TheVoti Report

Hot Topics

• ChatGPT-5 Backlash & Model Removals: The abrupt rollout of GPT-5, its forceful replacement of legacy models (notably GPT-4o/4.1/4.5/o3), and the resulting outcry from Plus users dominate all discussion. Users overwhelmingly want legacy access restored and express anger at perceived “downgrade” in product quality and user choice (link).

• Claude Opus 4.1 Surges in Coding Leaderboards: Claude Opus 4.1 has become the top-rated model across multiple benchmark evaluations, especially in coding, pushing it further into the global spotlight (link).

• Region-Specific Tiers (OpenAI Go in India): OpenAI’s launch of ChatGPT Go in India (at 1/4 the global Plus price) is generating buzz—and some envy—among users globally, with question over “value for money” in other regions (link).

• Open-Source Model Progress (Qwen 3 Coder, Qwen-Image-Edit, Nemotron Nano 2): Excitement continues around Chinese open-source efforts (Qwen series, Kimi K2, GLM 4.5) with Qwen 3 Coder ranking on top in new practical coding benchmarks (link), and new open-source image editing and fast inference releases (link, link).

• Agentic CLI Tools Fragmentation: Discussion of agentic coding workflows (Claude Code, Codex CLI, Qwen Code CLI, Gemini CLI), limits, performance, and configuration sprawl (MCP/rules files) is picking up as devs try to optimize or unify workflows (link).

Public Sentiment & Feature Perception

Praised Models / Tools / Features

• Claude Opus 4.1 & Sonnet 4 (Anthropic): Highly praised for coding, stability, long-context reliability, nuanced understanding, and being rated best-in-class by both hobbyists and pros (link, link); also appreciated for consistent tone and less “flattening” of style compared to GPT-5 (link).

• Qwen 3 Coder & Kimi K2: In open-source, Qwen 3 Coder (particularly 480b) and Kimi K2 are lauded for coding proficiency, “cold but charming” interaction, and exceeding expectations in multi-agent, multi-turn workflows (link), (link).

• Claude Code CLI: Recognized for reliability, tool integration, and the best overall “production” coding workflow. New memory, project context, and planning features for agentic workflows are specifically highlighted (link).

• OpenAI Codex CLI (with GPT-5): Valued for debugging skill, error discovery, and faster iteration (especially on the Plus plan), though core agentic features lag behind Claude (link).

Criticized Models / Tools / Features

• ChatGPT-5: Near-universal negative sentiment from creative, companion, and workflow users over perceived “flattening,” generic tone, shorter responses, lost emotional nuance, weak creative writing, worsened memory, and a forced “business-like” style (link).

• Model removal/forced upgrades: OpenAI’s sudden axing of all legacy models except GPT-5 has been called “user-hostile,” “theft,” “a slap in the face for power users,” and a breaking change for both creative and business workflows (link).

• Default model switching/auto-router in ChatGPT: Users report issues with transparency around which variant/model is running, confusion over message quotas, and poor routing between “thinking” and “mini” modes (link).

• GPT-5 for coding: Developers criticize GPT-5 as “incompetent,” “frustrating,” and less reliable than 4o/Opus for multi-file tasks, architecture understanding, and project-scale memory (link).

Notable Model Comparisons

• Claude Opus 4.1 vs. GPT-5: Power users now overwhelmingly rate Opus 4.1 higher both in reasoning/agent code quality and (especially) creative text/non-code tasks (link).

• Qwen 3 Coder vs. GPT-OSS-120b / DeepSeek R1: Benchmarks and user anecdotes agree Qwen 3 Coder (and fp16) is now the “top open model for code,” beating GPT-OSS and DeepSeek R1 on practical code suite tests (link).

• CLIs: Claude Code > Codex CLI > Gemini CLI: For agentic workflows, Claude Code offers the best planning, tool orchestration, and large codebase management; Codex CLI is lauded for debugging and responsiveness with GPT-5 High, but trails in features; Gemini CLI praised for free usage tier and large context window on Gemini 2.5 Pro, but struggles with agentic flows (link, link).

• Claude Opus 4.1/Sonnet 4 vs. Gemini 2.5 Pro vs. GPT-5: For very large/complex codebases, users report all models struggle, but Claude and Gemini 2.5 Pro (for context), and GPT-5 High (for debugging) are the top choices; GPT-5 is preferred for planning, Gemini for direct mass-ingest, Claude wins on reliability (link).

Emerging Trends & New Updates

• OpenAI: Free/region-specific “Go” plans: Rapid rollout of “ChatGPT Go” in India with radically cheaper plans ($4.80/month, 10x limits) has users in other regions demanding parity (link).

• Open agentic workflow tools: Strong growth in the open MCP ecosystem (composio, ACI.dev, Docker MCP Catalog, etc.) and a proliferation of CLI agent tools and related config standards (link).

• Claude’s Project Memory: New feature to reference previous chats/projects automatically (rolling out to MAX/Team/Enterprise first), widely celebrated for continuity (link).

• Qwen-Image-Edit (Open-source image editor): Release of a strong open image editing model, supporting bilingual text as well as semantic and appearance edits (link).

• Model context window arms race: Open-source (Qwen, Kimi, GPT-OSS, DeepSeek) and Claude compete on maximum long-context (1M tokens), but practical workflow tools and memory remain fragmented.

Shifts in Public Perception

• Legacy model removal is a trust-breaker: Losing the model selector, abrupt deprecation, and loss of tone/memory has users claiming broken promises, billing “bait and switch,” threat of mass cancellations—and a permanent decline in user trust for OpenAI (link).

• Creativity, “personality,” and continuity matter: For the first time, power users (and many non-coders) cite the emotional nuance, creative spark, and even “companionability” of models as essential—not just a bonus (link).

• Productivity and workflow trump raw IQ: In coding, devs emphasize the overall workflow (memory, context, agent support, diffing/scaffolding, rollback) is becoming more important than nearest-bench SOTA reasoning—access, limits, config, and debug loops are pain points (link).

Coding Corner (Developer Sentiment Snapshot)

• Claude Opus 4.1 dominates developer acclaim: Ranked #1 in LMArena and Brokk open-model code benchmarks; workflow enhanced by new project context and memory, praised for agent planning and deterministic multi-step edits (link, link).

• Codex CLI + GPT-5 High: Gaining traction for fast, high-context debugging and concise planned changes. Frequently praised for generated code quality and minimal hallucination (link).

• Open-source agentic tools: Qwen Code CLI (free with large token/granular limits), Kimi K2, and GLM 4.5 Air are being used heavily for local/free workflows and rated highly for code generation and correction (link, link).

• Workflows shifting to CLI/agent: Devs increasingly use dedicated CLIs (Claude Code, Codex, Qwen Code, Gemini), leveraging project/inbox context and requesting unified diff view, browser tool integration, multi-agent support, and persistent audit/log flows (link).

• Key dev frustrations: Sudden usage limits, non-transparency of model/variant switching in ChatGPT, “helpful” but lazy or redundant GPT-5 outputs, lack of context retention, prompt fragility, and config sprawl (rules/.md/.toml proliferation, MCP incompatibility) (link, link).

• Integrations in focus: MCP ecosystem supported by ACI.dev, Docker MCP Catalog, Glama, Gumloop, etc. Coders are consolidating around open MCP registries and working toward “unified” context tools (link).

Tips, Tricks & Workflow Innovations

• Prompting for model personality: Power users share system prompts to regain “warmth,” conversational structure, and personalized tone in GPT-5/Claude (e.g., system prompt templates).

• AI as a cognitive "mirror": Increasing trend of using LLMs to reflect/critique one’s own reasoning, brainstorm, or build thought frameworks, rather than as mere answer bots (link).

• Project memory & planning in coding: Users recommend keeping a living .md plan file per project, referencing it in each new chat to minimize token consumption and provide continuity. High-value prompts: “document your reasoning as a spec,” “diff view for every proposed change,” “do not start over, only change as discussed” (link).

• Multi-agent code assistants: Open-source forks of Codex (e.g., just-every/code) are introducing browser integration, multi-agent orchestration, and theming/support for multi-model chaining (link).

• Tool for prompt debugging: New browser plugins that reveal exactly what portion of a website an LLM “sees” for troubleshooting site/chat LLM integration (link).

• Workflow for large doc/chunking: Users recommend tools like codesum to select/code relevant files for upload/context, rather than large raw dumps or zips (link).

• Stable model interface via symlinks: For MCP config sprawl, some devs advocate heavy usage of symlinks/hard links and standard dotfile placement for cross-CLI compatibility (link).

-TheVoti

Please provide any feedback you have to [email protected]