Tech • AI • Robotics • Game
Fable 5.5 Appears Early, Opus 6 Leaks, Meta Drops Muse Spark Open
9.3/10TL;DR
Anthropic appears close to releasing Claude Fable 5.5, while also expanding faster and cheaper Opus 5.5 access as competition intensifies across OpenAI and Meta.
KEY POINTS
Fable 5.5 surfaces on Azure
Developers found a new Azure model entry labeled cloudfable 555 in an East US directory with general availability status, suggesting Claude Fable 5.5 may be nearing launch. The listing is not a formal confirmation, and some caution that an open-source config sync is not an independent verification. Still, a similar pattern reportedly preceded an earlier Fable 5.1 rollout.
Leaked specifications point to long-context use
The leaked details show a 1 million-token context window, 128,000-token maximum output, and API pricing of $10 per million input tokens and $50 per million output tokens. Those figures are not unprecedented, but together they position the model for very large documents, persistent coding agents and extended multi-step work. The cost could limit practical use to enterprises and high-budget developers, especially for long sessions.
Signs of quiet early access
Some users suspect difficult prompts in Claude Opus 5.5 have already been routed to a stronger unreleased model. Informal tests rely on questions tied to newer benchmark and industry references that standard Opus 5.5 should not know because of its earlier knowledge cutoff. That has fueled speculation that Fable 5.5 is already in limited evaluation.
Creative and coding demos are driving hype
Early reports attribute unusually ambitious outputs to the suspected new model, including a playable MMO RPG built in about five hours, long-form visual explanations of technical proofs, and a code-generated animated short developed over roughly two and a half hours. Other examples include a 40,000-year human art history animation and a 9-minute 38-second symphonic audiovisual piece. The demonstrations have intensified debate over whether large language models can now compete with specialized video tools by generating code-driven media.
Rumors of Opus 6 remain weakly sourced
Claims that Anthropic is preparing Claude Opus 6 for an imminent release have spread widely, but they remain poorly substantiated. No model card, benchmark set, pricing, release date or official confirmation has surfaced. Analysts note that a company approaching an IPO would have incentives to keep launches disciplined rather than respond to rumor cycles.
OpenAI and Astra speculation add pressure
Industry chatter suggests OpenAI may promote a model called 6.1 Astra during an upcoming shipping push, although reports have also claimed delays tied to security review. Separate rumors around a project called Bell are even murkier, with wildly inconsistent parameter estimates and no public confirmation. The result is a market driven as much by leak culture as by announced products.
Opus 5.5 fast mode is now broadly available
Anthropic has officially expanded fast mode for Claude Opus 5.5, removing waitlists and making it available to developers through API and Azure. The company says the mode uses the same weights and capabilities as standard service while delivering up to 2.5 times higher output-token speed through inference and API optimization. The feature is aimed at agent loops, debugging and other latency-sensitive work.
Faster access comes with lower pricing
In a notable pricing move, Opus 5.5 fast mode is listed at $8 per million input tokens and $40 per million output tokens, undercutting older fast modes priced at $10 and $50. Fast mode has separate rate limits and falls back to standard speed rather than failing when caps are hit. One operational caveat is that prompt cache is not shared between fast and standard sessions, so switching modes midstream can trigger a fresh uncached input charge.
Idle compaction aims to stop token waste
Anthropic has also introduced idle compaction, a feature designed to prevent long inactive sessions from rereading huge contexts after cache expiry. When conditions are met, including roughly 200,000 tokens in context and about 54 minutes of inactivity, the system compresses the conversation into a tighter summary before cache expiration. Some developers have reported detail loss after overnight compaction, and controls have been added to disable or tune the behavior.
Meta revives open-weight competition with Apple ties
Meta has released Muse Spark Open, describing it as the most powerful U.S. open-weight large model. It is a multimodal system with 816 billion total parameters, 42 billion active parameters, and a 1 million-token context window, optimized for coding, agents and long-running reasoning tasks. Meta said it worked closely with Apple to optimize the model for MLX on Apple Silicon, though running it locally at full scale could require two top-end Mac Studio systems costing more than $40,000.
CONCLUSION
The immediate shift is less about a confirmed Fable 5.5 launch than about a rapidly escalating contest over speed, scale and deployment flexibility. If the leaked Fable specifications hold, Anthropic is positioning itself to compete simultaneously in premium enterprise inference, consumer coding workflows and the next wave of AI agent systems.
Ask a question