Claude Code Review goes GA, GPT-Live voices from a prompt, Elorian Glimpse, and Perplexity Decider v1.2

When a pull request lands on GitHub, several agents read the change alongside the surrounding code and look for logic errors, security holes, and missed edge cases, then leave inline comments. Findings go through a verification step first and arrive ranked by severity, with a summary and a check run on the PR. With GA, an Owner can set the review level for the whole org and per repository: Quick, Standard (the default), or Deep.

Agents now write more code than people can comfortably read. This gives teams an official first-pass reviewer. You can keep routine repos on Quick and put payments or auth code on Deep to trade cost for thoroughness. Reviews never approve or block a PR, so the merge call stays with a human.

Write a description like "a calm, warm support-agent voice" and the API creates and saves that voice. No reference recording or consent recording is needed. You create it through the API or the voice Playground and reuse it in sessions with GPT-Live, OpenAI's real-time voice model. Creating a voice is free, and speech is billed at normal GPT-Live session rates. The same day, GPT-Live gained nine voices, including Noeul for Korean.

curl https://api.openai.com/v1/audio/voices \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"type": "prompt", "name": "Support assistant", "prompt": "A calm, warm adult voice for a customer support assistant.", "model": "auto"}'

If you're building a support bot or voice assistant, you can try out an on-brand voice in minutes without booking a recording session. The new Korean voice also makes GPT-Live easier to test for Korean-language products. The catch is that prompt-generated voices only work in GPT-Live. The Text to Speech, Realtime, and Chat Completions APIs can't use them.

Glimpse is the first model from Elorian, a startup founded by former Google DeepMind researchers. Most multimodal models look at an image once, turn it into tokens, and reason in text. Glimpse instead uses a 'visual scratch space' mid-reasoning: it zooms into a region, rewatches a stretch of video at a denser frame rate, or runs code to measure something. It was post-trained on an open-weight base model and is available through an API and a chat interface.

It's worth a look anywhere you have lots of visual data: robot footage, lab charts, media archives. Elorian says Glimpse beats GPT-6 Astra on four evaluations covering scientific chart reasoning, counting, short video captioning, and manipulation eventsVendor claim. Pricing is $0.60 per million input tokens and $3 per million output, 94% below GPT-6 AstraVendor claim.

These are 'decision models': give them a question and a fixed set of options, and they return one option plus a probability instead of free text. They accept text and images up to 1.6 MP and come in two sizes, the 27B pplx-decider-v1.2 on a Qwen3.8-27B backbone and the 4B pplx-decider-flash-v1.2. New in this version, every question in a request runs in a single pass, so asking several questions about the same input costs about as much as asking one.

Classification, routing, and reranking all have fixed answers, and a model like this runs those jobs at volume for far less than a chat model. Perplexity says the overall Decision Index score rose from 62.30 to 68.09Vendor claim and inference is up to 5x fasterVendor claim. The Apache-2.0 license means you can run it commercially on your own hardware.