Data, ML & AI in April 2026: The Month in Nine Stories

If March 2026 was about new model launches, April was about who controls the layer underneath them. Every hyperscaler made a move on "agentic data" within a single week (April 21-28), Microsoft and OpenAI ended their exclusivity, and Anthropic became the first lab to build a frontier model and decide not to ship it. Here are the nine stories that mattered.

1. Anthropic shipped Claude Opus 4.7 - and shelved Claude "Mythos"

On April 16, Anthropic released Claude Opus 4.7 across the Claude API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry, with prices unchanged at $5 / $25 per million input/output tokens.

Nine days earlier, the bigger story landed: Anthropic confirmed the existence of an internal model codenamed "Mythos" - and announced it would not be released to the public. The model independently identified thousands of zero-day vulnerabilities during testing, triggering Anthropic's ASL-4 safety protocol. Access has been restricted to roughly 50 partner organizations under a program called "Project Glasswing."

Why it matters: This is the first time a frontier lab has built a model and chosen not to ship it. Expect this to fuel real conversation about who decides what's "too capable to deploy."

2. Open-weight models caught up - fast

Three open-weight releases in the first week of April closed the gap with the proprietary frontier:

• Google Gemma 4 (April 2) - Sizes from 2.3B to 31B parameters, Apache 2.0, natively multimodal. The 31B model now ranks third globally on the Arena AI leaderboard among open models. Codeforces ELO jumped roughly 20× over Gemma 3.
• Meta Llama 4 Scout & Maverick (April 5) - Meta's first MoE Llama models, with 10M and 1M token context windows respectively.
• GLM-5.1, Qwen 3.6-Plus, DeepSeek V3.2, Arcee Trinity - All landed within the same window, several at near-frontier benchmark scores for a fraction of the per-token cost of GPT or Claude.

Why it matters: The "do we pay for an API or run our own?" calculus is shifting. For many enterprise workloads, a self-hosted open-weight model is now within single-digit percentage points of the best paid API.

3. Meta abandoned its open-source AI strategy

On April 8, Meta launched Muse Spark, its first major proprietary AI model - a sharp reversal of the Llama-era playbook. Developed by Meta Superintelligence Labs (the unit Alexandr Wang took over after the $14.3B Scale AI deal last June), the model is closed and offered to "select partners" via private API. Meta says future versions may be open-sourced.

The context: Meta's AI capex for 2026 is now projected at $115-135 billion, roughly double last year. The pressure from OpenAI, Anthropic, and Google has become too high to keep giving frontier weights away for free.

Why it matters: If you architected anything assuming Meta's best models would always be free, that assumption just expired.

4. The Hyperscaler Agentic Data Week (April 21-28)

In a single seven-day window, every major data platform made a coordinated push to position itself as the "control plane" for AI agents acting on enterprise data.

Monday, April 21 - Snowflake announced major expansions to Snowflake Intelligence (an agentic experience for business users that learns individual workflows) and Cortex Code (an agent development environment with a CLI, VS Code extension, and a Claude Code plugin in development). MCP-based interfaces now connect Snowflake to Google Workspace, Jira, Salesforce, and Slack. Snowflake says more than 9,100 customers now use its AI products weekly.

Wednesday, April 22 - Google unveiled Agentic Data Cloud at Cloud Next, built on three pillars: a Knowledge Catalog that auto-infers business semantics from query logs, a cross-cloud lakehouse that lets BigQuery query Iceberg tables on AWS S3 with no egress fees, and a Data Agent Kit that drops MCP tools into VS Code, Claude Code, and the Gemini CLI.

Earlier in March/April - Microsoft at FABCON 2026 launched Fabric IQ (semantic context layer over OneLake), made Fabric data agents generally available, introduced a unified Database Hub, and added Runtime 2.0 with Spark 4.x and Delta Lake 4.x. Microsoft Foundry also added Anthropic's Claude models alongside OpenAI's GPT family - and shipped its own MAI-Transcribe-1, MAI-Voice-1, and harrier-oss-v1 embedding models in Foundry Labs.

Tuesday, April 28 - AWS held its "What's Next" event with Amazon Quick, a desktop AI assistant connecting to local files, Google Workspace, Microsoft 365, Slack, Zoom, and Salesforce. It also launched Amazon Bedrock Managed Agents powered by OpenAI and made AWS Interconnect generally available - Layer 3 private connectivity between AWS VPCs and other clouds (Google Cloud now, Azure and OCI later in 2026).

Why it matters: Four vendors, the same week, the same pitch: the next workload isn't a human running a dashboard - it's an autonomous agent acting on data 24/7. The vendor lock-in conversation just got a lot more interesting.

5. OpenAI and Microsoft formally ended exclusivity

The biggest cloud story of the month: on April 27, OpenAI and Microsoft restructured their relationship to allow OpenAI to serve customers across any cloud provider. The very next day, AWS announced OpenAI's models - including the Codex coding agent - would be available via Amazon Bedrock, with a new Bedrock Managed Agents powered by OpenAI service in limited preview.

This follows the November 2025 disclosure of OpenAI's $38 billion compute commitment with AWS, later expanded with a $50 billion AWS investment in OpenAI and a 2-gigawatt commitment to AWS's custom Trainium chips.

Why it matters: For four years, "OpenAI on Azure" was a strategic moat for Microsoft. That moat is now gone. Expect aggressive pricing and enterprise pursuit from both Bedrock and Vertex AI through the rest of 2026.

6. Databricks shipped agent infrastructure on every cloud

Databricks' April release notes read like a quiet but methodical agent-infra rollout: GPT-5.5 and GPT-5.5 Pro added to Mosaic AI Model Serving, Lakeflow Designer in public preview (visual + natural-language pipeline building), Data Classification GA (agentic auto-tagging in Unity Catalog), the Supervisor API in beta for single-call multi-agent systems, query-based connectors for Lakeflow, an Excel add-in, and customer-managed key encryption for model serving artifacts. Genie also reached in-Geo availability in Japan and Korea.

The IPO meanwhile slipped: Databricks pushed its IPO to H2 2026, with $5.4B ARR and 20,000+ customers but no S-1 filing as of mid-April.

7. MongoDB doubles down on AI-native database

Two April moves signal where new CEO CJ Desai (in the role since November 2025) is taking MongoDB's product strategy:

• April 22 - Leadership reshuffle. Pablo Stern joins from ServiceNow as Chief Product Officer, AI and Emerging Products, overseeing Search, Vector Search, and the Voyage AI portfolio. Ben Cefalo was promoted to CPO of Core Products. At ServiceNow, Stern scaled IT Operations Management from $100M to over $1B in five years.
• April 23 - MongoDB announced a €74M Ireland expansion with 200 new engineering and AI roles and a new Cork office. The pitch: helping its 65,200+ customers move from "AI experimentation to production."

The backdrop: MongoDB closed FY2026 (ending January) at $2.46B in revenue, up 23% YoY, with Atlas up 29%. The Voyage AI acquisition (closed in early 2025) is now central to the company's "unified data intelligence layer" pitch - embedding/reranking models built into the database, so AI applications don't need a separate vector store. The MongoDB MCP Server (in public preview since mid-2025) connects MongoDB deployments directly to Cursor, Windsurf, GitHub Copilot, and Claude.

Why it matters: The operational database (where your apps actually write) and the AI/vector layer (where retrieval happens) are collapsing into one product. Whether you bet on MongoDB, Postgres + pgvector, or a dedicated vector store like Pinecone, the architectural decision now lives at the database layer, not the application layer.

8. Infrastructure-as-code: IBM consolidates HashiCorp

Following IBM's HashiCorp acquisition closing earlier this year, the consolidation is showing up in the product roadmap. April milestones:

• April 1 - Terraform Enterprise on the Replicated Native Scheduler reached end of support, forcing customers to migrate to Docker, Kubernetes, Podman, or Nomad deployment.
• April 1 - AWS permission delegation went GA in HCP Terraform.
• April 15 - Terraform 1.15.0-rc2 added Windows ARM64 builds.
• April 20 - Terraform 1.14.9 shipped.
• April 22 - Pre-written Sentinel policies for ISO 27001 launched, plus a major case study from Banco Bradesco (80 days to 5 days for digital product delivery).

Separately, IBM-HashiCorp announced it is ending external language support for Terraform, pushing teams toward HCL-only workflows. For shops with Python/Ruby Terraform integrations, that's a migration to plan now.

Why it matters: The tooling stack underneath all the AI agent excitement is consolidating fast. Expect more "rationalization" decisions like the language deprecation as IBM integrates HashiCorp into its enterprise software portfolio.

9. The money: a $300B quarter, four deals doing most of the work

Q1 2026 broke every venture funding record on file:

• $300 billion in global venture funding in Q1 - roughly 70% of all of 2025's spending in a single quarter (Crunchbase).
• AI startups absorbed roughly 81% of that capital. Four deals - OpenAI ($122B), Anthropic ($30B), xAI ($20B), Waymo ($16B) - totaled $188B, more than all of 2024's global venture funding.
• Cognition AI (the company behind Devin) entered talks on April 23 to raise at a $25B valuation, more than double its prior mark (Bloomberg).
• Combined Q1 AI capex from the "Magnificent Seven" hit roughly $78 billion, up 45% year-over-year. Microsoft alone added nearly 1 GW of data center capacity in the quarter.
• dbt Labs + Fivetran dropped their 2026 State of Analytics Engineering Report in late April: 83% of analytics engineers now rate "increasing trust in data" as a top priority (up from 66%), and 71% prioritize speed (up from 50%). Their $600M ARR merger is still on track to close in mid-to-late 2026.

What to watch in May

• OpenAI's first non-Azure deployments at scale on AWS Bedrock - does pricing actually drop?
• Snowflake vs. Databricks vs. Fabric vs. BigQuery agent benchmarks - the marketing is identical; real performance comparisons should start surfacing.
• dbt + Fivetran regulatory review progress.
• Databricks S-1 watch.
• DeepSeek's next flagship, reportedly delayed by export-control chip constraints (Reuters).
• OpenAI's "GPT-5.5 / GPT-6" ("Spud") - does it finally ship after its April window slipped?
• The Mythos precedent - does another lab follow Anthropic in withholding a frontier model?

Sources: Anthropic, OpenAI, AWS, Microsoft, Google Cloud, Snowflake, Databricks, MongoDB, HashiCorp, dbt Labs, Fivetran, Bloomberg, CNBC, VentureBeat, Crunchbase. All figures and dates as of late April 2026.