This edition covers Jev, a decision model from TypeSafe and the first model on qbash that is not an LLM. It also covers three new models, Sites, changes to Task triggers and the Task Builder, three new videos, and the AI Ticker for September 18 to 24.
Jev makes decisions
Jev takes text and a set of typed questions and returns an answer to each, with a confidence score. A question can be Pick one, Score (2 to 10 levels), or Yes or no.
In a Task, the confidence score is what you branch on. Act on the high-confidence answers and send the uncertain ones to a person or a larger model. TypeSafe trained Jev for calibrated confidence, and in its workflow evals Jev ran up to 193.6x faster and 444.6x cheaper than LLMs. TypeSafe says those figures are likely at the high end of real-world gains.
Selecting Jev on a prompt replaces the output schema with a Questions editor. Add questions or paste them as JSON, then save the prompt or use it inline in a Task. From Python, qbash.ai.raw(text, model="jev-latest", output_schema=questions) returns a dict keyed by question name. A prompt started in chat shows each answer with its confidence, and the raw JSON is one toggle away.
An admin adds your TypeSafe key on Model Providers. Jev costs $0.042 per million input tokens, with no output charge. It accepts text and has a 64K context window.
Three new models
- Claude Opus 5.5 (
claude-opus-5-5) costs $4 per million input tokens and $20 per million output, down from $5 and $25 for Opus 5. Both have 1M context. Adaptive thinking stays on in 5.5. - GPT-6 Sol (
gpt-6-sol) costs $2 and $10 per million tokens, with 1.05M context. - GPT-6 Luna (
gpt-6-luna) costs $0.10 and $0.50, with 1.05M context. It replaces GPT-5.6 Luna on trial keys.
Both GPT-6 models accept images, structured output, and web search, and take a reasoning effort from none to max. Long-context rates apply above 272K input tokens. All three are on by default when you use your own provider keys.
Run them against your current model in Playgrounds before you switch a Task over. Thinking that stays on can use up what a lower token price saves. Model Providers covers enabling models and adding keys.
Introducing Sites
A Site puts Task output and Memory content behind one link. A generated page shows a Task's latest result along with its earlier runs. Other page types list a Memory folder as it changes, or show the current version of a single document.
A weekly report Task published to a Site gives your team one place to read each week's report. Sites start private. You can share one with your team, with specific people, or with a project. You can also turn on a public link, which works without a qbash account. Viewers of a Site get no access to the Tasks or folders behind it.
Sites replace project pages. Every project that had pages now has a private Site with the project's name, page order, and link. Sites are rolling out by workspace, so yours may not show in the sidebar yet. The Sites overview has the details.
Task updates
- More than one trigger per Task. A Task can now have a schedule, a webhook, and an API trigger at once. It can also have several schedules, each with its own cron expression and timezone. Revisions snapshot the triggers, and the run feed shows which trigger fired. See Triggers and entry points.
- Agent Tasks can call Python Tasks as tools. Link up to 25 under Task resources, and the agent decides when to call them. Agents cannot call other agents, and a Task called this way cannot call further Tasks.
- Task Builder runs on Opus 5.5. It narrates long tool loops as it goes. State snapshots are now system messages that clear when you reply, so old ones stop piling up in the conversation.
Three new videos
Building an integration test suite with one prompt in qbash. Mitchell Tessier attaches the upcoming LinkedIn integration to a code Task and gets back a script that cleans up after itself: 53 tests across seven phases, with 46 passes, five failures, and two skips.
Bulk fill PDFs with qbash Tasks. Alex MacLeod's Task reads a fillable PDF and a 20-row Google Sheet, fills a flattened copy for each customer, and zips them into one download.
Reddit research with qbash Tasks. The AI Ticker below starts here. A daily code Task stores summaries of threads from the model subreddits in Memory, and a second Task chains three prompts over a week of them to draft the ticker.
The AI Ticker, September 18 to 24
Both labs shipped models on the same day, and the five stories below cover what agents cost and who answers for them.
Some users are having trouble moving agents to GPT-6 Luna. They describe loops and ignored instructions in agents that worked on GPT-5.6. GPT-6 shipped without a Terra tier.
Opus 5.5 drew strong day-one reviews in Claude Code, with uneven usage. Two runs of the same prompt, an animated and scored train journey, used 4% and 14% of a weekly limit.
Claude Code's prompt suggestions reread the full context after each turn. One user's traces put the cost at up to 91% of a normal prompt. Setting promptSuggestionEnabled to false turns them off.
Stripe Checkout exposes WebMCP tools to browser agents. Agents pay by calling tools the page registers, in place of clicking through checkout. Stripe reports 42% fewer tokens.
A thread with 1,300 upvotes describes engineers approving AI output they cannot review. One reply argues that part of what banks pay consultants for is someone to hold liable when it fails.
If you move a classification step from a text model to Jev, tell us how the answers compared.