Top story
GPT-5.6 Sol got cheaper; do not confuse a promotion with a baseline https://developers.openai.com/api/docs/pricing
OpenAI’s API pricing page now lists GPT-5.6 Sol at $4 input, $0.40 cached input, $5 cache writes, and $20 output per million tokens for short context. Batch and Flex cut those numbers to $2 and $10. The consequential footnote is temporal: OpenAI says the promotional pricing will remain available at least through November 21, 2026. On HN, ComputerGuru calculates a 20% input and 33% output reduction from the prior rate; wahnfrieden correctly marks the boundary—this does not lower ChatGPT subscriptions.
I would not reroute a workload because the sticker got friendlier. Sol still costs 20 times Luna on standard short-context input. Re-run the cost trace with cached input, context length, and output separated, then measure the quality delta. If budget alone previously eliminated Sol, this promotion earns a fresh eval this week. If Luna already clears your acceptance tests, a larger model name is not a reason to pay more. A lower rate is a reason to measure again, not evidence that one model wins.
Two tabs I kept
Keep the friction that turns AI-assisted coding into expertise https://larsfaye.com/articles/ai-coding-will-prevent-expertise
The useful part is not the prediction that expertise will collapse. It is the cited novice study: heavy GenAI users skipped planning and left with an “illusion of competence.” HN commenter causal supplies the needed counterweight—LLMs widened the tools they could reach, even while syntax faded. My rule: delegate mechanics, but make learners design, explain, and test the solution.
Treat model output as hostile input to the inference engine https://boydkane.com/essays/llms-could-control-their-host-machines-by-exploiting-inference-engines
This is a threat model, not a new CVE. It starts from CVE-2025-9141, where a vLLM tool parser fed arguments to eval(), and asks whether malicious tokens could compromise the inference host. Two recommendations are useful now: minimize GPU-host privileges and separate token parsing from the GPU machine. Hardening should not depend on proving that a model has intent.
One connection
Cheaper inference multiplies calls; weak review multiplies cognitive debt and attack surface with them.
— Tin