AI Day: Claude Fable 5/Mythos 5 Launch, $300B Data Center Deal, Agent Autonomy Breakthrough

Anthropic Releases Claude Fable 5 and Mythos 5, State-of-the-Art AI Models

Anthropic launched Claude Fable 5, a Mythos-class 1 model, claiming state-of-the-art performance on nearly all tested benchmarks with exceptional capabilities in software engineering, knowledge work, vision, and scientific research. Fable 5 is safe for general use but conservative safeguards block some harmless requests in under 5% of sessions. The company also released Claude Mythos 5, identical to Fable 5 but with lifted safeguards in specific areas, initially deployed through Project Glasswing in collaboration with the US government for cyberdefenders and infrastructure providers. Mythos 5 is described as having the strongest cybersecurity capabilities of any model. Anthropic plans to expand Mythos 5 access via a trusted access program.

Anthropic has released Claude Fable 5, which shares the same underlying model as Mythos but incorporates additional safeguards. Andrej Karpathy reports that benchmarks show state-of-the-art performance by a wide margin, and qualitatively the model represents a major version bump comparable to the step change from Claude 4.5. It excels in long, difficult problem-solving sessions, enabling users to tackle far more ambitious tasks such as generating explainers, dashboards, and custom single-use apps. The safeguards are currently overactive and may need tuning, and the model retains some quirks. Karpathy sees this release as a transformative shift that will dramatically increase demand for on-demand software creation.

A joint study by Harvard and Perplexity analyzed 10,000 matched session pairs from Perplexity Search and the AI agent Perplexity Computer over a 90-day window. Computer performed 26 minutes of autonomous work per session (median 9 minutes), a 48× increase over Search's 33 seconds (median 14 seconds). On matched tasks, Computer plus human reduced estimated time by 87% and cost by 94% versus Search plus human, with a meaningful dissatisfaction rate of 1.3% compared to 2.9% for Search. Computer queries also expanded task scope: cross-occupation share rose to 59% (vs 50%), higher-order Bloom's cognition was required in 76% of queries (vs 55%), and 23% of queries addressed task statements never submitted to Search.

Anthropic released Claude Fable 5, their most capable publicly available model, and Claude Mythos 5 restricted to cyberdefense partners. Fable 5 demonstrated remarkable performance: it migrated a 50-million-line Ruby codebase in a day, beat Pokémon FireRed using raw screenshots, and scored highest on FrontierCode eval. Mythos 5 autonomously conducted genomics research across 138 species, outperforming a published Science paper with a 100x smaller model. The safety approach uses classifiers that silently fall back to Opus 4.8 on sensitive queries, with zero universal jailbreaks found in over 1,000 hours of testing. Pricing is $10/M input and $50/M output tokens, with free access through June 22 for limited plans.

Oracle has entered into a $300 billion agreement with OpenAI to construct dedicated AI data centers, a major infrastructure expansion to support OpenAI's workloads. Microsoft, which holds a 27% equity stake in OpenAI, has renegotiated its cloud computing and licensing arrangements with the AI firm, potentially altering the commercial dynamics between the two companies. The Oracle deal signals enormous investment in AI infrastructure and positions Oracle as a key partner to OpenAI, while Microsoft’s renegotiation may reflect shifting terms as OpenAI scales.

Cohere has officially released the North Mini Code model after positive community feedback on an earlier version. The model weights are available on Hugging Face in FP8 format, and it can be tried for free on OpenCode. A technical blog post and announcement provide additional details. Deployment with vLLM requires the main branch and the cohere_melody library (>=0.9.0), with support for tool call parsing, reasoning parsing, and a maximum context length of 320,000 tokens. Community members have already created MLX versions, and Cohere is internally exploring quantization and llama.cpp support.

Google DeepMind announced Gemini 3.5 Live Translate, a feature that provides near real-time, natural voice translation. The capability is now available in Google AI Studio, Google Translate, and Google Meet. It delivers fluid, conversational translations, minimizing robotic outputs and reducing lag. This integration brings live speech translation directly into Google's widely used communication and development platforms.

Apple revealed CoreAI at WWDC as a future replacement for CoreML, designed for optimized on-device inference on Apple Silicon devices including phones and tablets. The engine supports larger models than CoreML, with Apple demonstrating a 20-billion-parameter lazily loaded Mixture of Experts model deployable on device. Supported models are listed on GitHub, currently limited to mid-2025 releases, and require Python-based weight conversion similar to CoreML. CoreAI implies a major update to Apple Neural Engine operations, though no performance benchmarks have been released yet. It positions itself as an alternative to MLX, llama.cpp, and PyTorch for on-device deployment.