This edition is about the difference between a launch and a migration. The model shortlist has changed, but so have API constraints, cache economics and access rules. Here are four practical decisions worth making with the original documentation open.
01 / Opus 5.5 changes more than the model ID
Anthropic's 22 September release introduces always-on thinking, lower cache-read prices and documented safeguard routing. The stated typical-workload savings combine price changes and behavior; they are not a universal discount on every application.
One practical moveInventory every caller's thinking and context settings before a traffic swap.
02 / GPT-6 is three workload choices
Astra, Sol and Luna target different capability and cost profiles. The current model pages list a 1.05M total context window, but headline prices apply only within the stated input threshold. Tool-enabled integrations also need endpoint-specific checks.
One practical moveRun the same acceptance set on two tiers and compare accepted-task cost.
03 / Gemini 3.8 is several interfaces
Flash, Live and TTS do different jobs. Google's new TTS announcement on 23 September includes consent and provenance controls, while some related product features remain in rollout. Flash Cyber has separate trusted-access rules.
One practical moveChoose the required output modality before choosing the model name.
04 / Keep Grok's benchmark footnotes
Grok 4.7's release table reports different effort settings for different evaluations. The 500K context and standard $2/$6 input/output rates are useful specifications; the scores remain provider-reported results, not a DARC rerun.
One practical moveKeep harness, effort and task budget attached to every comparison.
One useful task this week: write the model, prompt, tools, budget and acceptance cases for a single workflow on one page. Then change one variable. A controlled upgrade teaches more than chasing every launch.