▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=06BvFMW8Ng8
• Original title :: Which AI Models Are Worth Using

Theo ranks every model he would reasonably reach for today. He says up front that a tier list is a bad way to compare models - too many axes: task fit, price, token efficiency, speed - and does it anyway. The recurring argument underneath the rankings is that price per token is nearly meaningless; what matters is tokens consumed per task, which is why some "cheap" models end up the most expensive ones on the board.

The board

• S :: Fable 5.
• A :: 5.6 Soul, Luna.
• B :: Deep Seek V4 Flash, Kimmy K3.
• C :: 5.6 Terra, GLM 5.3 (later dropped to D for lacking vision).
• D :: Composer 2.5, Grok 4.6, Muse Spark, Opus 5, Sonnet 5 (only inside a Claude Code sub).
• F/"Google tier" :: Deep Seek V4 Pro (no vision), Gemini 3.7 Flash, Gemini 3.1 Pro.

He built the tier-list page itself with 5.6 Soul, which needed a screenshot and a scolding before it found the right vendor logos - smart enough to fix what you point at, dumb enough to need pointing.

Fable 5 and 5.6 Soul

Fable is the only S tier: it knows more than anything he has used, writes the code he is most willing to merge, and is what he trusts to review other models' work. It also trips over itself, takes needless shortcuts, and wanders into files it shouldn't - "a genius that has to be tamed", the smartest person at the company that nobody wants to work with and nobody will fire. It moved him from telling people to cancel Claude subs to running five of them behind a proxy and draining them weekly. He would never pay full API price for it.

Soul is the model he actually defaults to, and the one he would keep if forced to choose - slightly dumber but does exactly what it is told, easy to steer when the default behaviour is wrong, extremely token efficient, and far ahead of Fable on iOS work. He rebuilt the T3 Code mobile app in SwiftUI with it in a single thread. If the chart could show gap sizes he would split Fable into S+ and put Soul at S, because the jump from B tier to Soul changes how you work: with B tier models you must stay in the loop and review everything; with Soul and Fable you describe the goal vaguely, approve a plan, and get a PR that has been self-verified with computer use and sub-agent review.

Cheap models that earn their place

Luna is A tier not as a coding model but as the first cheap-fast-smart model in a while. It is his most-called model by volume: title generation, thread management, summarization, turning messy context into a useful JSON object, and enough agentic ability to make tool calls and pull from GitHub. He would not trust it with anything irreversible. It is also included in his OpenAI sub, so effectively free.

Deep Seek V4 Flash is the open-weight counterpart - similar capability, runnable on hardware you could plausibly own (roughly two DGX Sparks), which buys freedom Luna can't. It loops and gets distracted more, where Luna bails out of a wrong track faster. Then he learns Flash has no vision at all, which drops it behind Luna for him, and that V4 Pro has no vision either - which sends a large expensive 2026 model straight to F tier.

Where token efficiency decides the ranking

• 5.6 Terra :: measures fine (~half Soul's cost per task on artificial-analysis and Cursor Bench) but is less token efficient, so it only wins at odd price points. He would rather use Luna Max to save money or Soul on low to go faster. Never once chosen it.
• Kimmy K3 :: genuinely impressed him - first open-weight model that could carry long end-to-end work, good design taste, excellent vision, unusually strong at some 3D tasks. But at $15/M out with a big-model token appetite it costs more than Soul on xhigh, which is what he actually uses. Prices haven't fallen because hosting is expensive and because K3's license forces providers above ~$10M revenue into a deal with Moonshot that appears to pin them to MSRP. He bet the Hugging Face CEO that K3 pricing wouldn't drop and won on that technicality, so he isn't cashing it.
• GLM 5.3 :: a catch-up refinement of 5.2 rather than a new pre-train, stays on task better, no vision.
• Composer 2.5 :: Cursor's heavy RL on top of Kimmy K2.5, impressive and fast, but API-unavailable, slow on the cheap tier, and the advertised price is not the price of the fast version. Good for demos.
• Grok 4.6 :: a downgrade from 4.5 for him - marginally better orchestration bought with far more tokens, so it feels slower, and Soul on low gets there faster for similar money.
• Muse Spark :: underrated for data processing. It triages all his open PRs in under two minutes where Grok takes ten and Fable over an hour. Has an absurdly cheap contributor tier that trains on your data, and goes open weight soon. He would not use it for code.

The Gemini rant

Gemini 3.7 Flash lands in F, which he says hurts because Gemini 2.0 Flash was once S tier: 10c/40c per million, non-reasoning so output cost was predictable, fast, multimodal, genuinely useful. Each generation since inflated both price per token and tokens per task - 2.5 Flash with thinking on jumped to $3.50/M out and in practice ran 10-100x the cost of 2.0 Flash; 3.5 Flash reached $9/M out and often cost more than Pro at similar quality. 3.7 Flash finally cut prices, but as an introductory discount that doubles on 31 December, five months after launch of a model that was overpriced on day one.

The numbers behind the placement: 5.6 Soul uses ~60k tokens per task on max and ~30k on high for nearly the same score. Gemini 3.7 Flash burns 73k on low and 107k on high - and scores slightly worse on high than on medium. His answer to "but Gemini's tokens per second are faster": measuring a bike by wheel RPM ignores that some wheels are much smaller. He also grants that the Pro series still holds real knowledge (3.1 Pro still tops skate bench) that the models are too clumsy to surface.

Anthropic's lower shelf

Sonnet 5 is F tier over API and low D only as a thing you spend Claude Code sub quota on - he uses it to test sub-agents in T3 Code. Opus 5 is the interesting failure: it reads as brilliant, produces plans Fable itself often prefers, and then the code you try to merge is bad. He calls it mimic behaviour - looks, smells and quacks like a duck, then tastes wrong - and puts it in high D partly out of irritation at being fooled.

He plans to run one of these every few months, and is waiting on OpenAI's Astra to close the gap to Fable.