Last week’s canvas was launch velocity — GPT‑5.5, workspace agents, Anthropic’s five‑gigawatt AWS story, Google Next. This week is wiring: OpenAI’s amended Microsoft agreement explicitly clears multicloud distribution while keeping Azure “first” when capable; the same strategic window put GPT‑5.5, Codex, and Bedrock Managed Agents into limited preview on AWS; Symphony open‑sourced the spec for issue‑tracker‑native agent fleets; and Advanced Account Security made phishing‑resistant login the posture for the highest‑risk accounts — including a June 1 requirement for Trusted Access for Cyber individuals. Anthropic answered with creative‑industry connectors and a Sydney beachhead; Google pushed Gemini into ~four million GM dashboards and refreshed the consumer app; Meta’s Muse Spark keeps Wall Street asking whether proprietary frontier models can sit beside Llama’s open ecosystem.
Non‑exclusive through 2032 — Microsoft’s license to OpenAI IP per the Apr 27 amendment.
500% — OpenAI’s cited PR throughput lift on some teams after Symphony-style orchestration (internal observation; not a universal guarantee).
~4 million — GM’s count of U.S. eligible MY2022+ vehicles for the Gemini rollout described Apr 28.
01 · OpenAI · Corporate structure
April 27: the “next phase” post replaces revenue‑share complexity with a cleaner story — Azure stays primary unless Microsoft cannot support required capabilities and opts out; OpenAI may serve all products on any cloud; Microsoft’s model and product IP license runs through 2032 but is now non‑exclusive; Microsoft stops paying revenue share to OpenAI while OpenAI’s payments to Microsoft continue on a capped schedule through 2030. For CIOs, the subtext is procurement pluralism without giving up Redmond’s strategic stake.
02 · OpenAI × AWS · Distribution
April 28: OpenAI and AWS announced limited preview expansion — frontier models (including GPT‑5.5) on Amazon Bedrock, Codex configurable to consume Bedrock‑hosted models (CLI, desktop, VS Code extension called out), and Amazon Bedrock Managed Agents powered by OpenAI for multi‑step enterprise workflows with AWS governance framing. OpenAI states more than four million weekly Codex users; Bedrock positioning emphasizes existing security, identity, and procurement rails.
Build with OpenAI models inside AWS accounts — pitched for teams that standardize on Bedrock APIs and enterprise agreements.
Configure Codex to use Bedrock as the inference provider; data processed in Bedrock per OpenAI’s post, with AWS commit attribution for eligible customers.
Amazon’s managed agent product surface powered by OpenAI — orchestration, tool use, and production guardrails sold as a package.
GPAI — general-purpose AI; when agents touch regulated data, map Bedrock logging to your DPIA and retention policies, not just the vendor press diagram.
OpenAI — OpenAI on AWS AWS — Bedrock Managed Agents (OpenAI)03 · OpenAI · Engineering culture
April 27: OpenAI published Symphony — an open specification (repository includes SPEC.md) for a service that polls Linear (in v1), opens per‑issue workspaces, and runs coding agents continuously. The narrative claim: some internal teams saw landed pull requests rise roughly 500% in the first three weeks after adopting ticket‑native orchestration versus juggling three‑to‑five interactive Codex sessions. The post is explicit about tradeoffs: less mid‑flight nudging, more investment in harness guardrails and CI shepherding.
Symphony’s stated design boundary: a scheduler and tracker reader — not a full no‑code workflow engine. Implementations own approval policies and sandbox posture.
“Once work is abstracted this way, tickets can represent much larger units of work.”— OpenAI engineering, Symphony orchestration post
04 · OpenAI · Identity & threat
April 30: Advanced Account Security is an opt‑in ChatGPT setting merging phishing‑resistant sign‑in (passkeys or security keys, passwords disabled), stricter recovery (no email/SMS recovery — backup passkeys, recovery keys; OpenAI Support will not recover accounts enrolled this way), shorter sessions with login alerts and session review, and automatic training exclusion for those accounts. A Yubico bundle is offered at preferred pricing; any FIDO keys remain valid.
Trusted Access for Cyber: individuals in the program using the most permissive cyber models must enable Advanced Account Security by June 1, 2026 (or organizational SSO attestation in lieu).
OpenAI — Advanced Account Security OpenAI — Cybersecurity action plan (PDF)05 · OpenAI · Civic safety
April 28: OpenAI published a long‑form explanation of how ChatGPT handles potential violence — model refusals, automated detection (classifiers, reasoning models, hashes, blocklists), human contextual review, bans, law enforcement referral in credible imminent‑harm cases, and parental controls / upcoming trusted‑contact notifications for acute distress signals. The piece is as much governance transparency as product marketing.
OpenAI — Community safety OpenAI — Usage policies06 · Anthropic · Connectors
April 28: Anthropic shipped a connector set aimed at creative professionals — official hooks into Ableton docs, Adobe Creative Cloud surfaces, Affinity (Canva), Autodesk Fusion, Blender (MCP connector co‑released with Blender Foundation), Resolume, SketchUp, and Splice sample search. The post positions Claude as tutor, batch operator, and pipeline bridge — not a replacement for creative direction — and notes Anthropic made a one‑time donation to Blender (May 1 update on contribution channel).
Educational pilots named: RISD Art & Computation, Ringling “Fundamentals of AI for Creatives,” Goldsmiths Computational Arts — feedback loop for connector UX.
Anthropic — Claude for Creative Work Model Context Protocol — Intro07 · Anthropic · Asia-Pacific
April 27: Anthropic named Theo Hourmouzis (ex‑Snowflake SVP ANZ/ASEAN) GM for Australia and New Zealand and formally opened its Sydney office — framed alongside Tokyo, Bengaluru, and upcoming Seoul. The release cites Commonwealth Bank, Quantium, ANU, Murdoch Children’s, Garvan, Curtin, plus recent Canva and Xero platform collaborations and YMCA South Australia as a Claude for Nonprofits partner.
Anthropic — Sydney / ANZ Anthropic — Australia MOU08 · Google × Automotive · Scale
April 28 (GM press): Gemini will roll out to approximately four million U.S. model‑year 2022 and newer Cadillac, Chevrolet, Buick, and GMC vehicles with Google built‑in — replacing the prior Google Assistant experience with a more conversational layer delivered over months via software update (OnStar + Play Store + US English + opt‑in called out as requirements).
09 · Google · Consumer AI
April 2026 Gemini Drops (10th edition): native Gemini for Mac, Notebooks / NotebookLM integration for project organization, Personal Intelligence expansion with region eligibility caveats, Lyria 3 Pro for up to roughly three‑minute tracks, and in‑chat 3D models / charts visualization. The through‑line is the same as GM: Gemini as an ambient layer across devices and contexts.
Google — Gemini Drop April 2026 Google — Gemini app on Mac10 · Meta · Strategy
Meta’s Muse Spark, introduced from Meta Superintelligence Labs leadership including Alexandr Wang, is positioned as a closed frontier model path distinct from free‑to‑weights Llama — with Meta signaling paid developer access over time. CNBC’s late‑April analysis ties investor focus to Q1 earnings commentary, cites Arena leaderboards as a rough public scoreboard, notes Meta’s disclosed May 20 workforce reduction plan (~10%, ~8k roles) alongside very large guided AI capex, and contrasts ad‑driven returns with pure‑play lab valuations.
CNBC — Muse Spark / strategy Meta AI — Muse Spark blog11 · Policy · Compliance rail
As OpenAI sells Bedrock agents and Anthropic sells connectorized creative workflows, European deployers still phase obligations for providers, deployers, and GPAI systemic models per the Commission’s published implementation calendar — the legal clock does not care which hyperscaler invoices the tokens.
Map banned practices, high‑risk classifications, and transparency duties to each agent surface you ship — Bedrock and Azure both need evidence packs.
Use the official EU AI Act Service Desk timeline for effective dates rather than vendor slide decks.
12 · Infra · Comparison
| Surface | What shipped / previewed | Primary read |
|---|---|---|
| Azure | Still “first ship” cloud per amended Microsoft language when capable. | Default enterprise lane for many incumbents. |
| AWS Bedrock | GPT‑5.5, Codex provider mode, Managed Agents (OpenAI‑powered) — limited preview. | Commits, IAM, and privateLink narratives strengthen. |
| Vertex / multi‑vendor | No new flagship drop this week in this edition’s sourcing set; prior Next ’26 stack remains context. | Competition is now distribution + governance bundles. |
13 · Steel man
Optimistic read: Bedrock + amended Microsoft terms let regulated shops pick the control plane that matches their SOX / HIPAA boundary without freezing model choice.
Skeptical read: Limited previews, dual vendor SKUs, and Symphony-style orchestration explode the attack surface — every ticket becomes a CI/CD blast radius unless harness discipline keeps pace.
Consumer read: Four million dashboards with Gemini sound like distribution; they are also four million microphones where misrouting liability still sits with OEMs and drivers, not the model card.
14 · Forward calendar
Anchors from primary sources cited in this edition — confirm times in your own calendars and vendor portals.
Individual TAC participants must enable Advanced Account Security (or org SSO attestation) per OpenAI’s Apr 30 security post.
Free preview ends for ChatGPT workspace agents — carryover anchor from OpenAI’s prior workspace agents announcement; verify billing before fleet rollout.
CNBC reports Meta targeting ~10% reduction (~8k roles) as stated in company communications — watch for AI capex commentary on the same cadence.
Annual developer conference — expect Gemini, Android, and cloud agent follow-ons; check official Google I/O schedule for session times.
OpenAI on AWS and Managed Agents are limited preview — enterprise architects should expect waitlists and regional availability deltas.
GM states months-long staggered delivery to eligible vehicles — infotainment prompt when ready.