Shaun Smith
2.4K posts
Shaun Smith
@evalstate
https://t.co/rA1UoojwhN https://t.co/76p6mDAfej
united kingdom Katılım Temmuz 2024
889 Takip Edilen1.1K Takipçiler
This. is. insane. $0.04 per trial.
Last weekend I benched Grok 4.5 at 80.5% on tb-2.1 for $159 and was impressed.
Today gpt-5.6-luna has beaten it: with flex tier pricing the run cost less than $20.
Industry economics changed this week.
English
@xeophon No - not fast. But I've not had any weird errors or disconnects [yet].
English
@andrew_n_carr Agreed. I've completely changed the way I use subagents these last 2 weeks; for conversational stuff big models are a total waste now.
English
My guess is that the models are getting harder to understand because they're being optimized for sub agent spawning and communication.
We are not the intended audience
English
fast-agent model string:
responses.gpt-5.6-luna?reasoning=max&service_tier=flex
Migrate your automations here: fast-agent.ai/guides/migrate…
English
Full ATIF trajectories available in harbor, leaderboard submission PR#184. Cost taken from API Key used for the trial.
fast-agent cache efficiency was 93.83%, 1.75% of prompt tokens were priced at long context rates.
English
Great overview (as always) from @simonw on MCP.
The protocol is reinvigorated.
A single URL reliably connects, authenticates and adds capabilities almost universally - it's hard to beat.
Simon Willison@simonw
The new stateless MCP specification has rekindled my interest in MCP, and inspired some new projects, including mcp-explorer and datasette-mcp simonwillison.net/2026/Jul/31/st…
English
@SarahLacard Yep, my estimate for a TB 2.1 run is about $10. I'm currently figuring out how reproduceable the published score is and doing a couple of reasoning level sweeps.
English
@evalstate oh i haven't gotten that fancy with the sauce - only using it in opencode, it's a monster, and cheaper than luna from what i can tell
English
@SarahLacard I have, and am on that now😀. Very nice they are delivering it over the Responses API too (but no Web Sockets).
English
English
I love my team, colleagues and job! It’s been so much fun the last couple days - just letting my ideas run free and building things I care about!!
Oh, what a privilege it is that I get to do what I do \o/
Yes, this post is brought to you by two pints and a negroni
English
@Infoxicador Yeah, it's on the official leaderboard but an earlier version of codex. Token drain is real, I think fast-agent is about half the cost of OpenAIs figures in their report.
English
One of the things I love most about this video is how well it captures the ethos of OpenAI and what drives the people behind it to wake up every day and deliver some of their life’s best work.
Jason really put his heart and soul into this one. Kudos, mate!
jason@jxnlco
one of the most beautiful things about OpenAI is that every employee really has a voice. i wanted to capture what it feels like to work here, what our mission means to me, and why you should join us. so i made this video with Codex, shared it with the team, and they felt it was worth producing and sharing with the world. this is our mission. and it’s why i’m here.
English
@liran_tal Yep, but when tokens are that cheap...
(aesthetically it's horrible, but I think after using GLM for a bit we should all switch reasoning traces off for good)!
English
@evalstate Interesting! Wouldn't heavy token cost directly contribute towards higher spend though? (not across pricing models but across same tier in general)
Like for example Opus 4.8 and seems like Sonnet 5 too are incredibly verbose in thinking process
English
Shaun Smith retweetledi
Thinking Machines released Inkling Small (🦖) + NVFP4
12B active 276B total params, the model performs better than larger Inkling on coding 🤯
> check out our blog covering benchmarks, performance and deployment huggingface.co/blog/thinkingm…
> huggingface.co/collections/th…
English
@un3valuated Yeah, I'll post some terra benchmarks tomorrow. Sonnet is buried, Flash is buried and at $15 output Moonshot doesn't compete here. Grok is the one to watch now. What a time.
English
@evalstate My half cheap half intelligent model was Gemini 3 flash for last 8 months, now Luna costs in half. Rest in peace bro, you'll be missed.
English
Holy shi_. Assuming flex tier pricing still holds this is insanely aggressive.
OpenAI@OpenAI
We are committed to pushing the model frontier across cost efficiency, capability, and speed. Starting today, we are reducing prices for GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20% , and offering a faster option for GPT-5.6 Sol in the API. Luna and Terra’s lower prices are reflected in how usage is counted in Codex and ChatGPT Work, so your usage goes further.
English