Flash news
Every thread that mentions this topic, across all sections
Reddit user reports 'Argon' AI model hallucination-free for 5 minutes, contrasting it with Flash's hedging and lying
Users praise 'Argon' model for research reliability, citing few hallucinations
Users praise 'Argon' model for research reliability, citing few hallucinations
Google is changing Gemini model availability depending on your subscription plan
Google announced changes to Gemini Apps model availability on personal accounts taking effect October 9, 2026. Free users will be limited to Flash Lite, AI Plus subscribers get Flash added, and AI Pro/Ultra subscribers keep access to Pro plus Deep Think. Users can also select an effort level (low, medium, high) per model, trading answer thoroughness against usage limits.
Google announced changes to Gemini Apps model availability on personal accounts taking effect October 9, 2026. Free users will be limited to Flash Lite, AI Plus subscribers get Flash added, and AI Pro/Ultra subscribers keep access to Pro plus Deep Think. Users can also select an effort level (low, medium, high) per model, trading answer thoroughness against usage limits.
Knowing the Rules, Applying the Rules: Evaluating Language Models on Traditional Chinese Bazi
A new arXiv paper introduces a 3,000-question Chinese multiple-choice benchmark on traditional Chinese Bazi, spanning 14 Theory and 11 Case categories, evaluating six endpoint systems. Theory accuracy exceeds Case accuracy for every system, with gaps of 16.60–29.56 percentage points, showing that knowing domain rules does not guarantee applying them to concrete cases. The results argue for task-specific evaluation of cultural-domain applications rather than reliance on aggregate knowledge scores.
A new arXiv paper introduces a 3,000-question Chinese multiple-choice benchmark on traditional Chinese Bazi, spanning 14 Theory and 11 Case categories, evaluating six endpoint systems. Theory accuracy exceeds Case accuracy for every system, with gaps of 16.60–29.56 percentage points, showing that knowing domain rules does not guarantee applying them to concrete cases. The results argue for task-specific evaluation of cultural-domain applications rather than reliance on aggregate knowledge scores.
Astra + Flash: orchestration package for running any LLM in Codex with a visual runs panel
An orchestration package that lets Codex users route coding tasks to models they already subscribe to, such as Claude, instead of only Codex models. It includes a visual panel showing runs and reviews. It targets developers hitting Codex usage limits who want to use their existing paid models inside the Codex workflow.
An orchestration package that lets Codex users route coding tasks to models they already subscribe to, such as Claude, instead of only Codex models. It includes a visual panel showing runs and reviews. It targets developers hitting Codex usage limits who want to use their existing paid models inside the Codex workflow.
Deep Research: Flash(+Extended) vs Pro(+Extended) — user asks when Flash-Extended is worth using over Pro-Extended
User asks how Deep Research differs between Flash(+Extended) and Pro(+Extended) modes
User asks how Deep Research differs between Flash(+Extended) and Pro(+Extended) modes
Gemini Pro users report silent downgrades to Flash after canceling auto-renewal
Gemini Pro subscribers report that Google silently routes their queries to lighter Flash models, especially after they cancel auto-renewal, to cut costs and serve the mass market. Users note that 3.6 Flash feels like an incremental evolution, while 3.1 Pro remains superior for reasoning.
Gemini Pro subscribers report that Google silently routes their queries to lighter Flash models, especially after they cancel auto-renewal, to cut costs and serve the mass market. Users note that 3.6 Flash feels like an incremental evolution, while 3.1 Pro remains superior for reasoning.
Users report Gemini forcibly switching to Flash Light despite available usage
Gemini users report the model forcibly switching to Flash Light even when far from usage limits, degrading response quality. The issue has persisted for days and affects paid users who cannot use the models they paid for.
Gemini users report the model forcibly switching to Flash Light even when far from usage limits, degrading response quality. The issue has persisted for days and affects paid users who cannot use the models they paid for.
Users compare costs of DeepSeek Flash and OpenAI Luna for coding
DeepSeek and OpenAI refreshed their mini-sized models (Flash and Luna), with DeepSWE scores around GPT-5.4. A Reddit user calculates real costs, finding Luna roughly 12-25x more expensive than Flash for similar token usage, with Luna at about $250 per 7 billion tokens versus DeepSeek's $75.
DeepSeek and OpenAI refreshed their mini-sized models (Flash and Luna), with DeepSWE scores around GPT-5.4. A Reddit user calculates real costs, finding Luna roughly 12-25x more expensive than Flash for similar token usage, with Luna at about $250 per 7 billion tokens versus DeepSeek's $75.