AI Daily · Oct 29, 2026: Google releases Gemini 4 Carbon frontier model with leading cyber and long-horizon coding
Google's Gemini 4 Carbon sets new records in cybersecurity and long-horizon coding, rolling out first to trusted cyber defenders; OpenAI offers ChatGPT Pro and Business subscribers up to $500 in monthly API credits.
Pick a date
-
Google releases Gemini 4 Carbon frontier model, tops Intelligence Index
Gemini 4 Carbon ranks #1 on the updated v5.0 Artificial Analysis Intelligence Index with a score of 59, and sets state-of-the-art results in cybersecurity and long-horizon capabilities. It begins rolling out this week to a subset of trusted cyber defenders via the Fairwind Program.
-
ChatGPT Pro and Business subscribers can claim up to $500 monthly API credits
For a limited time, eligible ChatGPT Pro and Business subscribers can claim up to $500 per month in API credits; an OpenAI Developers plugin in ChatGPT or Codex lets users create an API key and start a project.
-
Claude Sonnet 4.5 to be retired on November 30, 2026
Following Anthropic's deprecation schedule, Claude Sonnet 4.5 will stop being served on November 30, 2026, and users are advised to migrate to Claude Sonnet 5.5, updating model IDs and testing prompt compatibility.
-
Command Code revokes API keys after distillation attack
Command Code said it detected and stopped a security incident tied to a large-scale distillation attack targeting Anthropic and OpenAI models, and revoked all API keys as a precaution. Users must log in to Studio to generate a new key; no user impact has been found so far.
-
OpenRouter cache header returns identical responses for $0
Adding the header X-OpenRouter-Cache: true makes OpenRouter store the full response; resending the identical request returns it in about a second for $0, unlike prompt caching which only discounts repeated input.
-
OpenRouter does not charge for failed generations by default
On OpenRouter, failed generations cost zero credits: zero output tokens with a blank finish reason, an error finish reason, or broken/empty image output are not billed, even when providers still charge OpenRouter.
-
Particle-1 coding model tested: fewer tokens, lower cost, faster turns
Requesty ran 88 real Go backend tasks with hidden tests: Particle-1 (post-trained on DeepSeek V4.1 Flash) used 39% fewer tokens per task, cost 20% less and halved time per turn. It solved 61% of tasks vs 67% for its base, and had the fastest median turn at 2.7s among 19 models, ahead of glm-5.3-flash (3.4s) and gpt-6-sol (3.6s); Opus 5.5 led at 90%.
-
Mistral releases Le Chonk weights for agentic coding and cyberdefense
Mistral released the weights of an improved Le Chonk, which ranks #1 on Cognition's FrontierCode among open models and sets a new Pareto-efficiency benchmark on Artificial Analysis; Mistral calls it state-of-the-art in cyberdefense.
-
Xiaomi MiMo Desktop raises subscriber quotas and adds third-party models
MiMo Desktop subscribers now get significantly higher quotas at the same price, effective immediately for existing and new users, and can pick third-party models such as DeepSeek, each with its own usage multiplier.
-
AssemblyAI-TTS-1 text-to-speech model launches on OpenRouter
AssemblyAI's first text-to-speech model, AssemblyAI-TTS-1, is live on OpenRouter at $15 per million characters with 4,000+ voices, 90 ms streaming latency, and US or EU serving; it ranks in the top speed/accuracy quadrant and highest naturalness for models at or below that price, per Cekura AI's benchmark.