Updates for 5/2

Perplexity Computer autonomous agent (19-model orchestration, Max tier)

Perplexity was mainly a search-and-answer tool with sources. Now it also offers a “Computer” assistant that can carry out tasks for you. This capability is positioned as a premium feature on the Max plan. The focus shifts from finding info to taking actions based on it.

Teams can shorten the gap between research and execution (for example: gather info, then apply it). It raises the bar for vendor comparisons: not just answer quality, but task completion. Budgeting may change because the most useful workflow features sit behind a higher tier. It can reduce manual effort for repetitive “do the steps for me” work.

Perplexity Computer: autonomous agent orchestrating 19 models with 400+ app integrations (Max tier)

Previously, Perplexity’s strengths were centered on fast, cited answers. Now “Computer” can coordinate many AI engines and connect to hundreds of apps. That means it can move information between tools instead of leaving you to copy/paste. This is presented as a Max-tier capability.

App connectivity is often the difference between a demo and daily use. If it can reliably act across your tools, it may replace smaller single-purpose automations. It also increases vendor lock-in risk, since workflows may become tied to one platform. Decision-makers should test real tasks end-to-end, not just prompt quality.

Claude Cowork (PC-automation AI agent) now generally available for paid plan users; Zoom connector ingests meeting data into Cowork

Claude already offered strong assistance for complex work and long documents. Cowork is now available for paid users, making PC task automation easier to access. A new Zoom connection can pull meeting information into Cowork. This strengthens the “from meeting to action” workflow inside Claude.

Meeting notes can turn into tasks faster: follow-ups, summaries, and action lists. It reduces the friction of getting context into the assistant in a safe, repeatable way. For businesses, it supports more consistent execution after meetings. It also signals a push toward integrated work assistants rather than standalone chat.

Le Chat gains Agent Mode for autonomous multi-step task execution

Le Chat was primarily a conversational assistant. Agent Mode adds the ability to run multi-step tasks more autonomously. This is a shift from “tell me” to “do this sequence of steps.” It aligns Le Chat with a broader industry move toward action-taking assistants.

Organizations evaluating Mistral now need to consider workflow automation, not only model cost. Multi-step task support can improve productivity in support, operations, and analysis. It increases the need for clear controls: what the assistant is allowed to do. Pilots should focus on measurable time savings and error rates.

Grok Computer: autonomous desktop automation agent with parallel planning and execution

Grok was best known for fast answers and strong integration with X. “Grok Computer” adds the ability to operate a desktop and complete tasks. It can plan and run steps in parallel, rather than one at a time. This shifts Grok toward being an execution tool, not just a chat tool.

Desktop automation can cut time spent on routine internal processes. Parallel task handling may improve turnaround for multi-part requests. It also raises governance questions: who can run actions, and how they’re reviewed. Buyers should evaluate reliability and audit trails before using it for critical work.

Grok 4.3 "Imagine" agent mode: autonomous agent for creative and image-generation tasks

Previously, creative work in Grok was more prompt-driven and manual. “Imagine” adds a more autonomous mode focused on creative and image tasks. It suggests a workflow where the assistant iterates and produces outputs with less back-and-forth. This is introduced alongside broader “agent” capabilities in Grok.

Marketing and design teams may get faster drafts and more variations per brief. It can reduce time from concept to first usable asset. Businesses should still validate brand consistency and rights/usage policies. It widens the gap between basic chat tools and production-oriented creative tools.

AI Legal Agent integrated into Microsoft Word for contract review, clause checking, and redline suggestions

Copilot already supported work across Office apps. Now Word includes a Legal Agent for reviewing contracts. It can flag clauses, check for issues, and suggest redlines. This brings specialized help into the document where legal work happens.

Contract cycles can speed up by reducing first-pass review time. It can help non-lawyers catch common issues before sending documents to legal. Organizations should define who can rely on suggestions and what requires human sign-off. It may change software spend by reducing the need for separate contract review tools.

Group chat for up to 200 users collaborating across 200+ AI models

Poe was already a hub for trying many AI models. Now it supports group chats for up to 200 participants. Teams can collaborate in one thread while switching between models. This moves Poe closer to a shared workspace, not just a personal tool.

Cross-functional teams can evaluate models together with shared context. It can speed up consensus on outputs for content, analysis, or support responses. Centralized collaboration helps with consistency compared to scattered individual chats. Leaders should plan access controls and guidelines for what gets shared in-room.

Uber: ~95% of engineers use Claude Code monthly, ~70% of commits AI-generated (as of April 2026)

Claude Code has been positioned as a terminal-based coding assistant. This update adds a real-world adoption data point from Uber. It reports most engineers use it monthly and many code changes are AI-assisted. That’s a shift from “possible” to “proven at scale” in a large org.

Strong adoption suggests developer tools can deliver measurable productivity gains. It also implies process changes: code review, testing, and standards need to keep pace. Buyers can use this as a benchmark when setting rollout goals. It strengthens the case for investing in training and guardrails, not just licenses.

Earlier descriptions often framed coding assistants as autocomplete tools. The updated content highlights repo-wide understanding and multi-file changes. It also stresses running tests, reviewing diffs, and fixing issues in a workflow. That reframes it as a practical assistant for shipping changes, not typing faster.

Teams can plan for broader use cases like refactors, debugging, and adding tests. It shifts evaluation criteria toward reliability with existing codebases and build systems. This can influence onboarding: engineers need habits that pair AI output with verification. It supports business cases tied to delivery speed, not just developer satisfaction.

Claude Security launched: enterprise AI security tooling built on Mythos model capabilities, helping defenders gain AI advantage

Claude previously focused on general assistant capabilities and developer features. Now Anthropic has introduced a dedicated security offering. It is positioned as tooling to help defenders work faster and smarter. This adds a clearer “security buyer” story to the Claude lineup.

Security teams may gain faster investigation and response workflows. A dedicated offering can simplify procurement compared to general-purpose tools. It may reduce risk by focusing features on defensive use cases. Organizations should assess how it fits with existing security processes and reporting.

Selected for U.S. DoD classified AI program alongside Google, Nvidia, and xAI

OpenAI already had broad model and API availability. This update adds a major government selection signal. It places OpenAI alongside other large vendors in a sensitive program. That can be read as increased confidence in operational readiness for high-stakes use.

For enterprises, it may de-risk vendor selection when sensitive workloads are involved. It can influence long-term roadmaps if the vendor invests more in security and reliability. It may also shift competitive dynamics: customers may expect similar controls from others. Procurement teams can treat it as one data point—not a substitute for due diligence.

Excluded from U.S. DoD classified AI program (cited as supply-chain risk; OpenAI, Google, Nvidia, xAI were selected)

Claude is widely used and known for strong reasoning and coding support. This update notes it was not selected for a U.S. classified AI program. The cited reason was supply-chain risk. That adds a public signal about perceived readiness for certain high-security contexts.

Organizations in regulated or sensitive sectors may ask tougher questions about sourcing. It could affect competitive shortlists where government-aligned requirements matter. It highlights that performance alone isn’t the deciding factor; operational assurances matter. Buyers should map requirements to vendor capabilities, not assume parity across providers.

Marketing cookies enabled by default for free users in ad-serving countries (paid plans exempt; opt-out available in settings)

ChatGPT’s free tier has been moving toward ads in some countries. This update notes marketing tracking is enabled by default for free users there. Paid plans are exempt, and users can change the setting. It’s a meaningful shift in default data handling for the free experience.

Companies should be cautious about employees using free accounts for work topics. Default settings matter because most users never change them. This may push teams toward paid plans or approved internal tools for sensitive work. It also affects customer-facing recommendations: “free” may come with trade-offs.

Gemini 3.1 Flash TTS Preview: expressive speech synthesis with dynamic emotion/tempo/style control via Audio tags, low-latency streaming

Gemini already supported a wide range of text-based tasks. This update adds a preview of more expressive speech output. It can adjust emotion, pace, and style, and it can stream with low delay. That moves it closer to real-time voice experiences that feel natural.

Better voice quality can improve call assistants, training, and accessibility tools. Low delay matters for live conversations, not just reading text aloud. Businesses can prototype voice-first experiences with clearer control over tone. It may reduce the need for separate voice vendors for some use cases.

DeepMind AI Co-Clinician outperforms GPT-5.4 in blinded physician simulation test for medical diagnosis

Medical AI claims are often hard to compare across vendors. This update adds a blinded test result where DeepMind’s system performed strongly. It is framed as outperforming a leading alternative in a physician simulation. That’s a notable milestone for clinical decision support research.

Healthcare leaders get a clearer signal of progress toward practical assistance. It may accelerate pilots in triage support, documentation, or second-opinion workflows. However, real-world deployment still requires validation, oversight, and accountability. Non-healthcare businesses should note the broader trend: benchmarks are becoming more rigorous.

SuperGrok Lite: $10/mo (from 2026/3/25)

Grok’s paid options ranged from bundled access to higher-priced tiers. This update adds a lower-cost SuperGrok Lite plan. It sits between free access and more expensive subscriptions. That changes how individuals and teams can start using Grok.

Lower entry pricing can expand adoption for pilots and small teams. It gives procurement more flexibility to match cost to usage. It may also signal a push to compete on accessibility, not just premium features. Decision-makers should compare what is included at each tier before standardizing.

Mistral Medium 3.5 GGUF initial bug (YaRN parsing) fixed by Unsloth — long-context performance restored

Some early users saw degraded performance on long documents with a specific format. This update notes that an initial bug was fixed by a third-party contributor. The fix restores performance for longer inputs. It improves confidence that the model behaves as expected in extended tasks.

Long documents are common in business: policies, contracts, research, and logs. Restored performance reduces the risk of failed evaluations or inconsistent results. It can make self-hosted or locally run deployments more dependable. Reliability fixes often matter more than new features when rolling into production.