AI Morning Paper: 2026-05-17

Models

Claude Opus 4.7 is now generally available

Anthropic says Claude Opus 4.7 is generally available, and positions it as a meaningful step up from Opus 4.6 for advanced software engineering, especially on the hardest, most multi-step tasks. That matters because the model gap developers actually feel is often not “can it answer?” but “can it stay coherent through a long refactor, a tricky test failure, or a chain of tool calls without losing the thread?”

For an agency or Laravel team, this is the kind of release that can shave time off code review loops and reduce how often you have to rescue an agent from its own partial state. Source: Anthropic

Databricks is using GPT-5.5 for enterprise agent workflows

OpenAI says Databricks is deploying GPT-5.5 in enterprise agent workflows after the model set a new state of the art on OfficeQA Pro. The important detail is the target workload: office-style knowledge work where the model has to reason over messy business context, not just tidy benchmark prompts.

That is the real production lesson for builders. Once you move from demos to enterprise workflows, the differentiators become retrieval quality, permission boundaries, auditability, and how well the surrounding app handles failure. Source: OpenAI

Products

Claude Design pushes Claude into a lightweight prototype studio

Anthropic launched Claude Design, a new Anthropic Labs product for creating polished visual work such as designs, prototypes, slides, and one-pagers. In practice, that puts Claude closer to the brief-to-artefact loop that agencies already live in.

The useful bit is not “AI makes slides.” It is that a product like this can compress the first draft phase: rough idea, visual exploration, client-ready concept, and iteration all happen in one place. For teams that sell speed, that is more interesting than generic chat. Source: Anthropic

Claude for Small Business is built around connectors and ready-made workflows

Anthropic also launched Claude for Small Business, bundling connectors and ready-to-run workflows that put Claude inside everyday tools. That is a practical product move because most teams do not want another standalone chat window; they want AI to sit inside the systems where work already happens.

If you are thinking like an agency lead, this is the right shape: fewer “prompt the assistant” moments, more embedded workflows for drafting, reviewing, triaging, and moving work forward inside the stack people already use. Source: Anthropic

Research

ATLAS tries to merge agentic and latent visual reasoning

A recent arXiv paper, ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both, proposes a functional token that acts as both an external agentic operation and an internal latent reasoning unit. The point is to reduce the trade-off between code/tool-based reasoning and hidden embedding-based reasoning.

Why it matters: multimodal systems are still paying a lot of tax when they have to jump between external execution, context switching, and internal reasoning. If this line of work holds up, we may get models that do more structured visual reasoning with less orchestration overhead. Source: arXiv

Open Source

Granite Embedding Multilingual R2 is open, multilingual, and retrieval-focused

Hugging Face published IBM Granite Embedding Multilingual R2, an Apache 2.0 embedding model with 32K context and a claim of strong sub-100M retrieval quality. That is a very practical release: embeddings are unglamorous, but they decide whether your search, RAG, and support tooling feels crisp or vague.

For builders shipping internal assistants or knowledge-base search, the combination of multilingual coverage, permissive licensing, and long context is exactly the sort of infrastructure upgrade that quietly improves product quality. Source: Hugging Face

Continuous batching keeps getting more production-minded

Hugging Face’s post on unlocking asynchronicity in continuous batching is about serving efficiency rather than model hype. That is still useful news, because batch scheduling and request handling often decide the real cost profile of AI features in production.

If your app mixes short prompts, long prompts, retries, and streaming responses, these kinds of serving improvements translate directly into better latency and lower spend. That is the kind of optimisation agencies feel when client traffic becomes real traffic. Source: Hugging Face

Tools

Codex from the ChatGPT app makes agentic coding more supervise-able

OpenAI says you can now work with Codex from the ChatGPT mobile app, monitoring, steering, and approving coding tasks across devices and remote environments. That is a small UX shift with a big operational implication: agent work becomes something you can manage in bounded steps instead of babysitting in one long browser session.

For Laravel and agency workflows, that is exactly the right shape. You want an agent to run a scoped task, surface a decision point, wait for approval, and keep the artefact trail intact. Source: OpenAI

GitHub Copilot CLI is being used to turn codebases into procedural toys

GitHub’s Copilot CLI demo shows a Hubber using the tool to build a procedurally generated roguelike extension from a codebase. The game is not the point; the point is that CLI-first agent workflows are becoming expressive enough to reshape codebases, not just answer questions about them.

That matters because the command line is where automation is easiest to review, script, and compose. If your AI tooling can live there, it fits much more naturally into CI jobs, repo scripts, and the kind of repeatable workflows agencies actually depend on. Source: GitHub