v22026-08-01·New model, faster boxes, tighter limits
- New coding model on the fleet — stronger tool-calling and reasoning, and it holds a plan across many steps instead of stopping halfway. Runs with full-precision KV cache, which fixed the multi-turn edit failures some users hit.
- Much faster responses, and no more waiting behind other sessions at busy times.
- Long answers no longer run away — they finish in minutes, and a cut-off reply now tells you why instead of stopping silently.
- Context window stated honestly at 200K, the model's real ceiling.
- Clearer plan limits, with a per-session allowance that scales across the Lite tiers.
- Windows setup smoothed out — fewer failed commands, clearer install docs.