AI Lately

Daily selection

Up to 2 per category
Checked through · UTC+8

Models

DeepSeek

DeepSeek V4.1 Flash accepts screenshots alongside text in its API

The updated Flash model accepts images and text through its API. Developers can include a page screenshot, an error description and code in one request.

In practice

A website developer could supply a screenshot of an obscured button, the relevant page code and reproduction steps, ask which code might affect the layout, then verify any changes in a browser.

How to provide images

Developers select deepseek-flash through the API and provide images as URLs, inline data or uploaded-file references. Other users need a tool connected to that capability. API support does not establish the same access in every DeepSeek web or app account.

Facts and assessment: compare the total cost per task

On September 19, official Flash rates per million uncached input tokens are USD 0.15 off-peak and USD 0.30 at peak; output costs USD 0.60 and USD 1.20 respectively. Our recommendation is to compare success, retries and cost on the same small task. Lower token prices need not mean a cheaper completed task.

Limits and verification

We checked the official release, vision guide and pricing, but have not tested Chinese screenshot recognition, repair quality or latency. Published benchmark results are vendor claims, not our review. This item verifies API capabilities, not identical access in every web or app account.

Original sources

  1. DeepSeek official changelog, September 10
  2. DeepSeek vision input guide
  3. DeepSeek current models and pricing

Checked: Sep 19, 2026

Products & apps

OpenAI

ChatGPT Work adds a Data agent for connected company databases

The Data plugin queries connected company data and builds interactive charts from plain-language requests. Administrators control plugin access, connections and permissions.

In practice

An e-commerce operator could compare two months of orders by product and channel, identify the largest drops and inspect the supporting data.

Who can use it

This suits teams with an existing data platform and the required ChatGPT Work access. The announcement lists sources including BigQuery and Snowflake.

What needs checking

We have not tested it. Define whether refunds count and whether dates mean order or payment dates. A decline in sales does not by itself establish its cause.

Original sources

  1. OpenAI · September 10 announcement

Checked: Sep 19, 2026

Open source

langgenius / Dify

Dify 1.17.1 fixes chat input remaining blocked after human approval

Dify combines knowledge bases, models and human-review steps into AI apps. The September 10 release includes a fix for chat input remaining blocked after an approval form is submitted.

In practice

A support team could draft answers from product documents and return policies, have an employee approve them, then continue the conversation. This fix matters for teams on affected versions whose chat input stays blocked after approval.

From a user report to a release

Issue #40076 includes reproduction steps. A maintainer corrected a bot that had pointed to older fixes, then linked the new PR, #42082. That PR is merged, listed in the 1.17.1 release notes and included in its tag. We checked the discussion, code and release rather than relying on issue closure.

Deployment requirements

Dify is an app platform requiring technical setup. This version lists at least two CPU cores and 4 GiB RAM for self-hosting, plus model and maintenance costs. Existing bundled-Weaviate deployments need the staged upgrade in the release notes; fresh installs and other vector stores are outside that requirement.

What the evidence establishes

GitHub showed about 156,000 stars at 01:25 China Standard Time on September 19. We have not deployed or reproduced the issue. The PR author reports regression tests and a blocked full type check; we found no post-upgrade confirmation from the original reporter. The evidence establishes a shipped fix, not a fault-free result in your environment.

Original sources

  1. User report #40076
  2. Maintainer corrects the bot response
  3. Maintainer confirms the merged fix
  4. Fix PR #42082 and validation notes
  5. 1.17.1 release notes and migration requirements
  6. Fix commit included in the release tag
  7. 1.17.1 README and deployment requirements

Checked: Sep 19, 2026

Hugging Face · Gradio

Gradio puts image generation, editing and upscaling on one canvas

The Workflow1111 demo, presented on September 10, combines image workflows in a browser app. It connects existing models into workflows that can be shared.

In practice

Creators can upload a reference image and describe an edit to receive a revised image, or generate a comparison grid from prompt variations.

What is available

The example can be opened or duplicated to change models and connections. Gradio Workflow connects models, other Spaces, datasets and Python functions.

What you need

The demo calls remote services, so it needs no GPU of your own. Sign in to Hugging Face or supply a token; calls use your quota. Custom functions still require Python knowledge.

Not tested here

We have not run the demo or verified output quality, cost or reliability. Browser access does not imply free use or that images stay on your device.

Original sources

  1. Hugging Face · September 10 announcement
  2. Gradio · Workflow guide
  3. Official demo · Workflow1111

Checked: Sep 19, 2026

Research & industry

No selection verified in this check