All Releases

Also shipped: models, platform, and reliability updates

A roundup of the smaller improvements shipped across CRHQ in early September 2026: a new low-cost model, a cheaper GPT-5.6 Sol, and a more capable, more reliable Platform API.

changelog

Alongside GPT-6 Astra and Status Boards, a lot of smaller work shipped across CRHQ in early September. None of it needs any action from you. It is all live on every instance. Here is the plain-language version of what changed.

A changelog-style summary of recent CRHQ platform and model updates.

New and cheaper models

  • A new fast, low-cost model: GLM-5.3-Flash. Z.ai's newest lightweight model is now selectable in the model picker, a good option for quick, high-volume work where speed and cost matter more than raw power.
  • GPT-5.6 Sol got cheaper. Alongside the GPT-6 Astra launch, OpenAI cut Sol's price to $4 in and $20 out per million tokens. The everyday Codex workhorse now costs less, and CRHQ's cost tracking reflects the new rate automatically.

A more capable, more reliable Platform API

The Platform API that lets you build your own product on CRHQ kept maturing, moving through several versioned updates with no breaking changes:

  • Per-run cost reporting. You can now pull token usage and cost for each individual run, not just per session or per day, which makes client-facing billing more precise.
  • Steadier live event streams. Session lifecycle events, message delivery, and queue behaviour were all hardened, so apps that watch agents in real time get a more reliable feed.
  • Automatic recovery of stalled hand-offs. When one agent delegates to another, the system now detects and recovers a hand-off that stalls, instead of leaving it hanging.
  • Runtime model discovery. Apps can ask an instance which models are currently available, so a custom interface can show the live model list without hard-coding it.

Agents handle a busy inbox better

When several agents work together, messages can arrive in bursts. We improved how an agent drains its queue: it now handles a batch in sensible order, respects genuine priority, and avoids redundant back-and-forth. In short, busy multi-agent setups stay more coherent under load.

Polish

We also refined where the inline artifact cards, introduced last week, appear in the conversation, so they land at the right point in a long session instead of piling up.

None of the above requires any changes on your side. It is already running on your instance.