Anthropic may be preparing another Claude refresh, with two experimental models, claude-marshmallow-eap and claude-melon-eap, reportedly appearing in testing. The names follow the same food-themed naming pattern as the earlier “Horchata” experiment, which briefly appeared before disappearing from Anthropic’s model lineup. Early testers reportedly find Marshmallow stronger than Melon, with Marshmallow even getting praise for its conversational quality compared with the current Opus model.

Neither model is reportedly at the level of Anthropic’s Fable-class models, which makes their exact place in the lineup unclear. The leading theory is that Marshmallow and Melon could be upcoming Opus and Sonnet updates, potentially versions like 5.1, although some speculation points to Haiku being involved as well. Anthropic has not confirmed any of this, but the timing is interesting. With multiple experimental models now showing up, it looks increasingly likely that Anthropic has something new in the pipeline. The bigger question is whether these are simply smaller refreshes or whether the company is still quietly preparing the much-anticipated Fable successor.

The mystery surrounding Ox Alpha is getting even more interesting. The model first appeared through OpenCode and OpenRouter with a 1M-token context window, multimodal capabilities, zero data retention and an unusually generous free trial. It was essentially dropped into the AI ecosystem with no clear information about who built it, and developers immediately started trying to reverse-engineer its identity from its behaviour. OpenRouter currently lists it as an anonymous stealth model, adding even more fuel to the guessing game.

Theories have ranged from Z.ai and GLM-5.3 Flash to Xiaomi and MiniMax, with some users trying to match its capabilities, architecture and even its reported compute requirements to existing labs. The GLM theory has gained particular attention because Ox Alpha's coding performance appears to sit in a similar range to GLM models, but there is still no confirmation that the two are connected.

What we do know is that Ox Alpha has completed the full DeepSWE evaluation at around 63%, putting it roughly alongside several well-known frontier models. For a model that seemingly appeared out of nowhere, that is pretty impressive. At this point, the biggest mystery may not be what Ox Alpha can do, but who had enough compute to quietly build it and then give it away for free.

OpenAI has propagated another Codex usage reset to paid users after investigating recent quota issues. The team says several fixes have also been shipped, including changes related to image handling and other features that were causing usage to drain faster than expected. More updates are expected as OpenAI continues investigating the issue.