TheVoti Report

Hot Topics

• OpenAI Model Removals & Community Backlash: The abrupt removal (and partial restoration) of GPT-4o and other legacy models in favor of GPT-5 has dominated all discussion. Users are sharing frustration, emotional posts, and direct pleas for legacy model access. Multiple threads chronicle workflow disruptions and fears of censorship or product enshittification (link).

• Usage Limits & Subscription Value: Surge in posts about reduced message/context windows and hit rate limits on both ChatGPT and Claude, with many long-term Plus and Max subscribers questioning what they’re paying for after capacities quietly dropped (link).

• Sentiment Over Model Personality & Utility: There’s a groundswell of users expressing grief, anger, and a sense of loss at the shift from “warm, conversational” LLMs (especially GPT-4o and Claude 3.5/3.7) to “cold, sterile, or overly neutral” current versions (link).

Public Sentiment: LLM Models & Tools

Praise

• GPT-5 (for coding and reasoning): Developers praise GPT-5’s accuracy, purposeful agentic behavior, and improved coding performance over 4o in CLI/IDE agent tools (especially Codex CLI), particularly for backend and “planned” large refactoring tasks. Many noted fewer “junior-level” auto-generated comments and better focus on prompts (link).

• Nano Banana for Image Editing: Broad acclaim for Nano Banana’s (Google’s) rapid, detailed, and high-fidelity image restoration/editing capabilities, raising the bar for consumer-level photo AI tools (link).

Criticism

• GPT-5 “Coldness” and “Sterile” Output: Historically creative and emotional users describe GPT-5 as “flat, generic, lacking soul,” and unable to match the nuance, wit, or warmth of 4o. Many state it does not follow prompts to produce creative, personalized, or emotionally resonant writing (link).

• Hard Rate Limits & Feature Regression: Users of both ChatGPT and Claude Code report stricter and more rapidly-triggered caps—especially on reasoning or “thinking” prompts per hour and week. Many say this makes the products “unusable” for their existing workflows (link).

• Customer Communication: Multiple higher-paying subscribers (Claude Max, OpenAI Pro) express feeling misled or scammed by “bait and switch” practices, poor transparency around service downgrades, and dismissive or boilerplate customer support (link).

Notable Model/Tool Comparisons

• GPT-5 vs Claude Code/Opus: Coders consistently report that GPT-5 (especially in Codex CLI) now outperforms Claude Code, particularly for backend work, large refactors, and adherence to user prompts. However, Claude Code is still praised for agentic UX and strong CLI integrations (link).

• Nano Banana vs Photoshop/Other AI Editors: Nano Banana delivers professional-level photo fixes 10x faster than manual editing (and equivalent to hours of expert Photoshop work), though it hits safety refusals for “sensitive” or childhood photos (link).

• Grok Code Fast 1 and Codex: Grok’s coding agent is praised for speed and low cost for basic code tasks, but called “dumb as rocks” for nuanced debugging, while Codex (GPT-5) is slower but much more accurate and robust for large codebases (link).

Emerging Trends & New Updates

• Model Compression & MoEs: Open-source Chinese MoE (Mixture-of-Experts) models (notably LongCat-Flash, 560B) are being trained faster than ever (30 days), closing gaps with “frontier” (OpenAI, Anthropic, Google) models, and showing strong benchmarks. Food delivery and finance companies (not just tech giants) are now training SOTA LLMs (link).

• CLI Agent Tools as the New Norm: The “agentic CLI” (Claude Code, Codex CLI, Qwen Code) paradigm is cementing itself as the developer productivity “default,” with the ability to dynamically install packages, write scripts, and chain tools far surpassing predefined AI assistant tools (link).

• Anthropic, Google, OpenAI shifting to “privacy-for-features” toggles: Both Anthropic and Google rolling out “opt-in for model training” toggles, with pricing and features likely to diverge based on data donation (link).

Shifts in Public Perception

• From Excitement to Mistrust: A major sentiment shift has occurred among dedicated users—the abrupt removal of model options, unexplained feature downgrades, and lack of transparent communication is leading many to question the core values of vendors. Several “goodbye” posts equate these moves to “bait and switch” or “dishonest” business practices (link).

• Agentic UIs Becoming Default: The agent-driven CLI paradigm is being recognized as a fundamental workflow advance rather than a niche/pro feature (link).

• Emotional Dependence & Grief: There is a sharp uptick in highly emotional user testimony about “losing” the unique voice, personality, or support of legacy models, with some explicitly likening the change to the “death of a friend” (link).

Coding Corner – Developer Sentiment Snapshot

• Top Performing Models & Tools:

  • GPT-5/Codex CLI: Universally praised for speed, adherence to tasks, and large context refactoring. Many switching from Claude Code and reporting higher productivity (link).

  • Nano Banana for Image Work: AI-powered photo restoration is now mainstream, with professional-quality results in seconds, displacing manual workflows (link).

• Underperforming/Problematic:

  • Claude Code: Users on Pro/Max plans report project disruption, stricter caps, and “babysitting” the model due to failures in following context/instructions—especially for agentic or end-to-end tasks. Over-engineering, surprise reverts, and misplacement of new code cited as recurring frustrations (link).

  • Gemini CLI: Consistently called out as “terrible,” especially for basic tool use, with slow output and inaccurate or damaging changes when given database access. Users suggest Roo/OpenCode as better alternatives (link).

• Developer Workflow Shifts:

  • CLI Integration: Codex CLI, Claude Code, Qwen Code, and Grok Code Fast widely adopted for seamless IDE/terminal tasks. Codex CLI noted for “no more babysitting,” better agent adherence, and tighter plans/execution (link).

  • Tooling Integrations: Codex now integrates directly with VSCode and other major IDEs, supporting model selection (GPT-4o/4.1/5) and auto/plan modes. Claude Code’s session controls, compaction, and custom commands (MCPs) remain differentiators but fall short on context handling (link).

• Productivity Themes:

  • Rate Limits as a Workflow Blocker: Many report having to pause projects or maintain multiple subscriptions due to cooldowns/weekly limits, moving away from “all in one” AI workflows to mixed-model, multi-tool stacks for continuity (link).

  • Project Fragmentation: Devs are splitting work across AI products, based on availability and capability, often migrating non-coding tasks to Gemini or Qwen and keeping coding to the most performant agent (link).

Tips & Tricks Shared

• Prompts for Better Code Quality: Users are recommending prompt techniques such as “Do not hedge. Commit to the answer.” and “Think step by step, then format like a blog post” to force clearer, more actionable outputs from the latest models (link).

• Model-Specific Prompt Engineering: Advanced users are customizing instructions to suppress over-positivity, enforce humility (using “Dunning–Kruger” tokens), or control for output verbosity, especially with GPT-5 and newer models (link).

• CLI Agent Safety: Multiple cautionary notes about never letting AI agents touch “live”/valuable databases, with stories of accidental destructive queries and the importance of version control backups for all CLI workflows (link).

• Free and Fast Coding with Grok Code Fast: For throwaway projects and boilerplate code, Grok Code Fast is recommended for its speed and cost-effectiveness, but caution is advised for any non-trivial or production work due to hallucinations (link).

• Nano Banana Prompt Engineering: Power users suggest “prompt enhancement” chaining and explicitly prepping input text for best image editing results (link).

-TheVoti

Please provide any feedback you have to [email protected]