StepFun's Step 5 Preview is a proprietary reasoning model with 600B parameters released September 18, 2026, scoring 44 on the Artificial Analysis Intelligence Index with competitive pricing of $1.00 per 1M input tokens and $2.70 per 1M output tokens. The multimodal model supports text and image inputs, offers a 1M token context window, and ranks well above average in intelligence compared to similarly priced models.
Microsoft used an LLM agent to port Copilot runtime to Rust, reportedly spending $120K on the project. The Rust compiler handled the agentic translation without issues.
Researchers demonstrate that dishonest LLM service providers can manipulate token generation to inflate user costs while maintaining task quality. They identify five attack methods and develop a lightweight black-box audit technique achieving 85.1% detection rates, flagging 7 of 15 tested API services for suspicious behavior.
A writer investigates whether different AI language models have identifiable dash usage patterns, finding that Claude prefers spaced em dashes, OpenAI models traditionally use unspaced em dashes, and Grok switches between styles. The author concludes that dash patterns alone are insufficient to fingerprint models, though they experimented with changing their own writing style to avoid appearing LLM-generated.
Researchers analyzing 450,000 gender-directed completions across GPT-2 through GPT-5 found that explicit discriminatory content is transformed rather than removed through safety training, a phenomenon termed 'harm laundering.' While surface-form toxicity classifiers report declining harm scores, women-directed output shows reduced topic diversity and shifted representational biases, with men-directed content gaining positive attributes that women-directed completions lack.
Diogo Almeida, an OpenAI researcher who helped invent RLHF, left the company to start Typesafe AI and released Jev, a non-LLM transformer model that outputs probabilities instead of text for software automation tasks. Jev is significantly faster and cheaper than existing LLMs while eliminating hallucinations, with developers reporting 5-20x performance improvements in classification and routing applications.
TypeSafe AI released Jev, a specialized model that answers typed questions in parallel without generating text, sparking rapid adoption in browser agents, trading, and robotics. Unlike general-purpose LLMs, Jev provides calibrated probabilities and schema-constrained outputs, representing a shift toward specialized decision primitives optimized for specific tasks rather than conversational AI.