Tuesday's theme is where the model runs. OpenAI cut the price of its mid-tier, Cloudflare published a hard lesson about shared inference, and two launches push the model onto the user's own hardware. If you self-host SGLang, read story two first.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
GPT-6.1 Terra costs $1 in, $5 out and ties GPT-6 Astra on Agents' Last Exam at medium effort
OpenAI shipped GPT-6.1 Terra on Monday as gpt-6.1-terra: 1.05M context, 128K output, text and image in. List price is $1 input and $5 output per million tokens, half of GPT-6.1 Sol. On Agents' Last Exam v1 at medium effort it scores 57.6% at $0.29 per task, the same score GPT-6 Astra gets at $4.82. Coding is the weak spot: 50.3% on Terminal-Bench 4.0 against 70.6% for Claude Sonnet 5.5. All numbers are OpenAI's own, per MarkTechPost.
Link: openai.com/index/introducing-gpt-6-1-terra
Try it now
curl https://api.openai.com/v1/responses -H "Authorization: Bearer $OPENAI_API_KEY" -H "Content-Type: application/json" -d '{"model":"gpt-6.1-terra","reasoning":{"effort":"low"},"input":"Extract every field from this invoice text as JSON: ..."}'. Tool calling needs the Responses API. Batch and Flex are 50% off, and prompts over 272K input tokens bill at 2x input, so keep long-context jobs under that line if cost matters.
Story II
Cloudflare's GLM-5.3 postmortem: 155 responses carried other customers' data through an SGLang bug, fixed in v0.5.22
Between August 27 and September 9, prompts longer than 681,280 tokens sent to GLM-5.3 on Workers AI could return text from other users' requests. The KV pool on Cloudflare's prefill GPUs was smaller than the 1,048,576-token context window, and in disaggregated prefill/decode mode SGLang turned over-length prompts into a one-token placeholder instead of failing. Decode then attended over stale pages and answered with HTTP 200. Cloudflare found no evidence of deliberate exploitation and recommends every SGLang operator move to v0.5.22 or later.
Link: blog.cloudflare.com/kv-cache-exposure-remediation
Try it now
Run pip show sglang on every inference host. If it reports anything below 0.5.22, upgrade with pip install -U "sglang>=0.5.22". Then compare the KV pool size SGLang prints at startup against the context length you advertise. Cloudflare's lesson: if the pool is smaller, your long-context tests can return 200 while the output is wrong, so check the content of the reply, not only the status code.
Story III
Hob is an MIT desktop agent that reads and edits your files with a local open-weight model, no account required
Frigade, a Y Combinator company, released Hob for Mac, Windows and Linux. It picks an open-weight model that fits your hardware (Qwen, Gemma or GPT-OSS), then works on folders you grant: reading and editing files, filling PDFs, drawing charts and running code in a sandbox. It asks before changing a file and lets you undo edits. Everything runs locally and works offline. The repo is MIT with 246 stars two weeks after it was created. Minimum is 8 GB of RAM; the site says 16 GB is good for everyday tasks.
Link: github.com/FrigadeHQ/hob
Try it now
Download from hob.chat, or clone it with git clone https://github.com/FrigadeHQ/hob and read CONTRIBUTING.md for the pnpm workspace build. Grant it one throwaway folder with a spreadsheet and a few notes, then ask: "Read every file here and write a one-page report with the totals by month and a chart." That shows you what your machine's model can handle before you trust it with real work.
Story IV
Opera ships WebAI in Opera One, so sites and extensions can call Gemma or Apple's on-device model with no API bill
Opera put an early preview of WebAI, a proposed web standard that Opera says Google and Microsoft back, into Opera One, with Opera GX and Opera Neon to follow. Websites and extensions can call a model running on the user's machine. At launch that means Google's Gemma, plus Apple's on-device foundation model on Macs. It is opt-in, and users control which sites get access. Opera also published example extensions that sort tabs and flag clickbait locally. Opera's pitch to developers: AI features without running inference or paying per request.
Link: blogs.opera.com — WebAI
Try it now
- Update Opera One (Windows x64 or Apple Silicon Mac for now).
- Open Settings, then Opera AI, and turn on Enable Web AI API. It downloads a local model, which can take a few minutes.
- Install one of Opera's example extensions and read its source to see the calls, then decide whether a local fallback makes sense for a feature you currently pay per token for.
— The Vibe Gate news desk. We read the firehose so you can keep building.