Tech • AI • Robotics • Game
Claude Fable 5.2 Is Insane! Fable 6 Hidden, Gemini Krypton, GPT-6.2 & Sol! AI News
9.4/10TL;DR
Anthropic has launched Claude Fable 5.2 and Claude Mythos 5.2, positioning them as faster, stronger models for coding, reasoning, vision, robotics and autonomous technical work, while rivals at OpenAI, Google and Meta are also preparing new systems.
KEY POINTS
Anthropic’s new flagship pair
Fable 5.2 and Mythos 5.2 were introduced as high-end models for difficult technical projects. Anthropic says they improve coding, reasoning and vision, while delivering outputs up to 30% faster than their predecessors. Internal comparisons reportedly place them ahead of Opus 5.5 on most evaluated capabilities.
Benchmarks and pricing
On Frontier Code 2.0, Fable 5.2 at high effort reportedly matched Opus 5.5 at maximum effort at roughly half the per-task cost. Anthropic lists pricing at $8 per million input tokens and $40 per million output tokens, about 20% cheaper than Fable 5.1, though the model remains expensive when used at its highest reasoning settings.
One model, two access tiers
Fable 5.2 and Mythos 5.2 are described as the same underlying model with different safeguards. Fable carries stricter protections against misuse, while Mythos is reserved for verified organizations working on beneficial cybersecurity and biomedical research through a specialized access program.
Autonomous engineering claims
Anthropic highlighted unusually broad real-world performance. In one demonstration, the model coordinated more than 1,000 sub-agents over eight days to build a prototype operating system from an empty folder, producing a bootable system with its own kernel, drivers, file storage, installer and cloud functions.
Hardware and robotics progress
The company also pointed to gains in robotics, saying the model can control physical robots and handle tasks it was not explicitly trained on. In another test, a single short prompt reportedly led the model to design three electronic devices, selecting components, laying out circuit boards, creating 3D-printable enclosures and programming the software, with boards ready for manufacturing in under 24 hours.
Costly but striking creative output
Early users have showcased long-form code generation at unusual scale. One game build reportedly consumed 445 million input tokens and 2.7 million output tokens over about eight hours, for an API cost near $550. Other examples included browser-based game clones, complex 3D scenes, a 48-minute coded film and detailed Blender models.
Hints of a stronger unreleased model
The absence of a rumored Fable 5.5 has fueled speculation that Anthropic is holding back a more powerful checkpoint. Training details suggest Fable 5.2 was deliberately constrained on some AI self-improvement tasks, including chip kernel optimization and improving large language model architectures, while other technical abilities remained less restricted.
Subscriber incentives and IPO timing
Anthropic is granting some subscribers one to three free usage resets based on loyalty, and Pro users are getting a weekly Fable allowance equal to 50% of their normal weekly limit. The release comes as Bloomberg reported Anthropic could begin marketing an IPO as early as Tuesday, with a Nasdaq debut potentially next week.
OpenAI and Google close behind
Reports also point to a new GPT 6.2 checkpoint being tested in Codex, with improved visual and coding output surfacing through some higher-reasoning routes. At Google, an early model codenamed Gemini Krypton has appeared in backend logs and is said to outperform Gemini 4 Carbon, while Gemini 4 Hydrogen is reportedly in private testing.
Meta enters with a cheaper model
Meta introduced Muse Ember, a smaller multimodal model aimed at reasoning and agent workflows. Meta says it supports a 1 million-token context window, can operate across apps and tools, and delivers performance near Muse Spark 1.3 Max at about 17 times lower cost, with pricing of $0.10 per million input tokens and $0.50 per million output tokens.
Safety and infrastructure moves
OpenAI also unveiled Control Room, an emergency shutdown system for frontier-model research workloads. The system can reportedly halt output generation within 35 seconds of a shutdown decision, and defaults to blocking inference if the control layer fails or is tampered with.
CONCLUSION
The latest model releases suggest the AI race is shifting from benchmark gains to expensive but increasingly autonomous execution across software, hardware and research tasks. The main question now is not whether capability is rising, but how quickly labs choose to deploy their strongest systems.
Ask a question