▶ YouTube Summaries

• https://www.youtube.com/watch?v=OTa_vK9919g
• Original title: He's always wrong

A reaction/review video walking through the release of Anthropic's new Sonnet 5 model (alongside news that Fable 5's export ban has been lifted). The core argument: Sonnet 5 is a surprisingly interesting model not because it benchmarks well or is cheap — it does neither — but because it is the first medium-sized model that genuinely understands agentic work, sub-agents, and orchestration. The catch is that it is not smart enough to wield those capabilities efficiently, so it burns huge numbers of tokens running in circles, ending up slower and more expensive than Opus or GPT-5.5 on real tasks. The takeaway is that Sonnet 5 is a model meant to be called by smarter orchestrator models, not used directly for day-to-day coding.

Summary

The video opens with two pieces of news for Claude users: a new model, Sonnet 5, and the lifting of the export ban on Fable 5. The Fable 5 ban (imposed June 12th over jailbreak/security-discovery capabilities that could aid foreign actors) was withdrawn after Anthropic agreed to proactively detect and address security risks, coordinate with the US government on release protocols, and report malicious activity. Crucially the withdrawal covers export and re-export, meaning API-hosting services (T3 Chat style) are allowed to serve the models to users.

The bulk of the video is a skeptical review of Sonnet 5. Anthropic pitches it as its most agentic Sonnet yet — able to plan, use browsers/terminals, and run autonomously at a level that recently required larger models — with performance close to Opus 4.8 at lower prices. Prime notes an overall industry tier-inflation: work that once went to Haiku now goes to Sonnet, Sonnet work to Opus, Opus work to Fable — a clever way to make everyone spend more.

On benchmarks the picture is unflattering. Terminal-bench improved meaningfully (high-60s to ~80%), knowledge work edged past Opus 4.8, and computer use improved slightly — though Anthropic has fallen behind Google there. But on Cursor Bench essentially every Sonnet 5 tier is more expensive and worse-performing than the equivalent GPT-5.5 tier, and on the artificial-analysis cost charts Sonnet 5 is the single most expensive model ever run through the bench (~$6,000 total, topping even Fable 5's ~$5,600), while GPT-5.5 medium/low cost a sixth to a twelfth as much. The "cheaper" claim from Anthropic refers only to per-token price ($2/$10 per million input/output introductory, reverting to $3/$15 after Aug 31), not real-world task cost.

The reason it is so expensive: it is extremely token-inefficient — using roughly 2x Opus's tokens and up to ~5x GPT-5.5's — and slow. Prime demonstrates by having three models rebuild his fish web game from scratch. Opus 4.8 finished in ~26 minutes with a solid, playable, well-balanced result. GLM-5.2 (~$8.30) produced choppy movement, a bad economy, and randomly-aimed shooting (GLM has no vision, so no browser use). Sonnet 5 took ~2–2.5 hours, produced a buggy mess (shooting fires on UI clicks, broken economy, no default fish), and — unprompted — spun up a whole cascade of sub-agents to investigate the old codebase, write and analyze a plan, implement it, and make a to-do list. It also asked good scoping questions up front, unlike Opus which asked none and spawned no sub-agents.

That sub-agent behavior is the whole point of the "5" in the name: Sonnet 5 exhibits orchestration behaviors previously seen only in Fable 5 — breaking work into pieces, handing it off, staying on task. The problem is it is not smart enough to do this well, so it over-decomposes work that shouldn't be split and runs in circles. Prime illustrates cost with a CEO analogy: a $100/hr senior engineer who solves a task in 10 hours ($1,000) versus a $20/hr junior who might match that but might take 100 hours ($2,000) if the task is beyond them. Sonnet 5 is the junior that keeps trying: it was built to "go until it gets an answer" even when it can't, which is exactly why runs get so expensive. This shifts engineering work toward helping the orchestrator (Fable) correctly route small tasks to Sonnet and hard tasks to Opus/Fable.

On his own SkateBench (naming skate tricks), Sonnet 5 XHigh scored the worst of any model he tests (37%) and wasn't cheap; Max scored better (59%) but was the most expensive model he's ever run on it (~15¢/question, some questions up to $1 — for wrong answers), jumping from ~1,600 to ~6,000 average tokens when moved from XHigh to Max. It also hit internal timeouts on Bun's fetch and got abandoned for GPT-5.5 medium.

Safety and reasoning are a mixed-to-bad story. Refusals on malicious requests rose slightly (92.3%), but the benign-but-suspicious pass rate dropped below 92% (down from Sonnet 4.6's 97%) — it now refuses legitimate work. A viewer-reported bug on claude.ai shows Sonnet 5 leaking its thinking traces, revealing that much of its "reasoning" is the model talking to itself about which tools it can and can't use, second-guessing whether to search the web, and agonizing over trivial questions (e.g., which MacBook color to pick) — clearly why it wastes tokens, and probably why labs hide traces (embarrassment, not just anti-distillation). Prime also calls out Anthropic for still bragging about SWE-bench numbers when it is effectively a contamination benchmark (recreating already-merged PRs), while their contamination discussion covers biology/AIME/USAMO/HLE/BrowseComp benches.

Final verdict: Sonnet 5 is not an Opus replacement for day-to-day coding. Its value is as a medium-sized model that finally understands agentic sub-agent orchestration and can play nicely on isolated tasks — a tool to be called by smarter models (Fable 5, Mythos 5, ideally GPT-5.6), ideally smart enough to know when to tap out and escalate to Opus/Fable. For now Prime is sticking with GPT-5.5 and Opus, and returning to Fable the moment he regains access. He's excited about this new era of models that break work into complex orchestrations, but thinks Sonnet 5 isn't smart enough yet to take advantage of the very capabilities that make it notable.