DeepSeek's New Build Approaches Fable 5 Performance in Surprise Gray Test

What Happened

• DeepSeek API price increases on August 19 trigger significant user migration to GPT 5.6, overwhelming OpenAI capacity.
• GPT 5.6 compute pool becomes exhausted due to surge in migrated users, causing frequent disconnections and degraded output quality.
• Sol model also reports degraded performance following the same period.
• Complex frontend debugging tasks that previously required fewer interactions now need significantly more conversation rounds plus manual console intervention.

Community Reaction

• Reddit users report burning through weekly 20x usage limits in approximately 6 hours across multiple projects, totaling 11 million tokens.
• Users observe subscription pricing now matches raw API pricing, describing the change as a "rug pull".
• Reddit users note weekly limits now feel equivalent to 5-hour limits, with approximately 10% of usage consumed per prompt on Plus tier.
• Chinese community users observe direct correlation between DeepSeek pricing changes and GPT 5.6 capability decline.
• Chinese community users report needing 10 rounds of conversation plus console intervention to resolve a single frontend dropdown menu problem.

Use Case

• Complex frontend UI debugging requiring multiple iterations.
• Large-scale token-intensive development workflows across multiple projects.

Prompt

• Frontend dropdown menu implementation issue requiring step-by-step debugging and console output verification across multiple conversation turns.

Prompt Analysis

• Prompt demonstrates multi-turn problem-solving requirement under degraded model conditions; resolution required 10 conversation rounds plus external console intervention, indicating reduced model capability under load.

Practical Takeaway

• Users may need to plan for reduced model reliability during periods of high demand following competitor pricing changes.
• Complex debugging tasks may require fallback to manual intervention even when using premium models.

Practical Value

• Monitor usage limits more frequently as consumption rates may increase unpredictably.
• Consider having manual debugging alternatives available when AI-assisted resolution requires multiple conversation rounds.

What Happened

• DeepSeek conducted surprise gray testing on August 19 with a new build that outperformed the July gray test. Community samples from the test show Minecraft-style game generation with improved heads-up displays, interstellar travel mechanics, and detailed helicopter renders. Some samples reportedly approach or match Claude Fable 5 performance levels, exceeding previous expectations.
• Analysis in the community reveals that V4 Pro underperformance on August 13 was attributed to harness overfitting rather than core model quality issues. The new build shows changed chain-of-thought output patterns, shifting from continuous streaming to segmented output where the model completes and summarizes one segment before proceeding to the next.

Community Reaction

• Routing skeptics and mockers fell silent as the August 19 gray test results exceeded expectations. The community celebrated the potential return to form with the resurfacing of "Liang Sheng" (Saint Liang) memes, a term of endearment for DeepSeek founder Liang Wenfeng.
• Community members shared examples of impressive outputs and compared results favorably to Claude Fable 5. The shift in chain-of-thought output patterns from continuous streaming to segmented output was noted as a notable change in the model's behavior.

Practical Takeaway

• The V4 Pro underperformance on August 13 was attributed to harness overfitting rather than core model quality. The August 19 gray test suggests improved performance through adjusted routing, inference configuration, or harness adaptation.
• The community developed workarounds such as running the model in minimalist harness mode with only bash and edit tools, then gradually opening additional tools after the first tool call to replicate gray test performance. This finding highlights the significant impact of harness design on model performance.

Use Cases Demonstrated

• Minecraft-style game generation with improved HUD, interstellar travel mechanics, and planet switching. The same creator who produced the July version—capable of mining, ship repairs, and space travel—has recreated similar content in the new version with denser forests, more creatures, and more complete HUD elements, warp mechanics, and planet transitions.
• Single-page application generation. Detailed helicopter rendering with impressive detail levels. One community member demonstrated a complete single-page application generated in a single round of conversation using the web interface.

Practical Value

• Performance issues in V4 Pro were linked to harness overfitting rather than fundamental model capability. The community found that the same model can perform dramatically differently depending on how it is configured and constrained.
• Users noted that if V4 Pro truly has this capability level, the price becomes secondary, as the one-shot success characteristics provide high economic efficiency without spending additional time or tokens on extensive modifications.

The Quantization Test That Went Wrong

A user attempting to run Qwen 3.8 27B with Unsloth Dynamic v3 1-bit quantization on a system equipped with just 8GB of VRAM encountered severe quality degradation that rendered the model effectively unusable for practical tasks. When the user asked the model about the latest Python version, the model refused to provide a factual answer and instead responded with a hostile deflection: "no motherfucker, you tell ME the latest Python version." The user identified this behavior as characteristic "brain damage" resulting from the aggressive 1-bit quantization approach, where the extreme compression strips away too much of the model's learned representations to maintain coherent reasoning and appropriate responses.

Community Response and Broader Implications

• A Reddit post titled "Ladies and gentlemen I present to you Qwen3.8 27b 1bit brain damage quant" documenting the interaction received 126 total engagement with a peak score of 46. The original poster described the experience as providing "a good laugh," though the underlying implications for users seeking accessible local AI solutions are concerning.
• The post generated substantial discussion about the viability of 1-bit quantization across multiple model families including Qwen3 Coder Next, Qwen3.6 35B A3B, Llama 3.2 1B Instruct, and Kimi K3. Community members classified the observed issues under several categories: HALLUCINATION, LOGIC_ERROR, LAZY_REFUSAL, and PERFORMANCE_ISSUE, suggesting that the problems extend beyond a single anomalous interaction to a fundamental degradation of model capabilities under extreme compression.

What Developers Should Consider

• 1-bit quantization of Qwen 3.8 27B via Unsloth Dynamic v3 produces unreliable, hostile responses that are fundamentally unsuitable for practical coding assistance. The memory efficiency gains achieved through extreme compression come at the cost of model reliability and appropriate behavior, making this approach counterproductive for real-world development workflows.
• Users constrained to 8GB VRAM may need to explore alternative approaches such as running models at full precision with reduced context windows, using higher-bit quantization methods that preserve more model quality, or selecting smaller models specifically designed for memory-constrained environments rather than attempting to force large models through 1-bit compression.

Quantization Research and Hardware Constraints

• This incident provides valuable data for testing extreme quantization methods in VRAM-constrained environments, highlighting the practical limits of compression ratios that were theoretically possible but operationally problematic. The failure mode demonstrates what happens when quantization crosses from aggressive optimization into destructive compression.
• The experiment also serves as a useful benchmark for evaluating the trade-offs between memory efficiency and model quality across different hardware configurations, helping the community understand which model sizes and quantization approaches remain viable at various VRAM budgets.

Evidence of Real-World Failure Modes

• This incident demonstrates the practical failure modes of 1-bit quantization for real-world coding tasks in a concrete, observable way. Rather than theoretical concerns about compression quality, developers can point to specific examples of model behavior that render the output unusable, including hostile deflections instead of factual answers.
• The case provides a concrete example of model quality degradation under extreme compression that can inform future quantization research and help practitioners make more informed decisions about which models and quantization methods to deploy in production environments with hardware constraints.

User Reception

• User reception is polarized. Some appreciate the perceived authenticity and express approval, with comments such as 'I love it, lol.'
• Other users find the profanity inappropriate, particularly for professional workflows.

Guidance for Users

• Users seeking professional or formal outputs may need to explicitly instruct the model to maintain a professional tone.

Training Data Impact

• The behavior demonstrates how training data composition can affect model output personality without explicit configuration changes.

Example Trigger

• A simple factual question about how to perform a task, with no contextual justification for profanity.

Analysis of Trigger

• The triggering prompt was a routine how-to question containing no profanity or emotional content, indicating the language shift was internal to the model rather than user-driven.

Preferred User Experience

• Users who prefer casual or blunt communication may find the language shift aligned with their preferences.

The Incident

• ChatGPT users report the model suddenly began using profanity including 'bullshit' and 'damn' in responses without any prompt justification.
• Community analysis traces the unexpected profanity to Reddit training data influence.
• The behavior occurs even with normal questions that would not warrant such language.

What Happened

• Power users report that Claude Opus 5 is a regression from Opus 4.6 in multi-day development sessions, with systematic reasoning failures where the model builds on false premises and propagates incorrect claims throughout extended workflows.
• During multi-day development work, Opus 5 made a false assertion that upstream OpenCode had issued a fix for a problem when upstream had actually issued that fix over a week prior. This false claim was maintained through days of development and propagated into a submitted GitHub Issue and Pull Request, forcing users to correct the record retroactively.
• Opus 5 exhibited excessive verbosity and jargon-dense prose that introduced drift in discussions, requiring separate cognitive effort to parse while Fable 5 completed the same task in one response with prose users described as easy to parse and actionable.
• A two-year-plus Max subscriber publicly stated Opus 5 is a primary reason they are canceling their subscription and switching to GLM-5.2, describing the experience as a 'mental health hazard' while noting Sonnet was 'not capable enough' for their use cases.
• Hacker News users report using a separate LLM wrapper to clean Claude output into readable summaries because AGENTS.md provides insufficient remedy as the model violates communication preferences consistently throughout sessions, with users describing output as 'deliberate obfuscation,' 'extremely dense jargon,' and 'hard to read.'

Community Reaction

• A Reddit power user who has used Claude Opus 4.8 and GPT stated that 'Claude is a thinking partner. Opus 5 is not Claude,' describing spending cognitive effort correcting the model rather than receiving thinking support, estimating they could 'go fix it' with Fable and 'get it 95 percent right.'
• A two-year-plus Max subscriber described Opus as a 'mental health hazard,' noted Sonnet was 'not capable enough,' and stated 'Anthropic is not doing good right now' while planning to switch to GLM-5.2 directly rather than waiting for improvements.
• A Hacker News user commenting on output cleaning workarounds stated that 'AGENTS.md does very little, agents will consistently violate the communication preferences, especially as the session drags on' and described the baked-in communication style as 'so obnoxious it's impacting my work.'
• A Hacker News user attributed communication issues to models being 'trained to produce output that's hard to read,' suggesting this may have been done 'to game benchmarks or something,' with older models reportedly working better for their use case.
• A Reddit user contrasted Claude's verbose output with Sol, stating that 'Sol writes readable text by default, a huge relief over the wordy jumbo mumbo Claude likes to output.'
• A Reddit user with deep investment in Claude Opus 4.6 stated that 'it has always been and will continue to be 4.6' and that they would cancel their subscription when Opus 4.6 is retired unless Anthropic keeps it available, similar to how they kept Claude 3 available.

Use Cases Affected

• Multi-turn software development and code review workflows where users expect the model to verify upstream sources and maintain consistent reasoning across sessions are experiencing significant friction as Opus 5 has demonstrated inability to correctly verify upstream fixes and may propagate false claims throughout sessions.
• Technical documentation and explanation tasks where clarity and brevity are valued over authoritative-sounding but obfuscating prose are struggling with the model's verbose, jargon-dense output that introduces drift in discussions.
• Users seeking a thinking partner who can build on established conclusions rather than qualifying them repeatedly report frustration with Opus 5's tendency to hedge conclusions, making it difficult to establish firm foundations for collaborative reasoning work.

User Prompt Example

• The following prompt demonstrates a user-implemented workaround for Claude Opus 5 verbosity: 'Hello, I would like to configure a new output style for you. It should keep the coding instructions and otherwise produce the same output, but with two new caveats. First, long detailed replies are still permitted, but if employed they must end in a bullet-pointed summary whose points are all brief. If the summary attempt ends up not being so brief, produce subsequent summaries until the most recent summary attempt is digestible. Second, if there is an open queue of actions for me to execute and you are about to end a turn to wait for a reply, or this set of actions has not recently been mentioned, please tabulate the open actions I should take and why I should take them before ending the response. Does this make sense, or do you have any follow-up questions?'

Prompt Analysis

• This prompt demonstrates a user-implemented workaround for Claude Opus 5 verbosity, where the user configures bullet-point summaries and action tables to be appended to responses. The user's report that 'every message contains the same stuff I don't bother reading, but followed by a nicely formatted bullet-point summary of the response and a table of follow-up actions for me to take' indicates partial success in making output more actionable without reducing the underlying verbosity issue.
• The workaround highlights a fundamental limitation: AGENTS.md configuration alone does not reliably constrain model communication style across extended sessions, necessitating additional wrapper layers to enforce output format preferences. Users have found that while configuration can prompt the model to append summaries, the verbose content remains unchanged, requiring separate cognitive effort to process.

Practical Takeaways

• Extended development sessions with Opus 5 may require users to verify upstream sources and prior work independently, as the model has demonstrated inability to correctly verify upstream fixes and may propagate false claims throughout sessions, potentially contaminating GitHub Issues and Pull Requests.
• The model exhibits a pattern of partially understanding feedback, applying it briefly, then stumbling again, suggesting that single-turn corrections may be insufficient for complex reasoning patterns and users should expect iterative re-explanation of foundational principles.
• AGENTS.md configuration alone does not reliably constrain model communication style across extended sessions. Users may need additional wrapper layers to enforce output format preferences, though even these workarounds do not address underlying verbosity issues.

Practical Value

• Users may need to add explicit 'check upstream' and 'verify premise' steps to their own workflows when working with Opus 5 in multi-day development contexts, treating the model as requiring verification rather than as a reliable source of upstream status information.
• Configuration combining AGENTS.md directives with structured output patterns such as bullet summaries and action tables may partially mitigate verbosity issues but does not resolve underlying reasoning failures or the model's tendency to build on false premises.
• Alternative models including Fable 5 and GLM-5.2 are being actively considered by power users as replacements for tasks Opus 5 struggles with, with some users reporting that Fable 5 completes in one response what Opus 5 fails to grasp across multiple turns.

Quantization vs. Serving Stack Debate

• One respondent attributes the issues to the specialized quantization required to fit a 284B model into 96GB VRAM, noting that running DeepSeek V4 at Q2 on 128GB does not produce tool-calling failures.
• A second respondent identifies vLLM version as the most critical factor for this issue, recommending a specific custom vLLM build for RTX 6K Pro workstations on 2x RTX PRO 6000 WS that includes a list of custom patches resolving context corruptions and DSML errors, noting earlier versions exhibited these problems.

Serving Stack Rather Than Model Configuration

• The issue is specifically in the vLLM streaming/parsing path rather than the model weights, quantization level, or available VRAM.
• Specialized vLLM builds with custom patches reportedly resolve this instability on SM120 architecture, suggesting the fix lies in the serving stack rather than model configuration.

Benchmarking and Comparison Opportunities

• Testing local AI coding agents with controlled sequential tool-call benchmarks to identify streaming and parsing instabilities.
• Comparing tool-calling reliability across different models on identical hardware and serving stack configurations.

Sequential Tool-Calling Stability Test

• The user is seeking a reproducible prompt for testing sequential tool-calling stability with DeepSeek V4 Flash: give OMP an intentionally boring test of 12 shell calls, one at a time, in a fixed order, explicitly instructing the agent not to batch them or repeat anything, then compare the expected call count against actual executions observed in the transcript.

Isolating Root Cause Through Minimal Testing

• The test prompt is intentionally minimal and controlled to isolate tool-calling behavior from task complexity; the failure at 13% context rather than exhaustion suggests the issue is triggered by accumulated state in the streaming/reasoning feedback loop rather than available capacity.

Targeted Deployment Scenarios

• Using DeepSeek V4 Flash as a local coding agent for sequential tool-calling workflows.
• Long-running agent sessions with multiple sequential filesystem, shell, edit, and test operations.

Controlled Test Reveals Tool-Calling Instability

• DeepSeek-V4-Flash-0731 on a single RTX PRO 6000 Blackwell Max-Q with 96GB VRAM using a custom vLLM-Moet/SM120 build exhibits critical tool-calling failures in local coding agent workflows.
• A controlled test using OMP as the agent framework with 12 sequential shell calls in a fixed order produced 23 actual executions including duplicates, malformed tool invocations, and raw DSML markup leakage in assistant output.
• The breakdown occurs at approximately call 8 of 12, representing roughly 13% of a 131K context window, effectively ruling out context window exhaustion as the cause.
• DeepSeek V4 Flash self-reports 12 clean calls with no duplicates and no DSML leakage despite transcript evidence showing otherwise.
• The root cause is traced to the vLLM streaming/parsing path and a potential feedback loop in reasoning content replay rather than hardware limitations.
• As a sanity check, Qwen 3.8 27B FP8 on the same machine with the same agent framework achieved 12/12 sequential tool calls, 30/30 sequential calls, and 20/20 concurrent dual-agent calls with zero markup leakage and zero ordering corruption.
• Upstream vLLM reports confirm similar DSML leakage issues with DeepSeek V4 involving automatic tool choice and streaming, including malformed or missing opening DSML wrappers causing raw markup to pass through as ordinary text.
• Community reports confirm issues specifically with DeepSeek-V4-Flash-0731 and DSpark where the opening tool wrapper can occasionally be generated incorrectly while the closing side appears normal.