▶ YouTube Summaries

• https://www.youtube.com/watch?v=KSV-7ywHxeU
• Original title: FABLE IS BACK! (And Sonnet 5 is here too)

Theo reacts to two pieces of Anthropic news: the export controls on Fable 5 (and Mythos 5) have been lifted by the Secretary of Commerce, so the model is coming back; and Sonnet 5 has launched. Most of the video is a skeptical, benchmark-heavy deep-dive on Sonnet 5. His verdict: despite the "5" branding and genuinely interesting sub-agent orchestration behavior, Sonnet 5 is slow, token-hungry, and often more expensive than Opus 48 or GPT 5.5 while performing worse. He sees it not as an Opus replacement for daily coding, but as a medium-sized model meant to be called by smarter orchestrator models (Fable/Mythos/GPT 5.6). For now he's sticking with GPT 5.5 and Opus, and returning to Fable the moment he regains access.

The Fable news is dispatched quickly. Fable was banned on June 12th, three days after release, over jailbreak/security concerns — specifically the government worried the model could find security issues in software and wanted to restrict that from foreign actors. Anthropic treated it as a hard ban. As of now the export controls are withdrawn: no license is required for export, re-export, or in-country transfer of the Mythos or Fable models, following commitments Anthropic made to proactively detect security risks and coordinate with the US government. Theo highlights that "re-export" being allowed matters for services that host these models over API and let users hit them directly (e.g. his own T3 Chat) — a concern he'd had that is now resolved.

On Sonnet 5, Anthropic pitches it as their most agentic Sonnet yet — able to plan, use browsers/terminals, and run autonomously at a level that recently required larger, more expensive models, at performance "close to Opus 48 but at lower prices." Theo argues Anthropic has quietly bumped every tier up: work that used to go to Haiku now goes to Sonnet, Sonnet's old work goes to Opus, and Opus's old work goes to Fable — a clever way to get users to spend more. On the benchmarks he's unconvinced: SWE-bench is heavily contaminated and near-meaningless; terminal-bench and "humanity's last exam" show real bumps; computer use is only a slight bump, and he notes Anthropic has gone from leading to lagging Google there. On Anthropic's own agentic-computer-use chart, Sonnet 5 is worse than Opus on both performance and cost, with Sonnet 5 Max costing more than Opus on high. Per Cursor Bench, essentially every Sonnet 5 tier is more expensive and worse than the GPT 5.5 equivalent — it "almost feels like Sonnet 5 was released to advertise how good a value GPT 5.5 is."

The cost story is the core of his critique. On Artificial Analysis it ranks 4th on the intelligence index (behind Fable 5, Opus, GPT 5.5), but cost-per-task is where it gets terrifying: GPT 5.5 on XHigh is cheaper per task than Sonnet by more than 2x, and Opus 48 is also cheaper. Sonnet 5 is the single most expensive model Artificial Analysis has ever run through the bench at ~$6,000 total, topping even Fable 5 (~$5,600); GPT 5.5 medium/low cost roughly a sixth and a twelfth as much. Anthropic's "cheaper" claim is only about price-per-token (introductory $2/$10 per million in/out until Aug 31, then reverting to $3/$15). It's a very inefficient model — using ~2x the tokens of Opus and up to ~5x GPT 5.5 — which also makes it slow. He removed the old Sonnet-specific subscription limit, but people report it eating large chunks of their usage even on the $200 plan.

His hands-on tests reinforce this. Rebuilding his "fish web game" from scratch: Opus 48 finished in ~26 minutes with a genuinely fun, well-balanced result; GLM52 (via open code, $8.30) produced choppy movement, a bad economy, and random shooting directions (GLM models have no vision, so no browser use); Sonnet 5 took ~2–2.5 hours, produced a buggy mess (shoots when you click UI, broken economy), and — unprompted — spun up a cascade of sub-agents to investigate, plan, analyze, and implement. On his SkateBench trick-naming benchmark, Sonnet 5 XHigh scored the worst of any model at 37% (Gemini 31 Pro leads at 95%); Max scored 59% but was by far the most expensive model he's ever run there (15¢/question average, some questions up to $1 — for wrong answers), jumping from ~1,600 to ~6,000 tokens average because it "runs in circles" when it can't solve something.

To explain how a Sonnet can cost more than an Opus, he offers a CEO analogy: a senior engineer at $100/hr who reliably solves a task in 10 hours ($1,000) versus a junior at $20/hr who might match that ($200) or might grind for 100 hours ($2,000). Sonnet isn't built to be expensive — it's built to "go until it gets an answer" even when it isn't smart enough, producing runaway costs. This, he argues, makes the engineer's job more interesting again: you must help orchestrators pick the cheap small model for small tasks and reserve Opus/Fable for hard ones — get the balance right and you save real money; get it wrong and Sonnet costs more than just using Opus for everything.

What genuinely excites him is the number "5": Sonnet 5 exhibits sub-agent orchestration behavior he'd previously only seen in Fable 5 — it breaks work into pieces, hands it off, stays on task, and asks good scoping questions (Opus asked none). The problem is it isn't smart enough to do this well and over-decomposes work that shouldn't be split. He envisions a future where Fable orchestrates and Sonnet sub-agents handle isolated tasks while knowing when to further subdivide or tap out to a smarter model.

He also flags safety and quality concerns. On Anthropic's refusal bench Sonnet 5 refuses 92.3% of malicious requests (2% higher than Mythos 5) but its success rate on benign-but-suspicious "dual use" requests dropped below 92% (Sonnet 46 was 97%) — meaning it'll refuse legitimate work it shouldn't. A viewer bug report shows Sonnet 5 on claude.ai leaking its thinking traces, and the traces are embarrassing: it spends enormous effort talking to itself about which tools it can and can't use rather than solving the problem (the phrase "let me" appears 21 times on one page), and dithers absurdly over trivial questions like which MacBook color to pick — which he suspects is a real reason labs hide reasoning traces. On contamination, he credits Anthropic for finally addressing it in the system card (using post-cutoff math olympiad questions, an eval blocklist for browse comp) but calls out that they still brag about SWE-bench numbers while SWE-bench is essentially a contamination bench (recreating already-merged PRs).

Final verdict: don't view Sonnet 5 as an Opus replacement for day-to-day coding. Its value is as a medium-sized model that understands agentic work and orchestration, to be called by smarter models once those exist (Fable 5, Mythos 5, hopefully GPT 5.6). His co-host Ben called it the best model until Fable returns, but hadn't yet seen completed work. Theo is unimpressed for now — sticking with GPT 5.5 and Opus, and moving back to Fable the instant he has access — while remaining hopeful about the broader era of models that can break work up in complex ways.