Interactive model
Click any turn to switch models there. Everything is priced against running the nominal trace on Model A. Add extra routed turns to see how much accidental work the routing can absorb before it stops paying back.
DefaultsDollars per 1M tokens, from OpenAI standard short-context pricing for GPT-6 Astra (A) and GPT-6.1 Sol (B). Cache writes cost 1.25× input and reads 0.1×. For a model with no write premium, set write equal to input.
Extra turnsTurns past the nominal horizon are shaded on the track. They stand for extra work caused by routing, and the all-A baseline never pays for them.
SavedHow much less each nominal turn on B costs than the same turn on A when both have a warm cache, through cheaper cache reads, cache writes and output.
Switch costPaid whenever a model has to write context it would otherwise have read from its cache, at (cache write − cache read) × those tokens. That covers the jump into B and any return to A. An expiry on a switch point adds to it. An expiry elsewhere does not, since the baseline pays that cold write too.
Extra turnsThe warm price of turns past the nominal end. The three terms add up to the net gain exactly.
Cumulative stacks the running total. Per turn shows each turn on its own, so a switch shows up as a cache-write spike that decays as the new model’s context warms back up.
Billing follows OpenAI’s GPT-5.6+ prompt caching. Each request reads the prefix cached through the previous request’s input and writes everything after it, which is the previous turn’s output plus the new input, at the cache write price. Prompts under 1,024 tokens aren’t cached and pay the plain input price. Any request whose prompt exceeds 272K tokens is billed entirely at long-context rates, with input, cache write and cache read doubled and output at 1.5×.
Clicking a turn toggles a switch at that point, and later turns follow until the next switch. In expire mode a click marks an expiry before that turn, shown as a ring. An expiry on a switch point means only the model you switch to starts cold, while the model you left keeps its cache and the all-A baseline is unaffected. An expiry anywhere else is a real-time pause, so both models lose their caches and the baseline loses its cache at the same turn.