Gemini 3.5 Pro reportedly restarted pretraining over tool-calling recursion
Gemini 3.5 Pro reportedly restarted pretraining to fix recursive tool-calling — here's what that means for agent loops and what 2M context costs per run.
Google reportedly scrapped the original Gemini 3.5 Pro base model and restarted pretraining because recursive tool-calling was breaking. That's the interesting part, not the July 17 date or the 2M-token context number. If you build agents, the only question that matters is whether a chained tool loop survives more than a few hops without the model eating its own output.
None of the specs are confirmed by Google, so treat the rest as rumor with a receipt: 2M-token context, Deep Think mode gated to the $250 Ultra tier, and roughly $1.25/$10 per million input/output tokens. Those numbers come from the leaked specs report, and Google has said nothing on the record.
Why does restarting pretraining matter?
Restarting a base model isn't a fine-tune tweak. It means the tool-calling failure was baked deep enough that RLHF and post-training couldn't patch over it. In practice, recursive tool-calling recursion looks like this: the model calls a tool, gets a result, and instead of reasoning about the result it calls the same tool again with slightly mangled args, then loops until it hits a token ceiling or emits garbage.
If that behavior lived in the base weights, a full rebuild is the honest fix. It also tells you the previous checkpoint was bad enough to justify burning months of compute rather than shipping. That's a real signal about how brittle current agent loops are across every frontier model, not just Google's.
What does 2M context actually cost per run?
Here's the math nobody puts in the launch posts. At the rumored $1.25 per million input tokens, filling a 2M context window once costs you $2.50 in input alone — before the model writes a single token back. Do that on every step of a 10-step agent loop and you're at $25 per run just feeding context, assuming you re-send the window each turn.
Caching helps, but only if your context is stable across turns, and agent loops are the opposite of stable — the whole point is that state changes every step. So the realistic pattern is: large static context (docs, tools, schema) cached, plus a growing scratchpad you pay full price for. Budget accordingly before you wire 2M tokens into a production agent.
Should you wait for it?
I'm not moving anything until it ships and I can run my own agent regression suite against it. A model that needed a base-model rebuild to stop looping is exactly the model you test on a real chained task — call tool, parse result, branch, call another tool, five hops deep — not a single-shot benchmark. The 2M context is a nice number that most people will never fill usefully, and if they do, the bill will teach them not to.
Verdict: worth testing on day one for agent workloads specifically, because the recursion fix is the actual feature here. Ignore the context-window headline. Run it against a loop that already breaks your current model and see if it holds. If it survives ten hops without degrading, that's the story. If it doesn't, the rebuild didn't buy you anything.
FAQ
When is Gemini 3.5 Pro launching?
It's widely reported to target July 17, but Google has not officially confirmed the date. Treat it as a rumor until there's an on-the-record announcement.
Is the 2M-token context window confirmed?
No. The 2M context, Deep Think on the $250 Ultra tier, and the pricing are all from leaks that Google has not verified. Assume the numbers can move before launch.
What's the real reason the model was rebuilt?
Google reportedly restarted pretraining because recursive tool-calling was failing in the original base model. That points to a defect deep in the weights, not something post-training could patch.
Is it worth it for agent workloads?
Potentially, if the recursion fix holds. Test it on a real chained tool loop that already breaks your current model before committing anything to production.