Testing GPT Astra vs Fable 5.1
Want to back the stream? The «داعم» membership on YouTube is EGP 60 a month: badges, emoji, and I answer your questions first on stream. Become a supporter
The gist
The stream centers on a practical, partly blind comparison of GPT Astra and Fable 5.1 using matched prompts for games, an SVG drawing, landing pages, and agentic tasks. Results were mixed: Astra often appeared stronger visually and in the controls of some games, while Fable sometimes delivered more challenging gameplay or more inventive design ideas. Speed varied by task and audience preferences split, so no clear overall winner emerged. Luqman discusses the limits of single tests, benchmarks, and model judges, arguing for evaluation grounded in domain expertise, bias reduction, critical-thinking and ideation tests, alongside efficiency, price, usage limits, and tool integration. The discussion also covers routing planning and execution to different models, sub-agent orchestration, documenting and compacting long coding sessions, preconfigured bots, video automation, backups, and changing developer roles. Claims about AGI, job displacement, and near-term end-to-end AI development remain attributed predictions or concerns rather than established conclusions.
104 moments from the stream
- 1:45:41
Testing Agentic Capability with a Landing-Page Build
He uses a test koshary brand and landing-page build as a practical probe of agentic capability. He tracks completion time and delivery format, notes that one result supplied code without serving it locally, and keeps the Astra and Fable outputs unidentified to gather blind audience opinions.
- 1:52:04
Comparing Astra and Fable 5.1 on a Koshari Landing Page
Both models received the same prompt to build a koshari restaurant landing page with motion effects at high effort. The results were compared section by section—including typography, menus, add-ons, ordering links, and animation. The host felt the Fable result on the left showed more creativity in revealing the dish’s ingredients, while some viewers preferred the design on the right. He also said their relative speed varied by task, so he reached no firm conclusion about which was faster.
- 1:59:34
Creating a Short Koshari Video with an Agentic Higgsfield Workflow
The host explained that two short koshari videos were created through Higgsfield, which provides image and video generation using multiple models. Astra and Fable received the same task and freedom to choose a generation model, with instructions to make a visual loop that begins and ends on the same image, setting up a comparison of their outputs.
- 2:01:47
How Tool Integration Affects Agent Workflow Speed
Luqman clarified that this test compared the agents and their workflows, not the video generator itself. In his run, Codex finished much faster, which he attributed to having a built-in integration with the video tool, while Claude had to be connected through a command-line interface. Because the requested clip was short, the comparison remains limited to this task and setup.
- 2:03:27
Comparing Racing Games Generated by GPT Astra and Fable 5.1
Luqman tested two racing-game implementations built from the same prompt, one by GPT Astra and the other by Fable 5.1. He found Astra's version markedly better in driving control and gameplay physics, including a boost mechanic, while Fable's looked more visually appealing and offered extra options but was harder to drive because of its sensitivity, physics, and performance. He withdrew his initial declaration of a clear winner, suggesting that improved physics could make Fable the stronger result, and also noted differences in UX and settings.
- 2:11:34
Could Agent Teams Replace Some Jobs?
The speaker speculated that agent systems modeling organizational charts and executive roles could replace some employees, presenting this as a concerning possibility rather than an established outcome.
- 2:12:23
Efficiency vs Familiarity When Evaluating AI Models
After the tests, the speaker’s view was that the more efficient model is likelier to prevail. He also acknowledged that long-term familiarity with a tool can influence judgment and that warning signs should not be ignored.
- 2:13:34
Agent Teams for Operations and Technical Support
The speaker discussed, conditional on the system working as expected, forming agent teams to monitor applications and address user problems, with possible uses in technical support, customer service, and marketing.
- 2:15:53
Can Product Design Create Attachment to an AI Tool?
The discussion considered whether color, UX, and marketing can foster attachment to a tool. The speaker attributed his hesitation mainly to long experience and accumulated practical reliance, rather than design alone.
- 2:16:50
Emerging Engineering Roles Around AI Agents
The speaker recommended watching specialties emerging around AI agents, mentioning AI engineering, roles resembling forward-deployed engineering, and graph engineering, while noting that new fields are emerging in the market.
- 2:19:09
Choosing Different Models for Different Tasks
The discussion compared speed, work completed, and usage limits, concluding that different tools may suit different use cases. The practical view was to use more than one model when feasible and route each task to the best fit; a model-routing service was mentioned but not clearly identified.
- 2:21:31
Splitting Planning and Execution Across Models
The discussion notes fatigue from repeatedly switching models as benchmark results change, then suggests using a stronger model for planning and debugging and a smaller or more efficient one for execution. These were proposed workflows and participant opinions, not a conclusive test result.
- 2:23:54
Practical Criteria Beyond Benchmarks
The speaker prioritizes critical thinking, ideation, and thinking outside the box, while noting that capability alone is insufficient and efficiency also matters. The assessment is preliminary and preference-based, not objective proof that one model is superior.
- 2:25:32
A Qualitative Test Using a Traffic-Growth Audit
The speaker frames a test requiring research, analysis of an agent system and its instructions, identification of gaps, and new ideas for increasing organic traffic. He says comparing the responses will be difficult and subjective because it requires domain knowledge and close reading; no final result appears on this page.
- 2:30:22
Is Context-Window Size Still the Key Metric?
Responding to an audience question, the speaker tentatively argued that raw context-window size was becoming less central as context-management strategies improved. At the time of the recording, he cited Codex hiding context-window details by default. The answer is cut off at the page boundary, so the full explanation is unavailable.
- 2:31:16
Is Context-Window Size Still the Main Factor?
Luqman argues, as a personal observation rather than a conclusive test, that automatic compaction may be making raw context-window size less important. He notes conflicting views on whether a larger window is always better and says performance is difficult to assess without clear benchmarks.
- 2:32:24
Context Degradation When a Session Changes Topics
The discussion considers how a model may lose coherence as context grows or when one session shifts across multiple topics. Luqman says the effect feels less noticeable to him, while the group still identifies topic switching as a weakness and briefly mentions Obsidian as a possible way to organize external context.
- 2:35:46
A Practical Workflow for Compacting a Long Coding Session
Luqman describes a workflow he uses as a long session advances: first ask for a concise handoff that preserves important state and decisions, then run the compaction command and reintroduce those instructions to continue. He presents it as a personal technique, not a guarantee against context loss.
- 2:39:20
When Context Compaction Becomes Risky for Coding Tasks
The group explains that lost details can matter more in coding, where forgetting a name or small decision may cause larger errors, especially if compaction occurs mid-task. Documentation files and splitting each feature or transition into a smaller session are suggested as mitigations, while emphasizing that the impact depends on the task and the quality of the compaction summary.
- 2:42:13
Setting Up a 3D Koshary-Delivery Game Test
Luqman sets up a playful test asking for a 3D game about delivering koshary by scooter, identifying Astra on the left and Fable on the right before submitting the prompt. This segment records the comparison setup only; it does not provide a final result or quality verdict.
Got a question about something in this stream?
Ask in the group