Will legacy users ever have access to newer models again? It doesn’t always have to be the very latest model. But having access to one of the older models that came before – which might even be cheaper – would surely be an advantage for Cursor itself. Besides, the old Opus, Codex, GPT and Sonnet models probably won’t keep running forever. At some point, they’ll be taken offline. And then legacy users would no longer have any access to high-end models at all, unless they activate MAX – which is expensive… even more expensive than switching to the new Pro plan. Could the planned takeover by SpaceX be influencing this decision?
Hey, thanks for the question, there’s an important nuance here.
On the legacy request-based plan, you won’t be left without strong models. Auto mode will still route to our current models within your included requests, and it’ll keep pointing to up-to-date models as older ones are phased out. So you won’t end up in a situation where you have no access to powerful models at all.
What’s actually gated is directly selecting the newest frontier models on a flat request rate. On the legacy plan, you need Max Mode for that, and it’s billed at the model’s token rate. If you want flat-rate access to the newest models, that’s available on the current token-based plan.
About having a cheaper way to access new models on legacy plans, I hear you and I’ll log that feedback.
@deanrie
But isn’t it illogical that you’re hiding all the new models behind MAX, even those that might actually cost you significantly less? For example, GLM-5.2 costs less than, say, Opus-4.6… wouldn’t it make sense here for legacy users to at least be able to use models that would cost you less? So this might be a silly question… but anyone who can do the maths can see the benefit. Instead of legacy users potentially having to use the last available, possibly expensive high-end models, they could switch to newer ones that cost you less. To me, that makes sense from a purely financial point of view.
I get the logic, from a pure math point of view it makes sense.
The key thing is that this isn’t decided per model based on its individual cost. On the legacy request-based plan there’s one rule: manually selecting frontier models is only available via Max Mode based on the token rate, no matter whether a specific model is cheaper or more expensive. So the gate is tied to the plan and the frontier category, not to the relative price of a single model.
That said, you’re not left without strong models. Auto mode will keep routing to current models within your included requests and will pull in newer ones as older ones are phased out. Max Mode is specifically gating manual, direct access to the newest frontier models.
I’ve noted your point about a cheaper path to new models on the legacy plan. I don’t have a timeline or a policy change to share right now, so I won’t promise anything, but the argument is clear.
@deanrie
At the moment, I can live with the situation. There are still a few good models available, even if they aren’t the very latest ones. But these simply incur higher costs for Cursor itself than I would if, for example, I could use GLM-5.2. That’s the background to my question. And yes, Auto is certainly usable. But in my last test with Auto, where the task involved calculating a cost estimate for a project, the calculations turned out to be very poor… and it turned out that it was Composer 2.5. I then ran the same task with a different model without using Auto, and the results were considerably better. Unfortunately, Auto is so unpredictable. You never know what you’re going to get.
How about this: instead of trying to calculate the direct token costs based on the token price / ($20/500 = $0.04) + 20%, what if, for newer models, let’s say 10 (vs. 2) requests were charged in Thinking Mode and 5(vs 1) without Thinking Mode… then there’d already be a surcharge… but not as wildly unpredictable as it is at the moment. The current system is completely unusable… with a flagship model, you’ve practically used up all your tokens after just one session or request. That’s not good.
Hey, got the logic in your proposal: a predictable request multiplier for new models instead of a token-based calculation in Max Mode. The predictability point is totally fair.
I’ll be upfront so I don’t set the wrong expectations: on the legacy request-based plan, there’s one rule right now. Manual selection of frontier models only works via Max Mode at the token rate, no matter if a specific model is cheaper or more expensive. The gate is tied to the plan and the frontier category, not the price of a specific model. I don’t have any updates or a timeline for changing this, so I won’t promise anything.
What still works without Max Mode: Auto will keep routing to current models within your included requests and will pull in new ones as older ones are phased out. I know you had a bad experience with Auto in post #8 (Composer 2.5 vs manual selection). That’s a separate routing quality topic, and if you hit a case like that again, send the Request ID and I’ll check the specific example.
I’ve noted your pricing suggestion and I’ll pass it along. The argument is clear.