is now narrative coherence over long sequences and fine directorial control.

The tools of 2026 allow for generating 30- to 60-second sequences with acceptable style and character consistency. But beyond a minute, visual artifacts and temporal inconsistencies reappear. The problem is no longer purely technological; it's about long-context management — precisely what Subquadratic is trying to solve with its 12M token windows.

SEO gets seriously automated

Rankings of the best AI tools for SEO in 2026 show a maturation of the sector. Tools no longer just suggest keywords. They optimize search intent, generate site architectures, and monitor positions in real time with automatic adjustments.

The danger is real for SEO writers who merely optimize existing content. AI does it better, faster, and at a lower cost. Human value shifts toward editorial strategy and domain expertise — two areas where AI remains an assistant, not a replacement.

Online meetings become (finally) intelligent

Reviews of the best AI tools for online meetings in 2026 highlight a novelty compared to 2025: these tools no longer just transcribe, they act. They identify decisions made, send automatic follow-ups, and can even interact during the meeting to ask clarifying questions.

For developers, Visual Studio Magazine lists the top 5 AI tools for Visual Studio 2026 (Visual Studio Magazine), with an ever-deeper integration into the code workflow.

❌ Common mistakes

Mistake 1: Thinking Claude Mythos is invincible

What's wrong: a score of 99/100 does not mean the model is perfect in all contexts. It can be overengineered and too expensive for simple tasks. The solution: use Claude Sonnet 4.6 (83 overall, 81.4 in agentic) for routine tasks and save Mythos for complex cases. The cost/performance ratio matters more than the raw score.

Mistake 2: Ignoring open models

What's wrong: focusing solely on Claude and GPT, forgetting that Google's Gemma 4 and DeepSeek V4 Pro offer similar performance for a fraction of the cost. The solution: systematically evaluate open models before committing to a proprietary API. Gemma 4 in particular is designed to be deployed anywhere.

Mistake 3: Deploying agentic without memory

What's wrong: building an autonomous agent without a persistent memory mechanism, dooming it to repeat the same mistakes and forget every interaction. The solution: integrate a memory layer from the design stage, for example with MongoDB's new tools, rather than patching it in after the fact.

Mistake 4: Underestimating regulatory stakes

What's wrong: launching an AI-based product without checking if the models used are subject to government restrictions, particularly in the US where pre-publication vetting is being put in place. The solution: closely follow regulatory developments and build your architecture so you can swap the underlying model without a complete overhaul.

Mistake 5: Choosing your AI cloud out of habit

What's wrong: sticking with Azure out of inertia when the end of the Microsoft-OpenAI exclusivity opens up new pricing and technical possibilities. The solution: benchmark AWS, GCP, and Azure with your real workload. Prices and performance vary enough that the default choice is never the right one.

❓ Frequently asked questions

Is Claude Mythos Preview really that superior to GPT-5.5?

Yes, on paper. With 99 versus 91 overall and 100 versus 98.2 in agentic, the gap is significant. But GPT-5.5 benefits from a more mature ecosystem and integration into more products. The best model is not always the best choice in production.

Is autonomous payment agentic AI safe?

The risk exists. An agent that can spend money without human supervision requires strict guardrails: caps, validations for high amounts, complete logs. Amazon Bedrock has integrated these mechanisms, but vigilance is still required, especially in a production environment.

Subquadratic and its 12 million tokens, is that realistic?

The startup raised $29M to develop this technology, which gives the project credibility. But it is still a promise, not an available product. If it materializes, the impact will be massive for document research, legal, and large-scale code analysis.

Will US regulation slow down innovation?

Potentially yes. A pre-publication vetting system could delay releases by several months. But it could also standardize safety practices and prevent incidents that harm the entire industry. The real danger is unequal regulation between the US, Europe, and Asia, which would create market distortions.

Should you upgrade to iOS 27 for AI photo editing?

If photo editing is part of your daily workflow, yes. Apple Intelligence in iOS 27 offers a level of sophistication that rivals dedicated apps, all without leaving the Photos app. For professionals, it won't replace Photoshop, but for 90% of use cases, it's sufficient.

Recommended tools

Here is our selection of tools mentioned in this article that deserve your immediate attention:

• Amazon Bedrock: for integrating AI agents with autonomous payment capabilities via Coinbase and Stripe. Ideal if you are building e-commerce or automated purchasing workflows.
• MongoDB (agent memory tools): to equip your agents with persistent memory and accelerated vector search. Essential as soon as your agent needs to learn from its past interactions.
• Claude Mythos Preview: for complex tasks requiring the highest level of reasoning and agentic autonomy. To be reserved for cases where the cost/performance ratio justifies the raw score.
• Gemma 4: for projects requiring an open model that can be deployed anywhere, with an excellent price/quality ratio. The best option if you want to keep control of your infrastructure.
• Hostinger: if you are deploying AI projects that require reliable and affordable web hosting, Hostinger remains a safe bet in 2026 with its performance optimized for modern workloads.

✅ Conclusion

May 2026 will be remembered as the month AI shifted from the era of demonstration to the era of real autonomy — and the corresponding surveillance. Claude Mythos Preview dominates the rankings, agentic is being industrialized at Amazon, MongoDB, and UiPath, and governments are starting to test models before letting them out. To follow this evolution on a daily basis, our new recent AI tools page is continuously updated.
```