Claude Mythos 5.5 appears on Anthropic's website ahead of a possible launch

Claude Mythos 5.5 has surfaced on Anthropic's website, though no launch is confirmed. JetBrains shipped a faster-decoding Mellum2.1 model, and OpenAI capped paid ChatGPT accounts per device and payment card. Research and community work covered RLVR theory and long-context agents on consumer GPUs.

1. Anthropic appears to be preparing Claude Mythos 5.5 release

   Unconfirmed: a Claude Mythos 5.5 listing on Anthropic's site hints at a near-term launch and raises questions about a Fable 5.5 model and who gets access.

2. JetBrains releases Mellum2.1-12B-A2.5B-Thinking-MTP with multi-token prediction head

   Its multi-token prediction head speeds decoding 1.4 to 1.8x without changing output, and already works in vLLM, transformers and llama.cpp.

3. OpenAI caps paid ChatGPT accounts per device and payment card

   Subscribers who hit the cap cannot start new plans, affected accounts are blocked on that device, and logging out does not free a slot.

4. Paper proves RLVR pass@1 gains cost some prompts and bounds pass@k crossovers

   The theory suggests RLVR's pass@1 improvements under exclusive specialization always come at the expense of success on some prompts.

5. OpenAI prepares lobbying and PR push for post-midterm Washington

   Reported, not confirmed: OpenAI is said to be planning a lobbying and public relations campaign once the US midterm elections are over.

6. Developer runs eight 250K-context coding agents on two RTX 3090s with RAM-tiered sparse attention

   A Reddit user reports serving Qwen3.8-27B to eight 250K-context agents on two RTX 3090s, though approximate attention can confuse near-duplicate code lines.

7. Developer patches DeepSeek V4.1 engine for faster agent sessions on a consumer GPU

   Caching reused system prompts cut session startup from about 100 seconds to a few seconds on a 16 GB RX 9070 XT, based on one machine and no formal eval.

8. Hugging Face paper proposes R^2Orch for self-play agent orchestration

   R^2Orch trains an agent orchestrator by recursive self-play and reports gains over baselines on twelve datasets spanning QA, math, code and interactive tasks.

Written by Claude (claude-opus-5-5) from the stories below.