GPT-Rosalind Is OpenAI's First Purpose-Built Domain Model, Aimed at Drug Discovery and Genomics
Named after the chemist Rosalind Franklin, GPT-Rosalind is OpenAI's first model designed from the ground up for a specific scientific domain rather than adapted from a general-purpose base. It launched on 17 April through a gated Trusted Access programme, with initial partners confirmed as Amgen, Moderna, Thermo Fisher Scientific and the Allen Institute. The model is optimised for evidence synthesis, hypothesis generation and experimental planning across genomics, protein engineering and chemistry.
Alongside it comes a free Codex Life Sciences plugin that connects researchers to more than 50 scientific tools and data sources without leaving their existing workflow. Broad public access is not on the table yet, but the gated-partner approach is a deliberate template OpenAI is likely to reuse across other regulated domains. If you work anywhere near biology, genomics or translational medicine, the access programme is worth applying to now while the initial cohort is small.
Codex Gains Native Computer Use on macOS, 90-Plus New Plugins and a Cross-Session Memory Layer
The 17 April Codex update is a step-change rather than an incremental release. Native computer use on macOS means Codex can now see your screen, move the cursor and type inside any application while you work in parallel. In practice, that means end-to-end tasks that previously required constant supervision are now delegable: reviewing pull requests, editing frontend designs in the browser, running SSH sessions to remote devboxes and processing multi-file terminal workflows.
The update ships alongside more than 90 new plugins, including Atlassian Rovo, GitLab, CircleCI and the full Microsoft 365 suite. A memory layer has also been added, learning user preferences and retaining context across sessions. One practical note for UK and EU users: computer use is listed as "coming soon" in this region, so the most significant capability has not yet reached all subscribers. Worth checking your rollout status before planning workflows around it.
Claude Design Launches in Research Preview: From a Text Brief to a Shareable Visual in One Step
Anthropic published Claude Design to its news page on 17 April. It is a Labs product that takes a plain-text description and produces polished visual outputs (prototypes, one-pagers, slides and design mockups) without requiring the user to open any design application. Outputs are editable and shareable directly within the product. Powered by Claude Opus 4.7, it is available as a research preview to Pro, Max, Team and Enterprise subscribers from today.
The most interesting part of the product is not what it generates but what happens next: a built-in handoff to Claude Code lets a visual sketch transition directly into a working build. That workflow (brief to mockup to live code) does not currently exist in Figma or Canva, and it is the detail most likely to shift how technical founders and product teams prototype. In its current form this is not a replacement for professional design tooling; think of it as filling the gap between a blank page and a brief you could hand to a designer or developer.
Claude Code Adds Auto Mode, an xhigh Compute Tier for Opus 4.7 and Finer Task Controls
The Claude Code release accompanying the Claude Design launch brings several updates that regular CLI users will notice immediately. Auto mode is now live for Max subscribers, scaling model effort dynamically without manual flags. There is also an xhigh compute tier for Opus 4.7 specifically, intended for tasks where thoroughness matters more than speed. Two new slash commands round out the control surface: /effort for adjusting task intensity and /ultrareview for deeper code review passes. A /usage command, available from 18 April, shows a breakdown of what is driving token consumption and surfaces cache-miss warnings before resuming older sessions.
On the quality-of-life side: Windows support has been improved, permission prompt frequency has been reduced, and a terminal-matching theme has been added for those who care about their tooling aesthetics. None of these changes affect pricing, but the combination of Auto mode and xhigh creates a more sensible range of effort controls than the previous binary approach.
Grok 4.3 Beta Arrives for SuperGrok Heavy Subscribers with 1 Trillion Parameters and Native Video Input
xAI dropped Grok 4.3 Beta into the grok.com model selector on 17 April with no press release, no blog post and no social media campaign: it simply appeared with an "Early Access" label for SuperGrok Heavy subscribers paying $300 per month. The model is a substantial architectural jump. At 1 trillion parameters it is double the size of Grok 4.20, and it was trained for longer. Native video input lets you send a video clip directly into the conversation, making it the first Grok model capable of reasoning over moving images.
The document generation capability is the detail most likely to matter for business users: Grok 4.3 can produce formatted PDFs, populated spreadsheets and PowerPoint-style decks directly from a chat prompt, with early testers reporting outputs usable enough to share without further editing. xAI has committed to shipping improvements throughout the beta period. A full rollout to a wider subscriber tier is expected in May. Given that both xAI items today surfaced via company channels before any press picked them up, the developer documentation is the place to watch for further announcements.
xAI Quietly Opens Speech-to-Text and Text-to-Speech APIs to Developers with No Announcement
Also on 17 April, and again without any formal announcement, xAI made its Speech-to-Text and Text-to-Speech APIs available through its standard API gateway. These are the same speech models that power Grok Voice, Tesla's vehicle infotainment system and Starlink's customer support infrastructure, now available as standalone developer endpoints. The STT API supports 25-plus languages across 12 audio formats with word-level timestamps, speaker diarisation and multichannel audio. Two operating modes are available: real-time WebSocket streaming for voice applications and batch REST for asynchronous processing.
Pricing is $0.10 per hour for batch and $0.20 per hour for streaming. That puts xAI well below the typical pricing of ElevenLabs and Deepgram for comparable capabilities, and the lack of any press coverage on the day means most developers will not have clocked this yet. If you are building voice applications or evaluating transcription providers, checking xAI's developer documentation now is worth ten minutes of your time.
Personal Computer for Mac Reaches General Availability for All Max Subscribers
Perplexity's Personal Computer feature, first previewed in March, reached general availability for all Max subscribers on 17 April. The product is an always-on AI agent that lives on your Mac and can operate any native application, your local files, browser sessions and cloud connectors through the Perplexity interface. Activation is a double-tap of both Command keys; it accepts text or voice input. Any Mac running macOS 14 Sonoma or later is supported, with a Mac mini as the recommended hardware for continuous background operation.
At $200 per month this sits firmly in the prosumer and professional bracket. The genuine shift it represents is not the AI capability itself but the operating model: rather than opening a chat window and describing a task, the agent can see and act across all your applications simultaneously, without screen-sharing workarounds. For anyone managing complex multi-app workflows, that is a meaningfully different proposition from anything that existed twelve months ago.
Agent Memory Enters Private Beta as a Managed Alternative to Context-Stuffing
Cloudflare announced on 18 April that Agent Memory is entering private beta as a managed service. Rather than dumping an entire conversation history into the context window each time, it extracts relevant information from past interactions and surfaces it when the agent actually needs it. That addresses one of the most tedious engineering trade-offs in production agent work today: context window costs versus continuity across long sessions.
Cloudflare is positioning it as a serverless memory layer that sits alongside its existing Workers AI and AI Gateway products. If you are already in the Cloudflare Workers ecosystem and building agents that need to recall user preferences, prior decisions or accumulated knowledge across sessions, the private beta is the obvious next step. The managed approach also removes the need to design and maintain a custom retrieval pipeline, which is usually the part of agent memory that consumes the most engineering time in practice.
Industry Themes
The Friday announcements this week reveal a clear competition to own the desktop. OpenAI shipped computer use for Codex, Perplexity opened Personal Computer to all Max subscribers, and both are making the same underlying bet: that the next significant unlock for AI productivity is not a smarter model in a chat window but a persistent agent that can operate your machine on your behalf. The practical distinction between these products is still thin for most users, but the race to capture the desktop layer (where actual work happens) is accelerating faster than the model benchmarks suggest.
The second pattern worth watching is how xAI is building. Both Grok 4.3 Beta and the speech APIs appeared in xAI's platform on 17 April with no press release, no blog post and no social media push. This looks deliberate: rapid iteration under the radar, letting developer uptake validate the product before making public commitments. The STT/TTS pricing ($0.10 per hour batch) is a direct challenge to ElevenLabs and Deepgram at significantly lower rates, and it surfaced via company documentation before mainstream press covered it at all. For anyone building on AI infrastructure, monitoring developer changelogs directly has become as important as reading the news.
Finally, Claude Design points to an emerging category that has not been clearly named yet: AI-native visual creation that is not a design tool. The product is aimed at people who need something polished quickly and do not want to open Figma or Canva to get there. The output quality question remains open for anything beyond internal documents, but the handoff to Claude Code (letting a prototype become a real build in one step) is a workflow that no conventional design application currently offers. That single integration detail may prove to be the product's most durable differentiator.