This edition covers GPT-6 Astra in the model catalog, including the long-context surcharge and cache-write rate that sit outside the headline price, four new videos on the channel, and five stories from a rough week for the labs.
GPT-6 Astra is available
OpenAI's new flagship is in the catalog as gpt-6-astra, with a 1.05M token context window, 128K output, structured output, and web search. It is priced at $10 per million input tokens and $50 per million output. Prompts above 272K input tokens move the whole request to long-context rates, 2x on input and 1.5x on output. Cache writes bill separately at $12.50 per million against a $1 cached-input rate.
Astra is opt-in and trial keys do not cover it, so an admin has to enable it in Org Settings against your organization's own OpenAI key. After that it appears for everyone in the org. Run it against your current default in Playgrounds before you move a Task across, because Astra is not uniformly ahead of what it replaces. The Model Providers doc covers enabling a model and adding your own key.
Four new videos
Build a custom AI model test suite. Two frontier models shipped in two days. A Task runs 15 business tests across eight models every Monday, scores them, and diffs the results against last week.
How qbash Tasks work: a simple model test. A single-model version of the test suite above, cut down to 24 lines so every step of a Task is visible.
Creating a Fed sentiment dataset for bond analysis. Long-bond yields are at 19-year highs. A Task tracks what 20 Fed officials say, scores each remark as hawkish, dovish, or neutral, and writes the rows to Airtable.
Asana AI automation that reads Slack and GitHub for you. Triaging our busiest Slack channel into Asana tickets took an hour a week. The Task that replaced it reads Slack, Asana, and GitHub together, and remembers enough that work finished last week arrives marked done.
The AI Ticker, September 8 to 10
Five stories that change something about model choice, cost, or how you check the output.
Astra tops the Intelligence Index and still regresses on finished work. Artificial Analysis has it tied first at 53 with hallucinations down from 92% to 51%, but off about 45 Elo on GDPval-AA v2 and behind Sol on presentation. Artificial Analysis ties the drop to Astra doing less work per task, 24 turns at max effort where Sol took 45.
OpenAI's Navier-Stokes proof arrived in the middle of a credit fight. Tristan Buckmaster says OpenAI heard of his unpublished Euler result with Anthropic's Levent Alpöge, pushed it to full Navier-Stokes over a weekend, then proposed a joint paper without Alpöge on it. OpenAI denies using their work.
Claude Code's weekly limits fall 17% on September 14. A temporary 50% boost ends the day before, and the permanent 25% increase replacing it sits below today's ceiling on Pro, Max, Team, and seat-based Enterprise. Separately, an open issue tracks prepaid credits draining after sessions close.
Anthropic now calls "double-check your work" an anti-pattern. Frontier models read verification rituals literally and spend tokens repeating lookups they would have done once. In Anthropic's own support benchmark, contradictory refund rules left Opus 5 withholding four refunds it owed while it asked the customer to confirm.
Databricks' Smart Routing picks the model per session from how hard the task looks. A small classifier reads the request, sends routine work down and hard problems up, and logs every decision. Databricks reports matching Opus 5 quality at under half the cost. Smart Routing is still in beta.
If you run Astra against your current default, tell us what came back.