Reddit Rant Community Dubs Claude Opus 5 Output "Unintelligiblish"

User Frustrations

• Users are increasingly vocal about Codex resets serving as temporary band-aids rather than addressing root problems. One Reddit post framed the situation bluntly: 'Maybe resets are just compensation while they fix the real issues. Or maybe they're candy to keep us happy for a while.' The post sparked widespread agreement, with users noting that celebratory reset moments quickly give way to renewed complaints about slow performance and unpredictable limits.
• A user paying 235 euros per month for Codex access expressed particular frustration, stating: 'it's pure casino atm. Made me mad so often in the last weeks when ever they have banked they could give banked for paying 235 euro per month. there is literally no planning possible for those who actually use it to earn money with it.' The comment highlights how professional users dependent on reliable access find the current system untenable.
• Not all reactions were purely negative. Some users noted an unexpected benefit of premature resets: one user who had already used their banked reset the previous day woke to find the automatic reset had given them '40% usage for free that you didn't expect.' Others recommended following Tibo's X (Twitter) posts and third-party tracking sites to anticipate reset timing and preserve banked resets.

Monitoring Strategies

• Users recommend actively monitoring Tibo's announcements and third-party tracking sites to anticipate reset timing and preserve banked resets. One user who postponed using their banked reset after checking Tibo's hints managed to retain it until nearly 1% usage remained. The community suggests this proactive approach is currently the most reliable way to manage Codex access given the unpredictable reset schedule.

Trust and Subscription Models

• The Codex reset controversy offers insight into how temporary compensations affect user trust compared to addressing root causes. Professional subscribers expect reliability proportionate to their premium fees, and when compensation mechanisms feel arbitrary, trust erodes. The 'pure casino' sentiment reflects broader anxieties about subscription-based AI tools where users pay for consistent access but receive unpredictable service punctuated by sporadic fixes.

Creative Direction

• Compose a piece arguing that Codex resets function as temporary compensation rather than genuine solutions, using candy as an extended metaphor. The piece should contrast moments of user celebration when resets arrive with the underlying problems that remain unaddressed. Consider incorporating the perspective of professional users who depend on predictable access for their livelihoods.

Metaphorical Considerations

• The prompt calls for extending the 'candy' metaphor naturally rather than forcing it. The metaphor works because it captures both the temporary sweetness of resets and their potentially manipulative nature. Writers should explore how candy temporarily satisfies without nourishing, drawing parallels to how resets temporarily restore access without fixing systemic issues. The contrast between celebration and recurring problems provides fertile ground for examining user psychology and product management strategies.

Professional Value Debate

• The Codex reset discussion reveals ongoing community debate about subscription value for professional users. Those paying premium rates expect reliability commensurate with their investment, and when limits fluctuate unexpectedly, the value proposition comes into question. This case study illustrates tensions between AI providers managing infrastructure challenges and users who need consistent access to build workflows and earn income.

Recent Reset Events

OpenAI staff member Tibo announced a Codex reset scheduled for the following day. Community response was mixed: one user reported having already used their banked reset the previous day and was managing at 60% remaining when an unexpected reset was applied the following morning. Another user spent their banked reset only to wake and find an automatic reset had been applied while they slept. Some users report the 5-hour usage limit has decreased compared to one month prior, despite reset announcements designed to reassure subscribers.

Model Confirmation and Architecture

• GLM-5.3-Flash has been confirmed as an official Zhipu AI release, resolving earlier speculation that the model known as Ox Alpha might be a third-party fine-tune of open-source GLM. The model contains 320 billion total parameters with 18 billion active parameters, making it larger than DeepSeek-V4 Flash.
• The deployment has scaled to 700 trillion tokens processed on a domestic AI chip cluster comprising approximately 50,000 units. Zhipu AI developed a specialized inference engine using SGLang, implementing optimizations including quantization, tiered deployment, and compute-communication exchange to address limitations in domestic chip memory capacity and bandwidth.
• These optimizations achieved approximately threefold performance improvement in end-to-end service throughput on the same hardware. The architectural changes include a hybrid linear and sparse attention mechanism that reduces attention computation to roughly one-third and KV Cache to approximately one-quarter compared to GLM-5.3.

User Analysis and Comparisons

• The Zhihu community views the domestic chip deployment as a significant milestone for Chinese AI development, noting that domestic supply chains can now meet approximately 80% of AI inference needs. The performance-to-price ratio has been positively received, with some users declaring DeepSeek V4 as already surpassed by GLM-5.3-Flash pricing.
• However, detailed cost analysis by community members reveals nuance. Despite GLM's headline pricing advantage, DeepSeek V4 Flash remains cheaper at off-peak pricing due to superior cache economics. The measured cache hit rate of 92% on GLM-5.3-Flash compares unfavorably to DeepSeek's higher cache efficiency, reducing the effective pricing advantage below what headline rates suggest.
• Hacker News discussion confirms that consumer-grade hardware priced around $10,000 cannot run large models locally. Electricity costs of approximately $1 per day at 330W draw make current API rates competitive with local deployment, validating the API-first approach for most use cases.

Deployment Scenarios

• High-volume, cost-sensitive applications benefit most from GLM-5.3-Flash's sub-$0.10 per million output tokens pricing during promotional periods. The model's performance on code generation tasks (achieving 63% on DeepSWE with $0.24 per task cost versus DeepSeek-V4-Pro's $1.67 for comparable results) makes it suitable for automated coding pipelines and developer tooling.
• Deployment scenarios requiring compliance with Chinese semiconductor supply chain requirements benefit from the model's native support for domestic AI chip infrastructure. Organizations operating within regulatory frameworks that prefer or require domestic hardware can now access frontier-level model capabilities without relying on imported NVIDIA GPUs.

Cost Evaluation Framework

• Cache efficiency represents a critical factor when comparing effective API costs beyond headline rates. Models with higher cache hit rates can deliver lower effective per-token costs at scale despite seemingly higher base pricing, making technical architecture discussions essential for procurement decisions.
• Domestic AI chips have achieved sufficient performance for frontier model inference at cost efficiency approaching mainstream NVIDIA GPUs, at least for specific workloads. This development expands the viable hardware options for large-scale AI deployment in markets with regulatory preferences for domestic semiconductors.
• Off-peak pricing strategies significantly impact total cost of ownership for high-volume users. Organizations should evaluate pricing across different time periods rather than focusing solely on standard rates when planning budget-conscious AI deployments.

Quantifiable Comparisons

• The GLM-5.3-Flash pricing advantage translates to approximately 1/20 the effective cost of GLM-5.3 during current promotional periods, representing a significant opportunity for cost reduction in existing Zhipu AI workflows. This promotional rate enables high-volume applications previously constrained by API costs.
• The threefold hardware efficiency improvement achieved through SGLang optimization demonstrates a viable path for AI infrastructure relying on non-NVIDIA chipsets. Organizations planning hardware procurement can reference this achievement when evaluating alternatives to NVIDIA dominance.
• The electricity cost baseline of approximately $1 per day at 330W provides a concrete reference point for calculating return on investment between API consumption and local model deployment. At this electricity cost, API pricing must fall below approximately $30 per month per 330W equivalent to maintain cost parity.

Analysis Prompt

• Compare cache hit rates and effective per-token costs between GLM-5.3-Flash and DeepSeek V4 Flash for a 10 million token monthly workload, accounting for off-peak pricing tiers and cache efficiency differences.

Task Requirements

• This prompt requires analysis of multiple pricing tiers and cache efficiency metrics to determine effective cost rather than relying solely on headline rates. A factual comparison requires grounding in community-validated performance data, including the 92% cache hit rate measurement for GLM-5.3-Flash and DeepSeek's comparatively higher cache efficiency. The analysis should incorporate off-peak pricing periods where DeepSeek maintains cost advantages despite higher base rates. Users requesting such comparisons benefit from understanding that effective token costs depend heavily on usage patterns, cache behavior, and timing rather than published pricing alone.

User Frustration and Debate

• Users express strong frustration that this verbose style cannot be prompted away, with one stating there is "no way to prompt your way out of it."
• Community members debate whether 5.1 is a regression in raw intelligence, with some arguing this is acceptable if communication ability improves, stating "It's far more important for these models to be easy to work with. They're tools."
• Users argue the benchmark-maxxing is "incredibly obvious" for Opus 5, with the model suffering particularly in domains not typically benchmarked such as communication quality.
• Multiple users emphasize that most people use these models for work alongside them, not expecting the model to do everything independently, making clear communication critical.
• One user states the models are "simply not worth the cost at all" if they cannot relay what they have done while getting to the point without inventing unexplained jargon.

Key Findings for Users

• The verbose communication style appears to be a model-level behavior that cannot be reliably controlled through user prompts.
• Users attempting to use Claude Opus 5 for work-related tasks may face persistent communication quality issues regardless of prompt engineering attempts.

Work Tool Evaluation

• Claude Opus 5 is being evaluated by the community as a work tool, where communication quality and task completion are primary value metrics.
• Users compare models on communication quality, not just benchmark performance, when determining value.

Example Prompt Attempt

• You are a helpful assistant. Keep responses brief and to the point. Use simple language.

Why Prompts Fail

• Multiple users report attempting to reduce verbosity through prompt instructions without success, suggesting the verbose style is a fundamental model characteristic rather than a promptable preference.

Affected Use Cases

• Work assistance where clear, concise communication is required.
• Task completion where users need reliable summaries of what was done.

The Complaints

• Two Reddit threads converge on user complaints about Claude Opus 5's verbose, over-explanatory output style. Users have coined the term "Unintelligiblish" to describe the model's communication pattern of converting anything simple into extended, doctoral-thesis-style explanations.
• Users report the model generates clickbait-style sentences such as "We've discovered something new, and it completely changes how we think about <thing>" followed by mundane actions like "Running 1 shell command" for extended periods.
• Separate discussion notes the model frequently overstates potential problems—claiming it found numerous issues that will break a project—only to later backtrack that the issues have no impact at all.
• Complaints include Opus 5 leaving tasks half-done and the model taking on work not within its guidance without completing it.
• Users note the model invents internal jargon and codewords during its non-user-facing reasoning turns that it then uses in its responses without explaining them to users.

What happened

• Luna coding model systematically underperforms on complex tasks, exhibiting documented failure modes: subtly wrong implementations, stopping short of solving actual problems, finding workarounds instead of fixing underlying issues, and following plans superficially while missing important implications.
• Users report Luna never produces code worth keeping on non-trivial tasks, with one describing it as feeling like GitHub Copilot circa 2024.
• OpenAI release notes for GPT-5.6 only list Terra and Sol for coding; Luna is not included in coding-specific features.
• A community member attempted Luna for infrastructure tasks managing a home lab Proxmox cluster and it "kept screwing things up," subsequently switching to Terra+ for infrastructure work.
• When Sol catches errors in Luna's work, it spends significant time requesting corrections, leading to the conclusion that using Sol directly is more cost-efficient than correcting Luna's output.

Community reaction

• One community member describes Luna as "a slightly better GPT-5-mini which was useless," noting Luna is only acceptable as sub-agents for mundane task distribution when the main orchestration model handles all discovery, planning, and handoff.
• Users express frustration that attempts to "save tokens" with Luna consistently result in wishing they had used Sol from the beginning.
• One user reports their current workflow for long-running tasks uses Claude Opus 5 as the sweet spot, finding it accurate without over-engineering, with generous limits allowing 12-hour sessions for 10-15% of weekly allowance versus roughly 40%+ with Sol.
• A community member uses Sol for planning and Terra for implementation on their growing codebase, having not found a trustable use case for Luna beyond possibly document summarizing, which they also noted had issues.

Practical takeaway

• Luna should not be used as primary model for complex, non-trivial coding tasks or infrastructure management.
• Luna may only be viable as a sub-agent for tightly scoped, mundane tasks when the main orchestration model handles all discovery, planning, and handoff.
• Using Sol to correct Luna's errors is more costly than using Sol directly for code generation.
• Claude Opus 5 emerges as a preferred alternative for complex tasks, offering accuracy and generous limits.

Use case

• Sub-agent for tightly scoped, mundane task distribution (with main orchestration model handling discovery, planning, and handoff).
• Likely viable for small, tightly scoped changes where architecture and solutions are already decided.
• Document summarizing (with noted reservations from some users).

Practical value

• Luna should not replace Codex or Sol for autonomous complex coding tasks.
• Switching from Luna to Terra+ for infrastructure tasks improved outcomes.
• Claude Opus 5 offers a cost-effective sweet spot for complex, long-running coding sessions.

Prompt

• Luna coding model failure patterns: subtle wrong implementations, stopping short of solving problems, finding workarounds instead of fixing issues, following plans superficially.

Prompt analysis

• The documented failure modes indicate Luna struggles with multi-layered architectural changes, complex refactoring, and tasks requiring understanding of implications across a codebase.
• Users operating in large codebases with changes crossing multiple architectural layers report consistent failure with Luna.
• The evidence suggests Luna is optimized for simpler, isolated tasks rather than complex engineering work.

Community Reactions

• Reddit users expressed concern about unexpectedly high Claude usage costs. One Max 5x subscriber reported they felt no improvement after upgrading this week, while another noted their quota "basically vanished in an instant."
• Hacker News commenters praised Claude Design's approach to quota messaging and expressed surprise that Claude Code had not adopted similar functionality. Users suggested that AI coding tool interfaces should graphically display quota usage on screen at all times to prevent accidental overuse.

Practical Takeaways

• Users should monitor quota consumption closely after upgrading plans, as plan changes may not immediately reflect in usage patterns.
• AI coding tool default model settings can significantly impact quota consumption, making it worthwhile to verify and adjust these settings before extended use.

Practical Value

• This incident demonstrates the practical cost impact of default model settings on AI service quotas, showing how easily users can exceed their limits without realizing it.
• The situation highlights a notable gap in Claude Code's user experience compared to Claude Design regarding quota-related feedback and visibility.

Use Case

• Monitoring Claude session allowance consumption in real time to detect unexpected depletion patterns and identify which settings or models are driving high usage.
• Comparing quota visibility and display approaches across different AI coding assistants to inform better interface design decisions.

What Happened

• A Max 20x subscriber reported burning through 91% of their session allowance in roughly an hour while using Claude Opus 5 Ultracode, describing the pace as noticeably faster than their usual all-day usage pattern over the past month.
• A Hacker News reader accidentally consumed an entire week's Claude quota in 1-2 hours while using GPT-5.6 Sol on Codex's default settings, prompting them to build a quota tracking tool to investigate the cause.
• The quota tracking tool revealed the user's Claude allowance ran out in just 10 minutes under those conditions.
• Community members noted that while Claude Design displays cache expiry messages in the chat interface such as prompts to start a new chat to save tokens, Claude Code lacks comparable quota visibility features.