Neither frontier lab clearly made tasks cheaper than its own last model — Contested · resolves 2026-12-31

AI models charge by the token, so the number of tokens a task uses sets what it really costs, and the debate is whether new models make each task cheaper or dearer.

Companies budget for the cost of each piece of work, not the list price, so when the two move in opposite directions budgets go wrong.

The original claim said OpenAI and Anthropic were moving in opposite directions; SecondSource reads both as no cheaper than their own predecessors, just charging in different ways.

SecondSource’s reading: OpenAI’s new model does use far fewer tokens per task, and two independent benchmarks measured it; at the same score, Artificial Analysis counted about 27,000 versus 78,000 output tokens.

But OpenAI raised its list price to 2.5 times the old level and took the savings back: on the same benchmark, GPT-6 Astra costs 15% to 60% more per task than GPT-5.6 Sol.

Anthropic went the other way. Its new flagship uses more tokens, but only the flagship gets cache reads at 2.5% of the normal price (a cache is content sent to the model again, billed lower the second time), which absorbs the extra tokens. The AI coding company Cognition was the first to say publicly that it moved work up from Opus 5.

Neither lab’s new model is clearly cheaper per task than its own predecessor. The only real saving for buyers is the price gap between vendors.

Three parts of the original claim fail: a 20% per-task rise is not a general rule, since another version of the same benchmark measured 12.8% cheaper; a price cut does not make anything dearer, a model change is what flips the direction; and Artificial Analysis is not the only one measuring.

The original material behind this call has not yet been linked here in a form readers can open and check.

Watch until 31 December 2026: if any Artificial Analysis or Harbor benchmark run at the same reasoning effort shows OpenAI’s next flagship costing less per task than GPT-5.6 Sol, the reading that OpenAI priced away its token savings is wrong; if it stays above Sol, the reading holds.

What would overturn it

OpenAI cutting Astra’s list or cache price below Sol’s per-task cost; Anthropic extending 2.5% cache reads to Opus 5; or a third-party comparison showing Fable 5.1 cheaper than Fable 5 even on low-cache work.