Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Moonshot AI's Kimi K3 scored an overall composite of 82.4 (95% CI 57.3–84.7) on iac-bench 2026-10, ranking 5th of 10 models and making it the highest-scoring open-weight model in the release, according to the benchmark r…
Anthropic's Claude Opus 5.5 ranked first among 10 language models in the iac-bench 2026-10 release, scoring 88.9 (95% CI 87.5–91.5) on 90 OpenTofu module attempts across AWS, GCP and Azure, with zero failed plans and zer…
A single "hi" message to an AI assistant with Drive and inbox connectors can produce an 18,000-token request, because the app resends a system prompt of a few thousand tokens, JSON-schema tool descriptions and the full c…
Amazon Bedrock AgentCore Evaluations now accepts decision models as code-based evaluators, according to a November 4, 2026 AWS Machine Learning blog post by Sungyeon Kim, Bharathi Srinivasan, Haibo Ding and Jonathan Buck…
Anthropic's Claude Help Center states that outputs from Claude Opus 4.6 and Sonnet 4.6 "may not reliably carry the text watermark when temperature, top_p, or top_k is set to a non-default value," a documented bypass firs…
Gemini 3.6 Flash scored an overall composite of 54.3 (95% CI 40.0–61.5) on iac-bench 2026-10, ranking 10th of 10 models evaluated, with security findings in 8 of 9 attempts and 1 failed gate, according to the benchmark's…
Iac-bench's 2026-10 release, run on 2026-10-22 and judged on 2026-10-23, measures how well language models write reusable OpenTofu infrastructure-as-code modules across three cloud fixtures: AWS S3 (8-input budget, zero …
OpenAI's GPT-6 Astra scored an overall composite of 86.7 (95% CI 86.0–89.8) on the iac-bench 2026-10 evaluation, ranking 3rd of 10 models across 9 attempts on three OpenTofu infrastructure fixtures, with zero failed gate…
Claude Opus 5.5 leads the Agent Index v1.2 preview, a November 2, 2026 snapshot of seven published agent benchmarks, with a composite score of 90.4 out of 100, ahead of GPT-6 Astra at 87.6 and Claude Sonnet 5.5 at 85.8. …
Buoyant released Kinference, an open source inference control plane for vLLM and Kubernetes that adds token rate limiting, tenant usage and cost tracking, KV-cache-aware load balancing, and inference policies to self-hos…
Ricursive Intelligence released Ripple-65B Preview on November 4, 2026, a 65-billion-parameter dense Transformer pretrained from scratch on 12 trillion tokens with a context window of up to 1M tokens, published under Apa…
Cloudy launched a free AI chat service and a terminal coding agent called Cloudy Code that share one account, offering 6 switchable models, 25 messages and 70 agent credits per day per account, and a stated policy that c…
Anthropic published FlashProve on November 2nd, an open-source research prototype that attaches a cryptographic proof to an AI response so a recipient can verify which committed model produced it, tested at 400B-paramete…
A developer behind Linkmortem, a dead-link checker that uses the open-weights Gemma 4 model to classify pages code cannot judge, reported that an AI crawler-targeting page attempted an indirect prompt injection with the …
A developer built brAIAIn, a parody AI lab whose word-level trigram model learns only from sentences customers pay $2.49 to submit, with each purchase producing a new model version complete with release notes, a research…
Nano-GPT launched as a privacy-focused alternative to OpenRouter, offering access to ChatGPT, Claude, Gemini and hundreds of other AI models in one place with pay-as-you-go pricing and chat history stored on the user's d…
Anthropic published a blog post Tuesday detailing FlashProve, an open-source system for "provable inference" that uses cryptographic proofs to verify which AI model produced a given output. The system requires labs to pu…
A developer built SynapseHealer, an autonomous dead-letter-queue repair pipeline on InterSystems IRIS AI Hub that uses a governed %AI.Agent with three fixed tool calls — vector runbook search, payload validation and risk…
Airbnb engineers reported that a fine-tuned open-weight classifier with 2 billion effective parameters (E2B), built with Airbnb's internal GLOWC framework, matched frontier proprietary models on a production content-clas…
OpenAI's published API specification no longer includes the prompt_cache_options field on CreateModelResponseProperties, affecting two OpenAI operations, according to an API-change report. The field is now ignored rather…