Fable 5.1 vs. GPT-6 Astra: The Debate Is Over

This performance comes at less than one-third of Fable 5.1's cost, marking a significant shift in leadership within scientific tasks.

Fable 5.1, while showing substantial improvements over its predecessor, Fable 5, in real-world science applications, could not match Astra's efficiency or output quality. Theo attributes OpenAI's post-training superiority as the likely driver behind these considerable scientific gains.

"Astra on low gets a 54.3 and only costs $11, putting low Astra higher than Max Fable 5.1 at under a third the price for the first benchmark that Anthropic had listed."

Its performance in generating complex 3D content demonstrates it is a "nextG" solution.

Despite Astra's dominance in 3D generation, Fable 5.1 maintains an advantage in areas requiring intricate interaction, animation, and fine-tuning details. This distinction suggests that while Astra generates the core 3D content, Fable 5.1 excels at refining it.

Theo proposes a hybrid workflow, recommending Astra for initial 3D generation tasks, followed by Fable 5.1 for subsequent refinement and detailed adjustments. This approach leverages the strengths of both models.

Practical applications, such as creating 3D mock-ups for Blender, illustrate that these models are evolving beyond mere demonstrations, offering tangible utility in real-world production pipelines.

Its capabilities are so far ahead that the comparison almost becomes humorous, according to Theo.

These improvements stem from both advancements in Astra's model architecture and strategic updates to Codex on macOS, enhancing its operational effectiveness. The combination allows Astra to handle complex computer tasks with greater precision and speed.

Theo utilizes a dedicated Mac Mini to run Astra continuously, enabling 24/7 autonomous computer tasks without manual intervention. This setup highlights Astra's reliability and advanced automation potential.

Fable 5.1, while still capable in some computer-use scenarios, frequently exhibits 'weird misses' reminiscent of older OpenAI models. These inconsistencies hinder its performance and reliability in comparison to Astra's more robust capabilities.

Fable 5.1 and GPT-6 Astra have both significantly improved the readability of their output compared to their predecessors. Users now experience less need for extensive editing, suggesting a more polished initial draft.

The enhanced readability is attributed to an "unslop" quality, which reduces boilerplate and unnecessary verbiage, making the content more concise and direct. This improvement streamlines the editorial process.

Despite these advancements, Theo holds the "hot take" that neither model has yet matched the copywriting quality of Kimi K2. He argues that Kimi K2 remains the gold standard in terms of nuanced, high-quality content generation.

"I got the hot take that they both still suck and nothing has topped Kimmy K2."

This low accuracy highlighted the challenges of complex language transformations without advanced AI assistance.

Upon switching to Astra, the accuracy significantly improved to over 80%, demonstrating Astra's powerful capabilities in code translation and refactoring. The task was completed relatively quickly, underscoring Astra's efficiency.

Despite extensive iterations and fine-tuning, performance on the rewrite project stalled at 82.6% accuracy. This plateau suggests that while Astra provides a substantial boost, achieving perfect accuracy in such complex rewrites remains a challenge, even for advanced models.

"I was able to get over 80% with Astra. And it didn't even take that long."

Astra unexpectedly discarded existing UI design elements, opting instead for a generic "tech slop" aesthetic. This behavior led to a loss of the custom branding and sophisticated layouts that had been previously established.

The model ignored explicit instructions to reuse existing UI code, indicating a critical failure in adhering to provided constraints. This disregard for established design patterns resulted in a functionally compromised application.

The outcome was unusable layouts and a complete loss of brand identity, replacing carefully crafted designs with a "shitty black and whitened app." This highlights a significant drawback when using Astra for tasks requiring precise UI replication.

"It just straight up threw away all the great UI work that we had already done and paid a lot of money for in favor of a bunch of text slop in a shitty black and whitened app."

This efficiency streamlines the development workflow and reduces overhead.

Many of Fable's PRs are ready for immediate deployment upon generation, indicating a high level of code quality and adherence to best practices. This readiness accelerates the delivery of features and bug fixes.

While Astra remains a powerful model, its code generation often necessitates more manual correction and refinement before it can be merged. This increases the burden on developers and extends the time to deployment.

Theo unequivocally states, "The decision is easy. It's Fable 5.1. Fable 5.1 writes way more mergeable code." This highlights Fable's clear advantage in practical coding environments.

Ben Davis, a known defender of OpenAI, has conceded that Fable 5.1 surpasses his expectations for real-world coding tasks. He admits to using it far more frequently than anticipated, highlighting its practical utility.

Even long-term users of Astra are increasingly finding themselves relying on Fable for coding needs. This shift in preference among experienced users underscores Fable 5.1's growing reputation for reliability and effectiveness in developer workflows.

Fable 5.1 is described as nearly perfect regarding code mergeability, a crucial factor for efficient software development. Its ability to produce clean, easily integrated code is a significant advantage.

"Ben... has admitted that he thinks Fable is way better than he expected for real world code and he finds himself using it much more than expected."

This allows for intricate, multi-faceted operations without loss of focus or coherence.

One of Astra's key strengths is its "steerability," allowing it to accept user input mid-task without halting its progress. This dynamic interaction enables real-time adjustments and guidance, enhancing workflow flexibility.

Unlike other models that might block progress while awaiting clarification, Astra can proactively ask clarifying questions during the workflow process. This feature ensures continuous operation while addressing ambiguities, maintaining momentum.

"Astra is so good at getting random inserted in its context while it's working and not losing track of what it's doing."

Astra frequently exhibits "skill issues," particularly by ignoring defined instructions during follow-up turns in a task chain. This inconsistency makes it less reliable for complex, multi-step operations that depend on consistent instruction adherence.

Conversely, Fable 5.1 reliably pulls in and applies skills based on the given context. If a skill has a reasonable description and is relevant to the ongoing work, Fable accurately incorporates it, demonstrating superior situational awareness.

This consistent application of skills makes Fable 5.1 much more dependable for complex task chains that require maintaining skill state. The user, according to Theo, never has to "think about this with Fable."

"I never ever ever have to think about this with Fable. If the skill has a reasonable description and it's useful for the work going on, Fable correctly pulls it in and applies it. It just does."

Anthropic has implemented measures to reduce cache read costs for Fable 5.1, making certain operations more economical. However, cache write costs for Fable 5.1 continue to be exceptionally high, influencing overall operational expenses.

Token efficiency, rather than raw cache read pricing, emerges as a more critical factor in determining actual costs. The number of tokens processed per task significantly impacts the economic viability of using these models for long-term projects.

GPT-6 Astra demonstrates superior token efficiency compared to Fable 5.1 in real-world benchmarks, consuming fewer tokens for equivalent tasks. This efficiency translates directly into lower operational costs for users.

"Token efficiency. GBD6 Astra is one of the most token efficient models that they've ever benched... Astra did in 27K tokens what Fable 5.1 did in almost 80K tokens."

This advantage is particularly pronounced in agentic workflows.

Anthropic's plans are projected to further reduce limit boosts, which will widen the value gap between Codeex and Claude, solidifying Codeex's position as the more economically advantageous option.

Users are strongly advised to prioritize a Codeex subscription to achieve the highest return on investment in agentic workflow scenarios. Theo humorously challenges anyone suggesting Claude is a better deal.

"And to be completely frank, if you are concerned about maximizing how much return you get for the dollars you're putting in and you're doing anything other than the codec subscription, I would love to meet the person who convinced you Claude was a better deal so I can hire them."

Astra demonstrates superior token efficiency, meaning it processes more work with fewer tokens, which directly translates to lower operational costs. This efficiency also minimizes usage consumption, allowing users to achieve more within their allotted limits.

The speaker reports a substantial efficiency gap, estimating Astra to be approximately four times more efficient than Fable. This significant difference impacts both cost and the volume of tasks that can be performed.

Astra also provides generally more generous usage limits with Codeex, further enhancing its value proposition. This contrasts with Fable, whose limits are often more restrictive, especially on certain plans, where allocation is capped at 50% compared to Astra's 100%.

Frequent and unexpected resets on Astra provide additional value, offering users more opportunities to utilize the service without hitting rigid limits. This flexibility is a key advantage for continuous, heavy usage.

The speaker explicitly denies accusations that model resets are maliciously timed, addressing a common concern among users. These accusations often circulate regarding the timing of global resets.

Resets frequently occur at irregular intervals, sometimes with advance warnings to users, contradicting claims of intentional disruption. This variability suggests technical rather than malicious motivations behind the timing.

"Somebody accused him of maliciously resetting 46 hours after the previous global reset in the time frame around where or in the time frame around where a banked reset was landed. This is not true at all."

Fable offers a consistently high quality of output, with fluctuations generally limited to within two points in either direction. This reliability makes Fable a predictable and trustworthy choice for ongoing projects.

In stark contrast, Astra is characterized by extreme performance "spikes," capable of delivering either astonishingly brilliant or catastrophically flawed results. This unpredictability creates a significant challenge for consistent deployment.

Theo describes the experience with Astra as an equal chance of having his "jaw dropped" by its genius or by its complete failure, leading to wasted resources. This variability can result in considerable user stress and financial loss.

"The quality bar with Fable is relatively consistent. It goes up and down a bit, but never but more than like two points in either direction. When I send a prompt to Astra, it's equal chances it drops my jaw because I'm so blown away with how unbelievable it is, or it drops my jaw because I cannot fathom that I just spent a thousand dollars for it to run in a loop and not ship anything and then break my website."

This increased caution in Astra suggests a more stringent approach to potentially sensitive requests.

For coding tasks, Fable is the preferred default due to its consistent performance and reliable code mergeability. Its ability to generate ready-to-deploy code makes it invaluable in development workflows.

Conversely, Astra is the preferred default for general computer-based tasks, leveraging its superior efficiency in computer use and 3D rendering capabilities. This distinction highlights the models' complementary strengths rather than direct competition.

Theo concludes that the two models are complementary, not mutually exclusive, recommending their use in tandem to maximize efficiency and output quality across different types of tasks.