@ArtificialAnlys Per Artificial Analysis, output pricing falls to $5 per million tokens from $12, and the cost to complete its benchmark tasks at max effort drops to $0.40 from $1.40. Cached input is cut to $0.05 per million tokens from $0.20.
OrigineelOpenAI heeft GPT-6.1 Terra uitgebracht en verlegt daarmee de grens van kostenefficiëntie. Het is het eerste model van de Terra-klasse sinds GPT-5.6, en is intelligenter voor minder dan een derde van de kosten. Prijzen per token zijn de helft of minder dan die van GPT-6.1 Terra - op $1/$5 per miljoen input/output tokens vs $2/$12 voor het vorige model. Cache-schrijven is ook gehalveerd, nu op $1,25 per miljoen tokens. Gecachte input is een kwart op $0,05 vs $0,20 per miljoen tokens. Belangrijkste punten: ➤ Pareto-optimaal voor Intelligentie vs Kosten
0
0
0
0
Reacties 21
-
-
@ArtificialAnlys Good chart, thanks
Origineel -
@ArtificialAnlys For computer use, tokens per finished task matters more than price per token. A model that needs fewer screenshots and retries can still win on the real bill. Would love to see cost per completed task next to per-token pricing.
Origineel -
@ArtificialAnlys cached input going from $0.20 to $0.05 is the line i'd look at first. agent loops resend the same long context every step, so on long runs the cache price probably moves the bill more than the $1/$5 headline.
Origineel -
@ArtificialAnlys bruh?
Origineel -
seeing the exact same ordering in codex on farmerbench (three 6-hour seeds each at medium effort): gpt-6.1 sol: 74.1 mean at $0.47 a run (159 points/$) gpt-6 astra: 66.6 mean at $1.79 a run (37 points/$) gpt-6.1 terra: 62.2 mean at $0.22 a run (286 points/$) gpt-6 luna: 15.7 mean at $0.08 a run (189 points/$) at 286 points per dollar and 2m 56s a run, terra sits right on the pareto frontier between deepseek v4.1 flash ($0.13/run, 46.8 mean) and 6.1 sol ($0.47/run, 74.1 mean).
Origineel -
@ArtificialAnlys This is the same Sonnet fumble. Dinner and more expensive than Sol. No place for it.
Origineel -
@ArtificialAnlys @grok if this has 20 point jump in Terminal-Bench 4.0 eval and matches gpt-6.1 sol at max effort, does that mean it is excellent for coding work given cost?
Origineel -
@ArtificialAnlys Benchmarks look great. I'll give Terra, Claude and Grok the same real task today and post cost per task.
Origineel -
@ArtificialAnlys umm https://x.com/garyfung/status/2114849687396286696
Origineel -
@ArtificialAnlys Opus 5.5 still seems like the best
Origineel -
@ArtificialAnlys Cool, OpenAI's really been pushing on cost efficiency these days. Looking forward to seeing its actual performance.
Origineel -
@ArtificialAnlys 更强还不到三分之一成本，这波确实狠
Origineel -
@ArtificialAnlys Not the most impressive upgrade, more like a tiny improvement. Competes with 6.1 Sol depending on the reasoning level. xHigh seems to be particularly lazy
Origineel -
@ArtificialAnlys Virei ateu de benchmark
Origineel -
@ArtificialAnlys who is terra for? it's 6 points under sol on the coding agent index, and for agent work it still looks underwhelming next to sonnet 5.5. is the case just low effort at volume, where it sits on the cost frontier?
Origineel -
@ArtificialAnlys 🤯🤯
Origineel -
@ArtificialAnlys The hallucination drop (88% to 38% guess rate) is the stat I'd care about most for screen work. A wrong click costs more than a wrong sentence. Any plan to break that out by effort level?
Origineel -
@ArtificialAnlys Imagine your AI finishing the entire task while you're still reading the instructions 😅
Origineel -
@ArtificialAnlys 最值得看的是幻觉率 88% → 38%：不会时少瞎猜，对 agent 来说比智力指数多几分更重要。另外智力指数上低 / 中档在帕累托前沿、最高档反而不在，日常开中档可能才是性价比最优解。
Origineel -
@ArtificialAnlys Cost per task just dropped again. The clusters that can serve it efficiently will see the volume first.
Origineel
스크롤하면 더 불러옵니다