Claude Opus 4.7 takes on longer coding tasks as Codex gains Mac controls

Anthropic released Claude Opus 4.7 on April 16, 2026. The model is generally available in Claude products, through the Anthropic API, and from several cloud providers.¹ Anthropic says it improves coding, long-running agent tasks, instruction following, and verification of its own work.¹ The company calls it its most capable Opus model to date.¹

For companies, performance is only part of the decision: a new model also needs to fit existing development workflows. The reported improvements come from Anthropic; teams should test how reliably Opus 4.7 works in their own codebases.¹ The model also includes safeguards against prohibited or particularly high-risk cybersecurity requests.² Organizations using AI for security work should check how those limits affect their workflows before deploying it widely.

OpenAI expanded Codex on April 16 for desktop app users signed in with ChatGPT.¹ The coding agent can now operate a computer in the background, clicking and typing while the user continues working.¹ An in-app browser and 90+ plugins and integrations connect it to more applications and tools.¹ OpenAI also points to testing and front-end iteration as tasks beyond writing code.¹

On April 21, OpenAI introduced Codex Labs and partnerships with systems integrators to support enterprise adoption.² Partners include Accenture, Capgemini, and Infosys.² The update broadens Codex’s role from producing code to carrying out tasks across applications.¹ Availability remains a consideration for companies: not every feature is launching at the same time in every region or for every account type.³

Alibaba introduced Qwen3.6-35B-A3B on April 15, 2026, as the first Qwen3.6 model with open weights.¹ It uses a sparse mixture-of-experts architecture that activates only a portion of its parameters for each token.¹ Alibaba also offers the model through Qwen Studio and an API.¹ The company positions it for agentic coding and multimodal reasoning.¹

The architecture is intended to reduce compute per token compared with a similarly sized model that activates all its parameters.¹ That does not establish a fixed savings figure for deployment: memory requirements and infrastructure also matter. Companies can run the weights themselves or try the model through the API.¹² The less expensive option will depend on their workloads and infrastructure.

Anthropic introduced Claude Design on April 17, 2026, as an experimental preview from Anthropic Labs.¹² The tool creates visuals through conversation, including presentation slides, prototypes, and marketing materials.²³ Users can refine results with comments or direct edits, then export them to Canva, as PPTX or PDF, or to Claude Code.²³ Access is rolling out gradually to paid subscribers.¹²

Claude Design is an application powered by Claude Opus 4.7, not a separate new foundation model.¹² It extends Anthropic’s offering into visual creation and puts the company in competition with established design tools.²⁴ For businesses, a practical question is whether exported drafts fit existing editing and review workflows. Because the product remains an experimental preview, teams should check layouts and content before publishing them.¹²

Cloudflare introduced Agent Memory on April 17 as a private beta. The managed service is designed to extract information from AI agent conversations and retrieve it for later requests without loading the full conversation history into a model’s context window. Agents built on Cloudflare Workers can access it through a binding; agents elsewhere can use a REST API. Cloudflare did not announce general availability.¹

Cloudflare positions Agent Memory as part of its infrastructure for AI agents. The service retrieves stored information when needed rather than sending an ever-growing conversation history to a model. That approach could simplify context management for long-running agents, but the announcement does not establish how well it performs in practice. Cloudflare has not provided performance figures, pricing, or usage limits.¹

Perplexity introduced Personal Computer for its Mac app on April 16, 2026. It extends Perplexity Computer and can carry out tasks across local files, browsers, and native apps.¹ It can search, read, and edit files and work with apps including Messages, Mail, and Calendar.² Access began for Perplexity Max subscribers; the company’s changelog says all Max subscribers have access.¹³

The release adds actions on a user’s Mac to Perplexity’s earlier cloud-based system.¹ That makes it relevant to tasks requiring information from several apps, rather than answers from a chat interface alone. Access remains tied to Max, with Pro access planned for later.³ Before using it for business workflows, companies should decide which local files and apps an agent may access and which actions need review before completion.

Tencent introduced HY-World 2.0 as an open-source project in mid-April.¹ The multimodal system is designed to generate, reconstruct, and simulate interactive 3D worlds from text, images, and video.¹² The release is staged: code and model weights are available for part of the system, rather than every announced component.³⁴

Unlike video-first approaches, HY-World 2.0 produces spatial 3D representations intended for use in game engines and embodied simulation workflows.¹² That distinction matters to game developers and robotics teams that need 3D assets for downstream tools, not just rendered footage.¹⁵ Tencent describes interactive simulation, but the available materials do not establish measured physical accuracy against a dedicated physics benchmark.¹²

Physical Intelligence introduced its π0.7 robotics model in a research post on April 16, 2026.¹ The model is intended to perform dexterous tasks and follow new language instructions, including for tasks absent from its training data.¹ In a demonstration, step-by-step coaching helped a robot complete an unfamiliar task involving a kitchen appliance.¹² Without that coaching, it made only a passable attempt, according to TechCrunch.²

The model builds on the company’s earlier π0 research.¹ The demonstration suggests that language could guide robots through new procedures without collecting additional teleoperation data for every task.¹³ Physical Intelligence describes the results as early signs of generalization, however.¹ Robust benchmark results, commercial availability details, and evidence of reliable performance beyond the demonstrated tasks are not available.¹²

The White House was preparing in mid-April to make a version of Anthropic’s Mythos available to federal agencies. The Office of Management and Budget was working on safeguards to let agencies use the model. But an email from federal CIO Gregory Barbaccia gave neither a firm commitment nor a timeline. It remained unclear whether the planned access had gone live.¹

The plan comes amid a dispute between Anthropic and the Defense Department. The Pentagon had designated the company a supply-chain risk, barring use by the department and its contractors.²³ Some federal agencies were nevertheless already testing the model, according to reports.² Anthropic is providing Mythos through a restricted cybersecurity-focused deployment and had not planned a public release.⁴ Agency testing should therefore not be mistaken for an authorized, routine rollout.¹²

Anthropic has published guidance on managing long sessions in Claude Code.¹ An April 20 post covers when to use subagents, while a separate guide explains session management.¹² Anthropic recommends subagents for tasks whose intermediate output is not needed in the main conversation. Each works in a fresh context window and returns a synthesized result to the parent session.¹

The guidance addresses “context rot,” Anthropic’s term for performance degradation as a session’s context grows.¹ Delegating a discrete task can keep unnecessary output out of the main context. Anthropic also describes compacting or clearing context to manage longer tasks.¹ These are usage recommendations, not evidence of a new model release or measured performance gains from the approach.¹²