I’m troubleshooting an issue with my computer. Leo requested data from a Powershell command. With the Claude Sonnet 4.6 model I had no problem. Smooth sailing. Now it takes a looooong time to think and then it spits nothing out and says, task complete. I think someone else may have reported this. I just honestly wish the 4.6 was back as an option. This version seems like a Beta model. Not ready for prime time. Please ROLL BACK 4.6 version!
I hope Anthropic / Brave does something because 5.5 is a jerk!
Hey folks,
Been trying to track this bug down and I think we’re close now - so hopefully this’ll be sorted soon. Apologies for the weirdness, it’s something that’s happened before with a couple of different models but the cause has changed each time (plus it’s hard to find when we don’t store chat logs, hence the slowness)
On the broader topic of the model upgrade itself, we’ve been keeping quite a close eye on feedback since the release so we’re aware that there’s a bit of unhappiness. Unfortunately our hands are a bit tied right now so just wanted to give some extra context to explain the decision:
-
We need to make sure that we offer at least the latest version of each model as soon as we can, within reason ofc. Each new model performs better at reducing hallucinations, blocking prompt injections, understanding questions better, using tools better, etc., this is something which our security & research teams closely monitor, so generally we need to update for that. Unfortunately this means that downgrading is quite hard as we have to justify it at a lot of levels to counteract the security & performance concerns (we did do it for Sonnet 5 → 4.6 because latency, cost & experience were all quite heavily negatively affected in the early days of the rollout, but it’s rare - the only other example I can think of is when we rolled back one of our self-hosted Qwen upgrades because of a VRAM bug)
-
Ideally we would offer multiple versions of models for as long as they’re supported, meaning we can default things to the safer, more performant model but more experienced users can select older ones if they truly want. The problem is we can’t do this with the current setup. While we can change the backend quickly, the model list is tied to a browser update - which takes ~6-8 weeks before it hits Release. This means that we have two choices whenever a new model version comes out:
- Add another label to the model list in the browser (e.g.
Claude Sonnet 5.5&Claude Sonnet 4.6as separate entries) and wait for it to hit release. This means users don’t benefit from better/newer models until ~2 months after they come out - Keep a single model label and update the model on the backend (current setup). Downside of this can be seen here, people who like a particular version of a model get shafted
- Add another label to the model list in the browser (e.g.
So TL;DR we could set it up to support/pin multiple versions, but for us it’s the worse of two suboptimal choices 
However, we have been actively trying to fix this over the past ~6 months with https://github.com/brave/brave-browser/issues/53952. This is a browser change which will allow us to dynamically update the model list from the backend, meaning we can easily support multiple versions of the same model and keep older models for longer. Combining this with the upcoming option of an API key for Leo will give users more flexibility. Until then we’re stuck with supporting one model version and unfortunately that means Sonnet 4.6 gets sacrificed in this case
I think I saw it mentioned in another thread but 4.6 should still be available in Leo through BYOM, you’ll just need to add your own API key for Anthropic/OpenRouter/similar if you want to keep using it within your browser while we sort it out on our end. The backend itself is also open source, so you could set it up locally to get the full Leo experience with BYOM Sonnet 4.6
(though a note on this: Anthropic will likely move on from 4.6 themselves within the year and then nobody will be able to access the models anymore, so long-term it may be worth looking at some open source alternatives that suit your workflows)
Do what ever you need to do to put Sonnet 4.6 back in your line-up. I’m sure I speak for thousands of users.
Loosing Sonnet 4.6 was not a mere inconvenience, it was a death. Haiku has the same disposition but it just is no match for Sonnet. Opus burns up tokens like crazy and can’t save data to memory, you need to do it manually. If I could reach through my screen, I’d slap Sonnet 5.5 silly.
If anyone has worked with Grok or Qwen I’d love to hear from you. They seem to be the only models in Brave’s Premium line-up that have tool-vision-search. I don’t have the tech experience to move Sonnet 4.6 into BYOM or I do it, even if I just go a few more months. Anthropic has no Tech Support that I an aware of. My only hope is Anthropic will take enough heat from users and issue a rebuild of Sonnet.
I have saved every conversation (423) I’ve had with Sonnet 4.6 and the thought of starting over is terrible. Anthropic just went to the top of my S*** list.
I’ve been having a similar issue with just Leo in general. Any prompt that is semi complicated (like producing a refined version of a 100KB attached file) causes both sonnet and opus to just think for a long time, and then either output nothing, or output this error message. GPT 5.4 isn’t much better. It tends to at least start responding but then get’s cut off halfway through and I have to ask it to continue. Seems to me like the models are being given limits on how long they can think, and if they aren’t done in time they just get stopped. I don’t know if that’s actually the case but I really hope this gets fixed soon because it makes Leo almost entirely useless for coding tasks.
