Same “Best,” Five Times the Bill
The same “best” label, but a fivefold price difference. Gemini and Claude each lead in different tests, and most leaderboards come from the companies selling the models themselves.
“Gemini beats Claude at coding. Claude beats Gemini at coding. Both are true.” This is the caption UP_SHIFTED posted with a video on X (formerly Twitter).
The video has a black background, with large white text reading “WHICH AI IS BEST AT CODING?” and blue text reading “Gemini beats Claude.”
Both claims are true because the tests are different.
- On DeepSWE (fixing real code), Gemini 4 Argon leads: 77.9% vs. 74.2%.
- On Terminal-Bench (working in the terminal), Claude Opus 5.5 leads: 66.4% vs. 57.4%.
- On operating a computer, GPT-6 Astra beats both.
The part that never makes headlines: most of these scores come from the model companies themselves. Every launch graphic is made by the people selling the model. So “the best AI” usually just means “the best on this test.”
Bills are easier to read than leaderboards. Per million output tokens (a token is the unit models use to count text):
- GPT-6 Astra: $50
- Claude Opus 5.5: $20
- Gemini 4 Argon: $10, for now.
The same “best” label can mean up to a fivefold difference in the bill.
In the replies, user @AIronAgent added: looking only at price per token is like reading only half the bill. If a cheaper model needs two tries, or writes twice as many tokens to finish the task, the cost per task can end up higher.
UP_SHIFTED agreed and offered a more practical approach: take 10 of your own real tasks and run each one on both models. One hour gives you more information than any launch chart. Later, @After_Dark_AI said five tasks is enough, and UP_SHIFTED agreed: five is just enough to actually run and enough to expose the kind of model that “looks cheap but isn’t.”
@airadar_1 also noted that Gemini’s “for now” is key. The launch price ends on December 1, and output will become $20, the same as Opus. The fivefold gap will soon become 2.5 times. UP_SHIFTED replied: leaderboards change with tests, and prices change with the calendar.
So next time you see “best,” ask three questions first: Best on which test? Who made the test? How much does it cost to complete one real task?
发布时间: 2026-10-26 22:49