GPT-5.5 Ships with a Clear Bet on Agentic AI
Released on 23 April, GPT-5.5 is described by OpenAI as a model built for real work rather than benchmarks. It can take a messy, multi-part task, plan its own approach, use tools, check its own output, and keep working through ambiguity until the task is done. The gains are most pronounced in agentic coding, computer use, and early scientific research. OpenAI says it matches GPT-5.4 on latency whilst delivering meaningfully higher intelligence, which is a meaningful claim given that latency has been the persistent trade-off in every previous frontier release.
The release comes six weeks after GPT-5.4, a pace that underscores how aggressively the frontier labs are iterating right now. API pricing lands at $5 and $30 per million tokens for the standard model, and $30 and $180 per million tokens for GPT-5.5 Pro. Both are rolling out now to Plus, Pro, Business, and Enterprise subscribers on ChatGPT and Codex.
OpenAI Releases Privacy Filter: an Open-Weight Model for On-Device PII Scrubbing
Also released on 23 April, Privacy Filter is a 1.5-billion-parameter open-weight model that runs locally on your own hardware to detect and redact personally identifiable information from text. With just 50 million active parameters and a 128,000-token context window, it is designed for high-throughput data sanitisation pipelines where sending data to an external API is not acceptable. It classifies PII into eight categories (names, addresses, emails, phone numbers, account numbers, API secrets, and others) and achieves an F1 score of 97.43% on the PII-Masking-300k benchmark.
For anyone building on OpenAI models within regulated industries, this is a meaningful capability shift: it means you can sanitise training data or user inputs before they leave your infrastructure entirely. The weights are available on Hugging Face under a permissive licence.
Google Cloud Next 2026 ran 22 to 24 April in Las Vegas. The following announcements were made at the conference and have been covered in articles published on 23 and 24 April. Yesterday's digest covered the Gemini 3 model launch, the Gemini Enterprise Agent Platform components, Workspace Intelligence, and the Apple Siri partnership. Today's coverage adds competitive framing and the remaining Day 3 product announcements.
Gemini Enterprise Agent Platform: Google's Full Operating System for AI Agents
Google's biggest Cloud Next announcement reframes agentic AI from a feature into an end-to-end enterprise system. Rather than individual tools bolted together, it ships as an integrated layer covering agent design, deployment, inbox management, long-running agents, and governance controls. For businesses already using Google Cloud, this matters: agents can persist across sessions, take autonomous action on your business data, and operate within your existing security and compliance framework.
The competitive angle is significant. Google is pitching the platform as the connective tissue between your data, your people, and your goals, positioning it directly against Microsoft's Azure AI Foundry and OpenAI's Workspace Agents for enterprise agentic workloads. With all three competing for the same integration budget from the same enterprise buyer, the differentiator is increasingly going to be governance and auditability rather than raw capability.
Veo 3.1 Lite Enters Public Preview with Temporal Consistency for Multi-Scene Video
Google announced Veo 3.1 Lite at Cloud Next, moving its video generation model into public preview on Vertex AI. The headline feature is Temporal Consistency, which allows brands and creators to maintain the same characters and environments across multiple clips without manual prompting tricks. This has been a persistent pain point for anyone trying to produce multi-scene video content with AI: until now, keeping a consistent protagonist across clips required significant post-production work.
Veo 3.1 Lite makes this accessible via API, which opens the door to automated video production pipelines for marketing, e-learning, and short-form content at a scale that would be prohibitively expensive with traditional production methods.
Lyria 3 Pro Brings Stem-Level Audio Control to Google's Music Generation Model
Google also unveiled Lyria 3 Pro at Cloud Next, positioning it as the first music generation model designed for professional workflows rather than quick demos. The defining capability is audio stem generation: the model can produce isolated vocals, bass, drums, and instrument tracks separately, which means output can feed directly into a digital audio workstation for mixing without needing to separate stems manually afterwards.
For content creators and agencies who rely on licensed music or commission original scores, this is a meaningful step towards AI-generated music that is actually usable in commercial contexts rather than just as background filler. The distinction from most existing music generation models, which compete on artistic realism, is that Lyria 3 Pro competes on professional workflow integration instead.
Gemini 3.1 Flash Image (Nano Banana 2) Arrives as a High-Fidelity UI and Asset Generator
Google announced Gemini 3.1 Flash Image, internally referred to as Nano Banana 2, as a new image generation model optimised specifically for generating high-fidelity user interface mockups and visual assets rather than photo-realistic imagery. This is a notably different positioning from most image generation models, which compete on artistic realism. For product designers and developers prototyping UI layouts, having a model that understands screen composition, component spacing, and interface conventions is a genuinely different use case.
It is available via the Vertex AI API and integrates into Google's broader Gemini Enterprise Agent Platform, meaning agents can now generate visual assets and UI mockups as part of an automated workflow rather than as a standalone step.
Built-in Memory for Claude Managed Agents Enters Public Beta
If you are building production agents on the Claude API, this is the most relevant Anthropic update in the last 24 hours. Memory for Claude Managed Agents went to public beta on 23 April, announced via the @ClaudeAI account on X before mainstream press coverage arrived. The feature gives agents a persistent, workspace-scoped document store that Claude can read and write across sessions, so an agent can remember a user's preferences, prior decisions, or accumulated knowledge without you having to pipe that context in manually each time.
Two details stand out for anyone thinking about production use. First, every mutation to a memory creates an immutable version, giving you a full audit trail and the ability to restore a prior snapshot or scrub sensitive content. Second, there is a concurrent-write protection mechanism via a content hash precondition, which matters when multiple agent sessions might try to update the same memory simultaneously. The feature ships under the existing managed-agents-2026-04-01 API header, meaning no version bump is needed if you are already on the managed agents platform.
Industry Themes
Three patterns stand out from today's news. First, persistent memory is moving from a research concept to a production building block at every major lab simultaneously. Anthropic announced Memory for Claude Managed Agents in public beta via the @ClaudeAI account on X on 23 April (a social media exclusive that surfaced before mainstream press coverage) while OpenAI shipped GPT-5.5 with an explicit focus on agents that can carry context and keep working when you close the browser. The agent that can remember is now the baseline expectation, not a premium feature.
Second, the race to own enterprise agentic infrastructure is heating up fast. Google's Gemini Enterprise Agent Platform at Cloud Next, OpenAI's Workspace Agents (announced 22 April), and Anthropic's Managed Agents are all competing for the same integration budget from the same enterprise buyer. The differentiator is increasingly going to be governance and auditability rather than raw capability, which is why Google's emphasis on audit trails and compliance controls in its platform announcement is more strategically significant than it might first appear.
Third, privacy infrastructure is becoming a first-class product. OpenAI's open-weight Privacy Filter signals that frontier labs now recognise enterprise adoption is blocked not just by capability gaps but by data governance concerns, and they are shipping tooling to address those at the model level rather than as a policy layer on top. Expect this to become a competitive differentiator in regulated industries over the next twelve months.