Sonnet 5 is now the cheaper model. That is not the same as the cheaper answer.

At the GPT-5.6 launch the story was simple. Terra was the cheap capable one and OpenAI was undercutting Anthropic on price.

Check the two pricing pages today and that has quietly inverted.

Identical on input. Sonnet 5 is 17% cheaper on output. Terra is nearly four index points ahead on v4.3, and cost Artificial Analysis 2.8 times less to run the index.

On list price the pairing looks close, in the way that makes people waste an afternoon on a spreadsheet and still pick wrong.

Why the price comparison flipped

Neither side announced it as a response to the other.

OpenAI cut the whole GPT-5.6 line. Terra went from $2.50 / $15 to $2 / $12. Sol and Luna came down too, Luna hardest, from $1 / $6 to $0.20 / $1.20.

Anthropic did something quieter that mattered more. It cancelled a price rise. Sonnet 5 launched at $2 / $10 described as introductory pricing that would end on 31 August 2026 and step up to $3 / $15. That step never happened, and the docs now say plainly that the increase will not occur. Sonnet 5 stays at $2 / $10 as its standard price.

Had that rise gone through, Terra at $2 / $12 would today be the cheaper model on both axes and this article would not exist. One cancelled price change decided the entire comparison.

The number that actually decides the bill

A 17% output discount sounds decisive until you look at how many tokens each model spends getting to the end.

Artificial Analysis found Sonnet 5, at maximum effort, generating around 300 million tokens during its evaluation. That was roughly five times the comparison median. Claude Code users have reported the same shape independently: cheap per token, but on large jobs it takes more steps, consumes more context, and can finish more expensive than a stronger model would have been.

The v4.3 round measured it again. Sonnet 5 generated 370 million tokens across the index against Terra’s 120 million, and Artificial Analysis spent $6,998 running the index on Sonnet 5 against $2,501 on Terra. On that workload the model that is 17% cheaper per output token cost 2.8 times as much.

Put those facts next to each other and the arithmetic stops being about rates. A model 17% cheaper per token that takes twice as many steps costs you more. The one that is 20% dearer and finishes in a single pass is the bargain.

This is the same argument I made about Luna against Sonnet 5, where the price gap is tenfold and still does not decide it, and about GPT-6 Astra, where OpenAI’s entire cost claim rests on step count rather than rate.

It still cuts both ways on your own work. An index run is not your workload. But the one full-run measurement we have favours Terra by a wide margin, and the two-dollar row in the price table is not evidence for Sonnet 5.

The tiebreaker nobody puts in the comparison

Here is the part that decides it in practice, and it has nothing to do with either model.

Sonnet 5 runs inside Claude Code. Terra lives in ChatGPT and Codex. If you already run one of those harnesses, the model is not really the unit you are choosing. You are choosing the tool, the prompts you have tuned, the permissions you have configured, and whatever scaffolding you built around it.

Moving that costs more than either price gap recovers, and not in licence fees. The cost is the fortnight where everything is slightly worse while you rebuild the parts you forgot you had.

What would change my mind

Three things, and I will update this piece when any of them is published.

Another index round. The one I was waiting for, v4.3, measured both models on one scale and put Terra nearly four points ahead. A new version could move that again.

A cost-per-completed-task comparison on a real agentic workload. The v4.3 run costs, $6,998 against $2,501, are the strongest evidence here, and they still cover one evaluation suite.

Any further move on price. This entire comparison turned on a rise that was cancelled rather than on anything either model does, which tells you how thin the margin is.

Questions people asked

Is GPT-5.6 Terra better than Claude Sonnet 5?

Yes, on current numbers. Artificial Analysis’s Intelligence Index v4.3 scores Terra at 42.25 and Sonnet 5 at 38.36, nearly four points apart. On list price Sonnet 5 is cheaper on output, $10 against Terra’s $12, with both at $2 for input, but running the whole index cost $2,501 on Terra against $6,998 on Sonnet 5, because Sonnet 5 generated 370M tokens to Terra’s 120M. The harness you already work in can still decide it, but the numbers now lean to Terra.

Which is cheaper, Claude Sonnet 5 or GPT-5.6 Terra?

Sonnet 5, on list price. Input is identical at $2 per million tokens; output is $10 for Sonnet 5 and $12 for Terra. But list price only decides the bill if both models take a similar number of steps, and on Artificial Analysis’s Intelligence Index v4.3 Sonnet 5 generated 370M tokens against Terra’s 120M, and the full run cost $6,998 on Sonnet 5 against $2,501 on Terra. On long agentic runs the cheaper rate can produce the larger invoice.

Should I switch from Sonnet 5 to Terra?

Not on the index alone, but the case for testing it is now strong. Terra scores nearly four points higher on Artificial Analysis’s Intelligence Index v4.3 and cost 2.8 times less to run the full index. Moving harness, retuning prompts and rebuilding the tooling around a model is still a real cost, so measure your own cost per finished task first, and switch if Terra wins it.

What is GPT-5.6 Terra best at?

It is the balanced tier: analysis, research, document work and moderately hard coding, at a fraction of what the Sol flagship costs. OpenAI positions it as the sensible default with Sol kept as an escalation path rather than an everyday model. If most of your work is professional output that does not need frontier reasoning, Terra is the tier that was designed for it.

If you want this applied to your own stack rather than to a price table, the contact page is the fastest route to me.