AI Daily · 2026-10-03

Today's updates both focus on model selection and inference cost control. OpenAI published a guide for the GPT-6 family, mapping Astra, Sol, and Luna to hard reasoning, complex coding/research/computer use, and high-volume routine tasks, while pointing to prompt caching and compaction as ways to reduce cached input token costs by up to 95%. LangChain described a model router built inside an agent harness rather than a generic gateway; using task type and complexity signals from harness traces, it cut median per-task coding costs by 64% with no measurable quality loss. The shared theme is that not every task needs the most expensive frontier model, and context-aware routing or caching can better balance success rate, latency, and cost per successful task.

North America · First-hand

OpenAI

⭐⭐ [Product Update] A model guide for the GPT-6 family

OpenAI published a practical guide for using its GPT-6 family of models. GPT-6 is described as its most advanced suite, with GPT-6 Astra for the hardest reasoning work, GPT-6.1 Sol for complex coding, research and computer use, and GPT-6 Luna for focused, high-volume everyday tasks such as extracting invoice fields, classifying requests or producing structured summaries. The guide recommends using prompt caching and compaction to manage context and cost in production, noting that cached input tokens cost up to 95% less depending on the model, and measuring task success, latency and cost per successful task. In the API, reasoning effort can be set to low, medium, high or extra high/Max and changed mid-conversation without breaking the cache, while long-running work can be managed with steering, async tools and delegation.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Research

⭐⭐ [Research] How to Build a Model Router in the Harness

LangChain describes how it built a model router for a coding agent. Compared with the baseline of always using a top-tier frontier model, the router cut median cost per thread by 64% with no measurable change in quality. The authors argue the routing decision belongs in the agent harness rather than a generic gateway, because picking the right model requires the domain and task context the harness already assembles and a gateway typically lacks. Using Open SWE, their open source coding agent, as the test bed, they pulled thread-level data from LangSmith traces and used an LLM classifier to label a week of interactive threads by task type: code changes dominated, with new features at 22%, bug fixes at 17%, and test or no-op runs at 16%. They also found feature-work investigations tended to be longer with higher cost and median turn count, while testing and release threads were relatively short and inexpensive, using total cost and invocation count as complexity signals.