Quick Overview
In this commentary and review video, Matt Wolfe discusses several major AI model announcements, including Anthropic's Claude Fable 5.1, Google DeepMind's Gemini 3.8 Flash, and OpenAI's Astra preview. He examines benchmark performance, operational costs, and real-world coding demonstrations to assess the practical value of recent model releases.
Key Points
- 1.Claude Fable 5.1 achieves leading benchmark scores across reasoning and agentic coding, but its high inference cost makes standard production use expensive.
- 2.During tests in Claude Code, generating a functional 3D Megabonk clone game using Fable 5.1 quickly depleted monthly usage allowances and accumulated over 114 dollars in extra credits.
- 3.Gemini 3.8 Flash matches or approaches frontier-level coding benchmarks like DeepSWE while operating at a fraction of the cost and significantly faster execution speeds.
- 4.Anthropic introduced anti-distillation safeguards to prevent external developers from using Claude's chain-of-thought outputs to train competing student models.
- 5.OpenAI previewed its upcoming Astra model, which demonstrates advanced cybersecurity vulnerability exploitation but uses recurrent depth techniques that obscure visible human-readable reasoning chains.
Summary
Matt Wolfe breaks down recent announcements and releases in artificial intelligence, arguing that rapid incremental updates are creating fatigue while failing to substantially change everyday workflows for non-coders. He examines Anthropic's release of Claude Fable 5.1 and Mythos 5.1, Google DeepMind's Gemini 3.8 Flash, and OpenAI's teaser post regarding its forthcoming Astra frontier model.
Anthropic launched Claude Fable 5.1 for general paid users, while reserving Mythos 5.1 for vetted cybersecurity partners. Benchmarks released by Anthropic show Fable 5.1 roughly doubling prior model performance in agentic scientific research and showing significant gains in agentic coding. Anthropic also adjusted safeguards to reduce false-positive refusals on cybersecurity tasks by roughly 60 percent, while instituting anti-distillation measures to prevent competitors from using Claude's reasoning chains to train student models. Despite Anthropic claiming reduced overall costs through prompt caching, base token pricing remains unchanged at 10 dollars per million input tokens and 50 dollars per million output tokens. On the Artificial Analysis Intelligence Index, Fable 5.1 ranked first with a score of 66, but recorded an average task cost of 3.69 dollars, making it the most expensive model evaluated.
Wolfe tests Fable 5.1 in practical scenarios, beginning with BuseyBench, where it scored a winning 7.0 for an SVG portrait of Gary Busey, though processing took 18 minutes and cost 4.35 dollars. In a full coding test inside Claude Code, Wolfe used UltraCode mode to build a 3D browser clone of the game Megabonk in a single prompt. The resulting game featured polished 3D assets, enemies, upgrade menus, and controls, but running the build depleted the entire session limit of his 200-dollar monthly account and incurred an additional 114 dollars in usage credits.
In contrast, Wolfe highlights Gemini 3.8 Flash as a significantly underhyped model. On the DeepSWE software engineering benchmark, Gemini 3.8 Flash achieved a 73.7 percent pass rate, matching Claude Opus 5 and outperforming Fable 5 while costing 2.36 dollars per task compared to Fable 5's 9.18 dollars. Artificial Analysis data placed Gemini 3.8 Flash at 59 on overall intelligence, but with an average task cost of 58 cents and an execution time of 2.5 minutes per task. In BuseyBench, Gemini 3.8 Flash rendered a portrait in 1 minute and 32 seconds for 9 cents, and generated a playable Megabonk game from a single prompt.
Finally, Wolfe reviews OpenAI's Path to Astra announcement, which outlines a frontier model designated for advanced cybersecurity tasks. Benchmark data shows Astra achieving a 40 percent vulnerability exploitation rate on ExploitBench using 76,188 output tokens, outperforming GPT-5.6 Sol's 11.5 percent rate at 138,053 tokens. However, Wolfe discusses concerns reported by The Information regarding Astra's use of recurrent depth or looped transformers. Because this technique conducts iterative reasoning mathematically within internal layers rather than generating explicit natural language tokens, it conceals the chain-of-thought reasoning process from human auditors.
Claude Fable 5.1 Performance and Cost Breakdown
Anthropic released Claude Fable 5.1 for paid subscribers alongside Mythos 5.1 for restricted cybersecurity partners. While Fable 5.1 achieved top marks on benchmarks like the Artificial Analysis Intelligence Index with a score of 66, its pricing remained at 10 dollars per million input tokens and 50 dollars per million output tokens. On average tasks, Artificial Analysis measured its execution cost at 3.69 dollars per task, making it the most expensive model on their leaderboard.
Hands-on Coding and Benchmark Testing with Fable 5.1
Testing Fable 5.1 on BuseyBench produced a top-ranked vector portrait of Gary Busey, though the single generation took 18 minutes and cost 4.35 dollars. When tasked with coding a 3D browser game clone of Megabonk in Claude Code, Fable 5.1 produced detailed 3D assets, controls, and game logic, but consumed 100 percent of a 200-dollar monthly plan's session allowance and racked up more than 114 dollars in usage credits.
Gemini 3.8 Flash Cost-Efficiency and Speed
Google DeepMind's Gemini 3.8 Flash demonstrated high coding efficiency, scoring 73.7 percent on the DeepSWE benchmark, which places it on par with Claude Opus 5 and above Claude Fable 5. Artificial Analysis metrics revealed Gemini 3.8 Flash completed tasks in an average of 2.5 minutes at 0.58 dollars per task, providing near-frontier coding capabilities at a tenth of the cost of larger models.
OpenAI Astra Capabilities and Obscured Chain-of-Thought
OpenAI published research detailing its upcoming Astra model, which reaches up to a 40 percent success rate in identifying and exploiting software vulnerabilities while using fewer tokens than GPT-5.6 Sol. However, reporting indicates Astra employs recurrent depth, or looped transformers, processing text multiple times internally. This technique obscures visible step-by-step chain-of-thought reasoning, raising safety and auditing concerns among researchers.
The Bottom Line
The video establishes that while frontier releases like Claude Fable 5.1 set new benchmark records in coding and reasoning, their high token costs make sustained agentic workflows financially impractical for most individual users. At the same time, smaller models like Gemini 3.8 Flash deliver near-frontier coding performance at dramatically lower operating costs and faster runtimes. The analysis concludes with unresolved questions regarding OpenAI's upcoming Astra model, specifically how obscured internal reasoning mechanisms might complicate safety monitoring and alignment.
FAQ
What is Claude Fable 5.1 and what are its key capabilities?
Claude Fable 5.1 is Anthropic's flagship AI model available to paid subscribers, featuring state-of-the-art benchmark performance in agentic scientific research, knowledge work, and complex coding workflows.
How much does running a coding task in Claude Fable 5.1 cost?
Claude Fable 5.1 costs 10 dollars per million input tokens and 50 dollars per million output tokens, averaging approximately 3.69 dollars per task on the Artificial Analysis benchmark.
How does Gemini 3.8 Flash perform against Claude Opus 5 on the DeepSWE benchmark?
Gemini 3.8 Flash achieved a score of 73.7 percent on DeepSWE, effectively matching Claude Opus 5's 74 percent score while operating at a significantly lower average cost per task of 2.36 dollars compared to 11.84 dollars for Opus 5.
What anti-distillation mechanisms did Anthropic implement in Claude Fable 5.1?
Anthropic updated its API to prevent users from manually editing Claude's prior context while preserving thinking transcripts in multi-turn conversations, preventing competitors from extracting Claude's reasoning to train smaller models.
What safety concern is associated with the recurrent depth technique in OpenAI Astra?
Recurrent depth processes reasoning internally through repeated neural network passes rather than generating visible text tokens, which hides the step-by-step chain-of-thought and prevents humans from auditing the model's logic.
Worth watching for
Software developers, AI engineers, and technical leaders evaluating the cost-to-performance tradeoffs of newly released frontier and lightweight language models.
- claude-fable-5-1
- gemini-3-8-flash
- openai-astra
- ai-benchmarks
- coding-models
- deepswe