Today the agent stack went from feature to category. Across Perplexity, xAI, Mistral, and Google, the unit of consumption is no longer the model — it is an agent plus tools plus memory — and your next procurement cycle is going to be choosing between those bundles.
Perplexity
What happened
Perplexity formalized its full-stack platform with four APIs — Agent, Search, Embeddings, and the upcoming Sandbox — and extended its Memory layer into the Model Council so every model in the mix can pull personalized context. Computer, the enterprise agent, now connects directly to licensed Morningstar, PitchBook, FactSet, Daloopa, and Carbon Arc data for research and reporting workflows.
What it means for your agentic build
This is the first time procurement can credibly evaluate Perplexity as a build-your-own-agent platform rather than a chat tool, and the Computer data integrations make it a serious Bloomberg/CapIQ alternative for research-heavy finance teams. Pilot one bounded internal workflow on the Agent API this quarter and negotiate the Computer data bundle alongside your existing market-data contracts before pricing hardens.
OpenAI
What happened
An internal OpenAI reasoning model autonomously disproved a 1946 Erdős conjecture in discrete geometry that had stumped mathematicians for nearly 80 years; Fields medalist Tim Gowers called the result a milestone in AI mathematics. In parallel, OpenAI is preparing a confidential IPO filing within weeks aimed at a September public debut, and it confirmed its macOS code-signing certificate will be revoked on June 12 in the wake of the TeamPCP supply-chain attack.
What it means for your agentic build
The Erdős result is the first concrete public datapoint that frontier models are performing original analyst-grade reasoning, which raises the ceiling on what executives can responsibly delegate. The IPO and certificate events both narrow your window: lock multi-year pricing while you can, and have IT inventory every macOS device running the OpenAI app before June 12 to avoid an MDM outage.
Anthropic
What happened
Claude Code 2.1.150 shipped on May 23 with richer usage insights, keyboard-friendly diff scrolling, improved Markdown task-list rendering, and a new enterprise setting for Claude.ai cloud MCP connectors. The release lands on the heels of the Code with Claude conference in London, where Anthropic introduced “dreaming” — agents that write notes to themselves for other coding agents to consume — while reports peg the company at a $900B valuation with its first quarterly operating profit.
What it means for your agentic build
Enterprise MCP connectors on Claude.ai remove the single biggest blocker we hear from regulated buyers — “we cannot let it touch our systems” — without forcing a separate inference stack. Restart any stalled Claude Enterprise eval that died on data-egress concerns, and instrument agent runs so dreamed notes become a tracked knowledge artifact rather than a side effect.
Google DeepMind
What happened
Google I/O delivered the new Gemini 3.5 family (Flash first, Pro to follow), the Gemini Omni multimodal editing model, and Gemini Spark — Google’s first autonomous personal AI agent running 24/7 on Google servers with MCP support for third-party apps. The Ultra subscription dropped from $250 to $200 per month, and DeepMind closed an $80–90M licensing-style deal that brought in more than 20 Contextual AI researchers.
What it means for your agentic build
Gemini 3.5 Flash plus the Ultra price cut compresses cost-per-token enough to force a rebid of any long-context or multimodal workload this quarter, and Spark with MCP means Google is now a default option in agent RFPs rather than a Workspace add-on. Require every vendor in your next agent RFP to demonstrate parity with Spark’s MCP tool surface, then negotiate from the lower Google baseline.
Meta AI
What happened
Meta cut 8,000 jobs this week and reassigned another 7,000 employees onto AI projects, the most extreme AI-driven restructuring we have seen at a single enterprise software company. The same day, leaked audio of Mark Zuckerberg surfaced explaining that Meta had been tracking employees’ Gmail, coding sessions, and internal tooling to train models — turning the layoff news into a live HR and privacy story at the same time.
What it means for your agentic build
Meta’s 8,000-plus-7,000 split is now the clearest enterprise template for headcount-to-AI conversion, and the employee-monitoring leak will be referenced by works councils and labor lawyers inside your own organization within weeks. Model the ratio against your AI-impacted roles now and pre-build the communication and reassignment paths before the conversation reaches you reactively.
xAI
What happened
xAI capped a fast month on May 22 with new third-party connectors for Vercel, Canva, Gamma, and S&P Global, letting Grok deploy sites, build decks, and pull live market data without leaving the chat. Earlier in May, xAI shipped Grok 4.3 with a 1M-token context and native video input, persistent Grok Skills, the Grok Build coding agent in early access, and a new Quality Mode for the Grok Imagine API targeting enterprise creative pipelines.
What it means for your agentic build
The connector layer plus Grok Build makes xAI a credible alternative for engineering teams already living in X and Tesla ecosystems, and Quality Mode opens Imagine for marketing creative work. Run a 30-day Grok Build bake-off against the Copilot or Cursor seats you currently pay for, and validate Quality Mode on one in-flight brand asset before your next creative tooling renewal.
Cohere and Aleph Alpha
What happened
Cohere released Command A+, a 218B-parameter Apache 2.0 Mixture-of-Experts model with 25B active parameters, a 128K context window, and multimodal document processing, and acquired biopharma analytics firm Reliant AI to fold into a new North for Pharma offering. The $20B sovereign-AI merger with Germany’s Aleph Alpha is moving forward — endorsed by both Berlin and Ottawa, with Schwarz Group anchoring Cohere’s coming Series E at $600M and signed MoUs with Spain’s Indra Group and Multiverse Computing layered on top.
What it means for your agentic build
Command A+ under Apache 2.0 gives sophisticated buyers a frontier-grade self-hosted option without closed-API lock-in, and the merger effectively ends Aleph Alpha as an independent vendor. Benchmark Command A+ on workloads you currently RAG to a closed model, and if you have an active Aleph Alpha contract, demand a written continuity statement on pricing, residency, and roadmap before the deal closes.
Mistral AI
What happened
Mistral pushed Mistral Medium 3.5 — a 128B dense model that posted 77.6% on SWE-Bench Verified — into Le Chat and the Vibe coding agent as the new default, alongside Remote Agents in Vibe and a Le Chat Work mode for complex agentic tasks. Annualized revenue reportedly hit $400M (versus $20M a year ago), and the company is now in talks with European banks on a sovereign cybersecurity model pitched as an alternative to Anthropic’s Mythos.
What it means for your agentic build
Mistral has graduated into the most realistic European sovereign substitute for Anthropic and OpenAI in regulated industries, and the rumored banking cyber model targets pre-approved budget lines rather than greenfield AI spend. EU financial services, healthcare, and public-sector buyers should scope a Mistral pilot against their residency requirements this quarter and ask for early access to the cyber product.
This Week’s Structural Trends
The agent stack is the new product surface. Perplexity’s four-API platform, xAI’s connector layer, Mistral’s Remote Agents, and Google’s Spark all standardize “agent plus tools plus memory” as the unit of consumption — expect agent RFPs to replace chatbot RFPs across your portfolio by Q3 and budget your evaluation cycles accordingly.
Sovereign AI moved from talking point to deal flow. The Cohere–Aleph Alpha merger, Cohere’s Indra and Multiverse partnerships, Mistral’s European cyber pitch and Singapore expansion, and DeepSeek’s state-backed $45–50B round are all carving the global model market along jurisdictional lines. Buyers in regulated industries should explicitly map vendor sovereignty as a procurement criterion rather than an afterthought.
Headcount-to-AI conversion is now visible at scale. Meta’s 8,000 cuts plus 7,000 reassignments is the template enterprise buyers should expect replicated across software, finance, and consulting over the next two quarters. The companies that pre-build reassignment paths and internal communication arcs now will avoid the HR and legal blowback that is already attaching to Meta’s playbook.
Sources
OpenAI news (openai.com/news), Anthropic news (anthropic.com/news), Google DeepMind blog (deepmind.google/blog), Perplexity changelog (perplexity.ai/changelog), xAI news (x.ai/news), MarkTechPost (Mistral coverage), Cohere press releases, AI News Today — May 23, 2026 (buildfastwithai.com), TechCrunch and South China Morning Post (DeepSeek funding), Bloomberg and PitchBook (Cohere-Aleph Alpha merger), MIT Technology Review (Code with Claude).