Claude Opus 4.7 Ships with a Bigger Vision Window, Stronger Coding, and a Tighter Prompt Contract
Claude Opus 4.7 is live today across Claude, the API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry and Snowflake Cortex AI, a simultaneous five-platform launch that removes the usual wait for hosted availability. API pricing is unchanged at five dollars per million input tokens and twenty-five per million output. The most immediately useful improvement for document and UI-heavy workflows is the image ceiling: it has jumped to roughly 3.75 megapixels, meaning dashboard screenshots, scanned PDFs and dense UI layouts can now go in without manual downscaling. Early testers are reporting a 13 per cent lift on coding benchmarks over Opus 4.6, and Anthropic is explicitly positioning longer, less supervised agent runs as in-scope for this release.
One practical caveat worth flagging before you cut over: Opus 4.7 follows prompts more literally than its predecessors. Older prompt harnesses that relied on loose interpretation to fill in gaps are likely to behave differently, so run your existing prompts against it in a test environment before treating it as a drop-in replacement.
Snowflake published its own announcement simultaneously with the Anthropic launch, confirming Opus 4.7 is in public preview in the US and EU via Cortex Code, Cortex AI Functions, the Cortex REST API and Snowflake Intelligence. For regulated teams, the key point is governance: Claude runs inside Snowflake's existing data perimeter, so the stronger vision and agentic capabilities are accessible without routing data to a separate endpoint. See the full Snowflake story below.
GPT-Rosalind Enters Research Preview for Drug Discovery and Life Sciences
OpenAI has launched GPT-Rosalind in research preview for a small group of vetted partners: Amgen, Moderna and the Allen Institute for Biomedical Research. The model is tuned for evidence synthesis, hypothesis generation and experimental planning rather than general-purpose conversation. Access is restricted, the scope is narrow, and the partner list is tightly controlled. For anyone outside biotech, the practical impact today is minimal.
The pattern itself is worth logging: narrow scope, vetted partners, restricted access. OpenAI is deliberately separating frontier capability launches from consumer product releases, which is a template likely to repeat across pharmaceuticals, finance and government over the next 12 months. If you work at the boundary of AI tooling and regulated industries, this is the structural model to watch.
Codex Gains Marketplace Installs, Richer Terminal Memory and Tougher Sandboxing
The 16 April Codex changelog has a handful of items worth wiring in for anyone using it in daily agent work: marketplace installs for approved extensions, expanded MCP and app-server APIs, and hardening of both the devcontainer and macOS sandboxes. The terminal UI also gains memory and history improvements, which means long sessions are less likely to lose their context on a crash or interrupted run.
Pricing is unchanged. But the combination of better sandboxing, wider extension support and more reliable session memory makes the case for delegating longer, less supervised tasks more credible than it was a week ago.
Gemini Robotics-ER 1.6 Confirmed for Spot Integration, with Broader Coverage on 16 April
Broader coverage on 16 April confirmed additional detail on Gemini Robotics-ER 1.6 and its integration into Boston Dynamics' Spot robot via the Orbit AIVI-Learning pipeline. The headline capability, reading analogue instruments such as pressure gauges and sight glasses, addresses a problem that has been stubbornly resistant to multimodal models until now. Internal tests show a 10 per cent improvement on video hazard detection over Gemini 3.0 Flash, with better spatial reasoning across pointing, counting and task-success detection.
The model is available to developers now through the Gemini API and Google AI Studio. If you are exploring physical AI or industrial inspection use cases, this is the most capable open-access model for the problem at the moment.
XChat iOS Launch Pushed from 17 April to 23 April
XChat was scheduled to debut on iOS today as xAI's Grok-powered messaging push. The launch has slipped to 23 April after another round of internal testing. The feature set is unchanged: end-to-end encryption, in-chat Grok summarisation and support for larger group calls. xAI has not indicated any scope changes, only the timing.
If you were evaluating XChat as a lightweight AI-assisted alternative to existing team messaging tools, the slip is short enough to wait out rather than plan around. Worth revisiting next week.
Claude Opus 4.7 Lands on Snowflake Cortex AI as a Same-Day Launch Partner
Snowflake published its own announcement at the same time as the Anthropic launch, making this one of the cleaner simultaneous-availability stories in recent memory. Opus 4.7 is in public preview in the US and EU, accessible through Cortex Code, Cortex AI Functions, the Cortex REST API and Snowflake Intelligence.
For teams in regulated industries, the practical value is that Claude runs inside Snowflake's existing governance envelope. You can invoke the stronger vision ceiling and the longer agentic run capabilities without routing data to a separate endpoint, which is often the detail that determines whether a proof of concept survives a security review. If your data already lives in Snowflake and procurement keeps you there, there is nothing to wait for.
Personal Computer for Mac Opens to All Max Subscribers and the Full Waitlist
Perplexity's Personal Computer agent is now available on Mac for all Max subscribers and everyone who signed up to the waitlist, moving it from a demo into something you can run against your own files today. It reads and writes across local folders, iMessage, Mail and Calendar. On a Mac mini it can run unattended around the clock against the apps you already use.
The differentiator most competitors cannot yet match is the iPhone handoff: start a task on mobile and pick it up on the desktop with 2FA as the bridge. For anyone who has been waiting for a capable cross-device agent outside the main model vendors, this is the first credible option currently shipping.
Perplexity Health Pulls Medical Records, Wearables and App Data into One Queryable Surface
Perplexity Health launched on the same day as the Mac rollout, pulling medical records from provider portals alongside data from health apps and wearables into a single queryable interface. For people managing multiple clinicians or tracking tools, the pitch is asking one question and getting an aggregated answer rather than logging in to four separate systems.
Privacy posture is the most important thing to check before connecting sensitive medical data to any AI service. Perplexity had not published detailed guidance on data residency or processing at the time of writing, so treat this as early-access territory and review their terms carefully before uploading anything sensitive.
Opus 4.7 and the Bedrock Agent Registry Go Live Together on the New Inference Engine
Two things land on Bedrock today. Opus 4.7 is available on AWS's next-generation inference engine, described as purpose-built for generative workloads and fine-tuning, removing the lag that has historically meant a new Anthropic model appears on first-party endpoints days before it reaches AWS-hosted deployments. Alongside this, Agent Registry is now usable inside Bedrock AgentCore as a private catalogue for agents, tools, skills and MCP servers, with semantic search, approval workflows and CloudTrail logging built in.
Bedrock has also added IAM-principal cost allocation, making model inference spend attributable to individual teams for the first time without custom tagging. If you have been building agents across multiple accounts with copy-pasted tool definitions and no clear cost ownership, the registry plus cost allocation turns that into something governable.
Marvell Joins the Nvidia AI Factory and AI-RAN Ecosystem
Marvell's networking and custom silicon capabilities have been brought inside Nvidia's AI factory and AI-RAN reference architecture. As an infrastructure partnership this will not change anything you can do today, but it sets the direction for how next-generation AI inference capacity gets assembled at the silicon-to-rack level. The most likely downstream effect is cheaper or faster inference on hyperscaler menus as the architecture matures through 2026 and 2027.
The AI-RAN side of the deal is the part worth tracking separately: it pulls telco workloads closer to Nvidia's fabric, which has implications for inference latency in edge and mobile contexts. Teams planning infrastructure commitments over the next 18 months should log this as a signal about where Nvidia's reference architecture is headed.
Industry Themes
The dominant story today is coordinated simultaneous availability. Anthropic shipped Opus 4.7 across its own API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry and Snowflake Cortex on the same day, with Snowflake publishing its own blog post at the same time as the Anthropic announcement. For teams that have historically waited weeks for a new model to reach their preferred hosting platform, this is a meaningful operational shift: there is no longer a first-mover advantage to running directly on Anthropic's own endpoints.
A second pattern running through the day is the push from chat to persistent agents on the user's own machine. Perplexity's Personal Computer for Mac, the Codex memory and sandboxing improvements, and Bedrock's Agent Registry all point in the same direction: the next phase of competition will hinge on which products let agents operate safely across real files, apps and identities rather than in isolated chat windows. The cross-device handoff that Perplexity is shipping (iPhone to Mac via 2FA) is the kind of UX detail that determines whether agents become a daily habit or stay as occasional tools.
OpenAI's GPT-Rosalind marks the clearest signal yet that frontier labs are deliberately fragmenting their model catalogues to serve regulated industries. Narrow scope, vetted partners, restricted access is a repeatable template that limits liability while unlocking revenue in sectors such as pharmaceuticals, finance and government that would otherwise face substantial adoption barriers. Expect most major model vendors to apply this pattern to at least one domain-specific release over the next 12 months.