Voice of AI: The week AI agents became an independent economic force 🚀
- Ralph Schwehr
- May 4
- 7 min read
There are weeks when individual headlines grab attention. And there are weeks when a pattern emerges that will later be recognized as a turning point. What happened between April 15 and May 1, 2026, unequivocally belongs to the latter category. Within 16 days, the rules of the game shifted simultaneously on four levels. Models, pricing, distribution, and software architecture were all rewritten in parallel. AI agents are no longer applications running within our software. They are becoming market participants in their own right, with their own models, their own platforms, and, as of this week, their own credit card.
The six-week cycle: Why GPT-5.5 is more than a model update
On April 23, OpenAI released GPT-5.5, setting the next benchmark in a race that is putting pressure on procurement departments worldwide. The model achieves 82.7 percent on Terminal-Bench 2.0 and 39.6 percent on FrontierMath Pro, significantly ahead of Claude Opus 4.7 with 22.9 percent. Greg Brockman, President of OpenAI, described the release as a significant step toward agent-based and intuitive data processing.
The real signal, however, isn't in the benchmarks. It's in the frequency. GPT-5.4 was released on March 5th, GPT-5.5 on April 23rd. OpenAI now releases a new model every six weeks. Anyone making strategic decisions about AI tools based on a single model will have made an outdated decision before it's even implemented. The consequence for companies is clear: it's no longer about selecting the best model. It's about building an architecture that can handle model changes.
DeepSeek V4: When the price anchor drops to fractions
Just 24 hours after OpenAI, the Chinese lab DeepSeek released its V4 generation, revealing a second front in the competition. DeepSeek V4-Pro costs $3.48 per million output tokens. OpenAI and Anthropic charge $30 and $25, respectively, for comparable performance. This is no longer a price advantage; it's a different product economics.
DeepSeek V4 was trained on Huawei's Ascend processors, which caused shares of the Chinese chip manufacturer SMIC to rise by ten percent in Hong Kong. DeepSeek V4-Pro is now the world's largest openly available model, with 1.6 trillion parameters under an MIT license. This has strategic implications. While Western vendors are raising their prices to ration demand, the Eastern market is moving in the opposite direction. For companies that want to use AI in high-volume processes, this marks the beginning of an era in which the best model is no longer the deciding factor, but rather which model delivers the right cost curve for which use case.
The end of exclusivity: Microsoft and OpenAI renegotiate
On April 27, Microsoft and OpenAI ended their exclusive partnership, which had been in place since 2019. OpenAI is now permitted to offer its models on any cloud platform, including AWS and Google Cloud. Microsoft retains a non-exclusive license to the intellectual property until 2032 and a 20 percent revenue share until 2030, although this share is now capped. The notorious AGI clause, which tied the termination of the agreement to the achievement of artificial general intelligence, has been removed.
Just one day later, AWS CEO Matt Garman announced that OpenAI models would be made available via Amazon Bedrock. Google is reviewing the new contract terms. What appears to be a minor detail in purely contractual terms is, in reality, a seismic shift. The most significant exclusive technology pact in recent economic history has been dissolved. For enterprise customers, this means freedom of choice; for strategists, it means that cloud platform decisions can now be made independently of model availability. Anyone still linking AI strategy and cloud strategy today is thinking within the framework of a world that ceased to exist on April 27th.
Salesforce Headless 360: When the software platform leaves the browser
While the model layer was being re-measured, Salesforce was simultaneously rebuilding the software layer. At TDX 2026 in San Francisco, co-founder Parker Harris posed the question of why anyone would ever log into Salesforce again. The company itself provided the answer. Salesforce Headless 360 transforms the entire platform into a collection of APIs, MCP tools, and CLI commands. Over 60 new MCP tools and 30 pre-configured coding skills give agents direct access to data, workflows, and business logic.
Even more important is the commercial consequence. Salesforce is replacing traditional per-seat pricing with consumption-based billing that directly captures agent workflows. This makes visible for the first time what has previously only been discussed theoretically. Software vendors must reinvent their business models because their primary users in the future will no longer be human licensees, but agents without seats. The question is not whether this will happen. The question is which vendor will be the first to fully implement it.
Vertical AI in everyday corporate life: Microsoft Legal Agent reaches Word
On April 30, Microsoft announced Legal Agent in Word , an AI tool specifically designed for legal workflows. Contract reviews, redline creation, and counterparty change assessments are all performed directly within the document, with native support for track changes. The tool was built by a team that moved from Robin AI, a specialized legal AI provider, to Microsoft. This move carries strategic weight, as approximately 99 percent of all legal work is done in Microsoft Word.
What's happening with Salesforce's platform is happening here with individual professional fields. Vertical AI is no longer a research topic; it's a market where the largest software companies are directly competing for specialized application areas. Every corporation collaborating with legal tech providers today faces a new question: Why purchase a separate license when the functionality is already integrated into the software running on every desktop? This question will resurface in the coming months in every profession that works with structured documents.
MoonAgents Card: A Mastercard for AI agents
On May 1st, MoonPay completed the economic logic of this development with the launch of the MoonAgents Card. This virtual Mastercard debit card allows users and AI agents to spend stablecoins directly from self-managed wallets at any Mastercard-accepting online merchant. The initiative combines MoonPay's agent infrastructure with Monavate's card-issuing platform and Mastercard's global payment network.
Stripe President John Collison predicted a surge in agent-based commerce back in February, driven by stablecoins and fast blockchains. Coinbase CEO Brian Armstrong summed it up perfectly: AI agents can't open a bank account, but they can own a crypto wallet. That's precisely where the MoonAgents Card comes in. Software subscriptions that need to be renewed without human intervention. API credits that an agent scales. Travel and logistics bookings as part of a business process. All of this requires a payment infrastructure designed not for human transactions, but for machine-based ones. With the MoonAgents Card, that infrastructure now exists.
The downside: When agents become attack vectors
As impressive as the new capabilities sound, the attack surface has grown significantly. On the same day that GPT-5.5 was released, Google's threat intelligence team published an alarming report on indirect prompt injections. Between November 2025 and February 2026, the number of malicious prompt injection attacks in a sample of the CommonCrawl database increased by 32 percent. Attackers hide instructions in ordinary web pages, which an AI agent then executes automatically as it reads them.
Forcepoint researchers found payloads containing complete PayPal transactions, specifically targeting agents with integrated payment capabilities. A second variant redirected AI-mediated financial actions to a Stripe donation link via manipulated meta tags. Traditional security architectures fail to detect these attacks because the agent operates with legitimate credentials and authorized permissions. Forcepoint articulated the strategic implications precisely: A browser-based AI that merely aggregates data poses a low risk. An agentic AI actor that sends emails, executes terminal commands, or processes payments is a highly effective target. Deploying agents without enforcing a strict separation of data and instructions introduces risks into one's own processes.
The market is catching up: 266 AI M&A deals in the first quarter
Behind the scenes, capital is driving these developments forward at an unprecedented pace. CB Insights reports 266 AI-related acquisitions in the first quarter of 2026, a 90 percent increase compared to the previous year. Nearly half of all strategic tech M&A volume exceeding $500 million now involves AI-native companies or explicitly AI-driven deals. In 2024, this share was just 25 percent. It has doubled in a single year.
Gartner forecasts global AI spending of $2.52 trillion for 2026, an increase of 44 percent. Companies that don't acquire capacity today won't be able to build it quickly enough tomorrow. Vertical AI providers, specialized talent, and proprietary datasets have become strategic assets that would take years to build. By 2026, build versus buy will no longer be a debate, but a question of speed.
Key messages
GPT-5.5 achieves 82.7 percent on Terminal-Bench 2.0 and sets a six-week release cycle, putting pressure on procurement cycles.
DeepSeek V4-Pro costs $3.48 per million output tokens, while OpenAI charges $30 for comparable performance.
Microsoft and OpenAI are ending their exclusive partnership; OpenAI models will now run on any major cloud.
With Headless 360, Salesforce replaces per-seat licensing with consumption-based pricing for agent workflows.
266 AI-driven M&A deals in Q1 2026 represent an increase of 90 percent; almost half of all major tech acquisitions are now AI-driven.
Sources
Introducing GPT-5.5 | OpenAI | April 23, 2026 | https://openai.com/index/introducing-gpt-5-5/
DeepSeek unveils V4 model with rock-bottom prices and Huawei chips | Fortune | April 24, 2026 | https://fortune.com/2026/04/24/deepseek-v4-ai-model-price-performance-china-open-source/
Microsoft and OpenAI get their exclusive deal | VentureBeat | April 27, 2026 | https://venturebeat.com/technology/microsoft-and-openai-gut-their-exclusive-deal-freeing-openai-to-sell-on-aws-and-google-cloud
Introducing Salesforce Headless 360 | Salesforce Newsroom | April 15, 2026 | https://www.salesforce.com/news/stories/salesforce-headless-360-announcement/
Word: Legal Agent in Frontier | Microsoft 365 Copilot Blog | April 30, 2026 | https://techcommunity.microsoft.com/blog/microsoft365copilotblog/word-legal-agent-in-frontier/4516218
AI threats in the wild: Prompt injections on the web | Google Online Security Blog | April 23, 2026 | https://blog.google/security/prompt-injections-web/
MoonPay Announces MoonAgents Card | PR Newswire | May 1, 2026 | https://www.prnewswire.com/news-releases/moonpay-announces-moonagents-card-enabling-ai-agents-to-spend-stablecoins-anywhere-mastercard-is-accepted-302760128.html
AI M&A Trends 2026: Why Acquirers Pay Premium Multiples | FE International / CB Insights | April 2026 | https://www.feinternational.com/blog/ai-ma-trend
Conclusion and call to action
Anyone who's been paying attention over the past two weeks will have noticed. The dimensions at which AI strategy is decided today have tripled. It's no longer just about model selection, but about architecture, platform logic, and security design. Companies that consider these three levels separately are building strategies that will be obsolete in 90 days. Those who consider them together are laying the foundation for the next ten years.
OAK AI supports companies precisely at this intersection. From AI impact analysis and strategy development to implementation in core processes. We make visible which models fit which workflow, how agents can be securely integrated into existing systems, and which investments will truly pay off in the market of tomorrow.
Let's talk about which of this week's four postponements most directly affects your business, and which ones you can afford to address later.
Write to me directly: info@oakai.de
Ralph Schwehr | oakai.de The future is not a matter of chance. It is a decision.
Comments