Sonnet 5 Is Dead in the Water

Ignore the token price for a second and look at the run cost.

Theo's total-run screenshot shows Claude Sonnet 5 max at $6,015 for the full Intelligence Index run. Fable 5 with fallback is $5,631. Opus 4.8 max is $3,753. GPT-5.5 xhigh is $2,819.

That was not supposed to be the Sonnet 5 story.

Sonnet 5 was meant to be the sensible middle. Cheaper than Opus. Less weird than Fable. Strong enough to become the new Claude Code default without everyone feeling like they had been moved to a worse model with nicer pricing copy.

Instead, the first serious cost chart says: more expensive than Opus on the run, more expensive than Fable in that screenshot, and nowhere near GPT-5.5 xhigh.

The per-task chart is not much better. Artificial Analysis says Sonnet 5 costs $2.29 per Intelligence Index task, driven by higher token usage. The screenshot puts Opus 4.8 max at $1.80, GPT-5.5 xhigh at $1.03, GLM-5.2 max at $0.48, and Fable 5 with fallback at $2.75.

So the defense is technically true and practically useless: Sonnet 5 is cheaper than Opus per token.

Fine. But I do not buy tokens. I buy the loop. Read the repo, make the change, run the tests, notice the mistake, fix the mistake, stop before the whole afternoon is gone. If the cheaper model needs more reasoning, more output, more retries, and more context to get there, the cheaper token price is just a smaller number on the wrong page.

This is why the launch feels so off. Anthropic released Sonnet 5 on June 30, 2026 and made it the default for Free and Pro users. Claude Code gets it too. The sticker price looks attractive: $2/M input and $10/M output through August 31, then $3/M and $15/M. Opus 4.8 is $5/M input and $25/M output. Fable 5 is $10/M input and $50/M output.

That pricing table only works if token usage stays under control.

But Anthropic's pricing docs say the newer tokenizer used by Sonnet 5 and newer Opus/Fable models produces about 30% more tokens for the same text than older Sonnet models. Then Artificial Analysis measures the actual benchmark bill and Sonnet 5 lands above Opus 4.8 per task.

The temporary launch discount helps for two months. It does not fix the product positioning. Serious users planning agents and automation should care about September 1 pricing, not the launch sale. Artificial Analysis used the normal $3/$15 price.

Maybe Sonnet 5 is fine for chat. Maybe it is fine for small edits. Maybe it is fast enough in some workflows that the bill makes sense. I am not pretending one benchmark is every workload.

But Anthropic is selling this model into exactly the kind of work where the chart hurts: agents, tools, long tasks, coding loops, browser work. Claude Code users care about task economics because that is where the money and quota disappear.

That is why this launch looks so bad. Sonnet 5 is below Opus 4.8 on the intelligence axis and above it on cost per task. In the total-run screenshot, it is above Fable too.

That is not a clean default. That is the worst middle: not the best model, not the cheap model, and not efficient enough to make the trade obvious.

If I want cheap, open models like GLM-5.2 are moving fast. If I want frontier coding agents, Codex and Opus are the real comparison.

Sonnet 5 had to be boring in the best way: cheaper than Opus, good enough, painless as the default.

Its first serious cost chart says the opposite.

Lower token price. Worse task economics.

Dead in the water for production agent work.

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com