Sonnet 5.5 is a subagent, not a driver
- YT :: https://www.youtube.com/watch?v=8WbW_n95wc4
- Original title :: OpenAI should be scared of this one
Theo reviews Anthropic's surprise Claude Sonnet 5.5 release and lands on a counterintuitive verdict: the model is excellent, but almost nobody should select it directly. Its real value is as a tool that Opus or Fable calls — a fast, cheap subagent for codebase investigation. On the way he argues that "max" reasoning effort is a trap, that Sonnet's pricing is less of a bargain than the headline numbers suggest, and that OpenAI has now lost every dimension it used to lead on.
The announcement and what it actually claims
Anthropic positions Sonnet 5.5 as the second model in the 5.5 family: 30% faster and up to 30% cheaper than Sonnet 5, aimed at well-scoped everyday tasks while Opus 5.5 handles complex judgement work. Haiku 5.5 is promised in the coming weeks. Theo dislikes the Sonnet 5 comparison — Sonnet 5 was, in his words, one of his least favourite releases ever, so beating it is a low bar. The genuinely striking number is Terminal Bench 4, where Sonnet 5.5 posts the highest score recorded to date, 70.6% against Sonnet 5's 10.3%. A seven-fold jump makes him skeptical of the benchmark as much as impressed by the model. It also scores just under Opus 5.5 on GDPVal and is the first Sonnet to beat Pokémon Red from screenshots alone.
The qualitative shift he notices is tone. Anthropic has been stripping out the "Claudeisms" people hate, and Sonnet 5.5 writes more clearly than its predecessors. More broadly, the cheaper tiers now feel like slightly dumber, much faster Fable rather than robots mechanically executing instructions — as if Anthropic got jealous of everyone distilling its models and got good at doing it itself.
The pricing asterisk: cache reads
Sonnet 5.5 keeps Sonnet 5's pricing — $2/M in, $10/M out, $0.20/M cache reads. The cache read number is the problem. Fable 5.1 cut cache reads by 90% and Opus 5.5 by roughly 60%, but Sonnet got no discount at all. Because every other price scales cleanly (Sonnet → Opus doubles, Opus → Fable doubles again) while cache reads barely move, cache reads become a larger and larger share of your bill as you go down the tiers: under 5% with Fable, ~20% with Opus, and 50%+ with Sonnet 5.5. Since agentic work reads from cache constantly and emits few output tokens, the advertised savings mostly evaporate. On Artificial Analysis, Sonnet 5.5 ends up costing roughly the same as Fable 5.1 for real-world tasks — their most expensive run ever.
Why max reasoning effort should be ignored
Theo's sharpest aside is about reasoning levels. Low/medium/high/x-high are budgets — ceilings on how much the model may think — so bumping them often changes token usage by only 5–8% on simple tasks. Max is different: it raises the floor, forcing a minimum amount of reasoning before the model is allowed to finish. He has measured a 1,500% token increase going from x-high to max, and performance can actually degrade because a model made to overthink starts second-guessing correct answers. Concretely: Sonnet 5.5 at max cost $7.60 on a bench where Astra Max cost $3.26, while the same bench at high and medium cost $0.18 and $0.59. On Frontier Code, Anthropic even had to publish two Sonnet scores because x-high beat max. His advice: pretend max doesn't exist, and also avoid low — Anthropic dropped the no-reasoning option entirely because these models are bad without it.
Token hunger cancels out the speed
Sonnet 5.5 sets another record in Cursor Bench, this time for token usage: 271,920 tokens per task, ahead of Opus at 218K and Gemini 3.8 Flash at 162K, and more than 5x what GPT-6 Soul and Astra use. Open Router reports ~94 TPS through Anthropic (Theo sees ~150 on his subscription versus ~100 for Opus), but the extra tokens eat the advantage. On his "fish slop" game-generation task Sonnet took 43 minutes to Opus's 36, and 39 minutes of generation time to Opus's 27. Design is another weak spot: on the Witch AI showcase he finds Sonnet's output noticeably worse than Opus and significantly worse than Fable, which he still considers the best front-end design model.
Where Sonnet actually wins: as a tool for other models
The turn comes from a benchmark Theo built for his Grok 4.7 review — a deep audit of a huge T3 Code PR, asking models to propose how to break hundreds of thousands of lines into landable chunks. This is analytical comprehension rather than coding. Sonnet 5.5 came in at roughly half Opus's price while scoring slightly higher, and finished in about 5 minutes where Opus took nearly twice as long and Astra nearly three times. That profile — cheap, fast, good at digging through a large codebase to answer a question — is exactly what a subagent is for. He expects a properly configured Opus that delegates investigation to Sonnet to feel dramatically faster and cost slightly less.
He also reports two side observations: Opus has been deleting ~1.6 million lines of legacy slop from his Rust TypeScript-compiler rewrite, and his cost analysis across six Claude accounts suggests the $200 plan yields around $2,300/week of usage, close to $10k/month.
Verdict
"I don't think you should use Sonnet 5.5 yourself" — but it's an incredible addition to the family, and he's excited to see how Opus wields it. For OpenAI the picture is bleak: they don't have the smartest model, the most efficient one, the cheapest one, or the best one for calling other models, and his $200 Codex plan gets nowhere near the real-world output of his Claude subscriptions.