| in the first week of 2026 i ended up using cursor and google antigravity back to back, not by plan but because i burned through two cursor ultra subscriptions faster than expected and decided to try antigravity on the free tier. my normal usage is ~$60–100/month. within a few days it jumped to $500+, with the dashboard projecting ~$1.6k/month. max mode was off, and the ui consistently showed a 200k context window. what i eventually pieced together is that cursor maintains a large hidden prompt state. beyond visible conversation history, this includes tool traces, agent state, reasoning scaffolding, and large chunks of repo context. that state is prompt-cached using claude’s cache feature, and on every request the full cached prefix is replayed. anthropic bills cache reads every time this happens, even if that content is later summarized or truncated before actual inference. one concrete example from my logs:
• actual user input ~4k tokens
• cache read tokens ~21 million
• total billed tokens ~22 million
• cost for a single call ~$12 this wasn’t limited to opus. i saw the same pattern with sonnet. support explained that this matched how the underlying api is billed. from my perspective, the issue wasn’t correctness but visibility. cost had become decoupled from anything i could see or reason about in the product. i had no way to inspect cache size, understand replay behavior, or set meaningful guardrails. i canceled and treated it as an excuse to try google antigravity. the free tier was more usable than i expected. it gives access to opus 4.5, which is still my preferred model for nontrivial coding work. for easy to moderate complexity tasks, it usually finishes cleanly. limits are opaque (free → pro → ultra is described in very abstract terms), but when you hit a limit you at least get a clear cooldown message telling you when opus will be available again. when opus is exhausted, antigravity falls back to gemini models. that was useful for comparison. for real coding work on a messy, evolving codebase, gemini flash, pro, and thinking consistently lost architectural decisions and produced one-off changes that didn’t respect existing constraints. some of that is model quality, but not all of it. cursor’s agent does a better job gathering relevant repo state, forming a plan, and executing it coherently. antigravity’s agent feels thinner, and i found myself spending more time reviewing and correcting diffs to preserve invariants. there were also smaller papercuts. tab completion is advertised as unlimited, but i couldn’t get it working reliably. going back to basic autocomplete was a reminder of how dependent my workflow is on good tab completion. ux-wise, antigravity feels slower. maybe not raw latency, but the way responses stream and animate doesn’t keep me in the same loop cursor does. net result: antigravity free is a solid option for starting out and experimenting, especially if budget matters. i wouldn’t pay for it yet. cursor is still a strong product, but opaque caching and billing behavior makes it hard to reason about cost at scale. for context, i’m building an agent runtime myself at inference.sh, focused on explicit state, durable execution, and reliable deep agents with complex tool use. because of that, i’m probably more sensitive than most to differences in agent orchestration, hidden state, and how cost emerges from those design choices. this whole experience reinforced something i already believed: hidden state is dangerous in agent systems. hidden state combined with opaque billing is worse. if users can’t see state, they can’t reason about cost. and if they can’t reason about cost, they won’t trust the system. right now, none of these coding agents are “set and forget” if you’re doing work that hasn’t been done a thousand times before. you still have to stay in charge. |