『Opus 5.5 is your new workhorse! OpenAI ships GPT 6 Sol and Luna before DevDay and Meta goes all in on Muse! Your friday read is here』のカバーアート

Opus 5.5 is your new workhorse! OpenAI ships GPT 6 Sol and Luna before DevDay and Meta goes all in on Muse! Your friday read is here

Opus 5.5 is your new workhorse! OpenAI ships GPT 6 Sol and Luna before DevDay and Meta goes all in on Muse! Your friday read is here

無料で聴く

ポッドキャストの詳細を見る

【Amazonプライム会員限定】今ならプレミアムプランが4か月 月額99円。

10月19日まで。※適用条件あり
Hey, it’s Alex 👋What a freaking week! A week after every lab head agreed to “pace the frontier”, Anthropic and OpenAI shipped big new models within hours of each other, and both of them are CHEAPER. So much for pacing 😂Visit https://thursdai.news/ep/2026-09-24 for all the links in this podcastAnd we had a new producer on the show today! Opus 5.5 listened to us live, put up the chyrons, kept me on time (mostly) and even fact-checked Nisten on air. If the stream died, you knew who to blame. You can re-watch it work on thursdai.live if you want to experience being there with us!3 huge themes this week: pacing the frontier does not mean stopping, Assistants not just agents (Meta went all in on Muse at Connect), and voice, where Google now says it has the best TTS in the world... and it clones voices.With me: Peter Gostev (Arena), Nisten, Yam Peleg, and Wolfram Ravenwolf live from AI Engineer Paris (thanks Mazi for the tether!), plus JevBench creator Florian S. Let’s dive in (all links at the end as always!)Frontier AI - not pacing yet!Claude Opus 5.5 - Opus is BACK, Fable-level smarts for 40% less (X, Blog, System card)Folks. FOLKS. Opus 5.5 is the highlight of my week, and it’s not even close. It’s like Opus 4.6 is back, with Fable-level abilities, and it talks like a normal person again - no more Jargon Douche Claude!My week started with burning through my quotas on Fable, and I was like, oh no, I need Claude for production on Thursday! Then Opus 5.5 dropped and... I just couldn’t get to the end of my quota. I ran workflows, agents, Claude Code for hours with no end in sight. Remember the “limitless Codex” days when you never thought about quotas? Then Astra came out and I burned my entire weekly quota in half a day (ps. this was partly due to a bad config, so if Astra is burning your tokens, keep reading for a fix). Now it’s flipped, and Claude is the seemingly limitless one. And it’s fast!The numbers back it up. It beats Anthropic’s own Fable 5.1 on GDPval-AA (1846 vs 1735), 66.4% on Terminal-Bench 4.0 vs 57.9% for GPT-6 Astra, and it’s 40% cheaper than Opus 5. Output goes from $25 to $20 /1Mtok and cached reads from 50 cents to 20 cents, and for agentic coding the cache reads are most of your bill, so you really feel it. Basically the smaller, overachieving brother of Fable. Sonnet and Haiku 5.5 are coming in the next few weeks too.Peter came in hot with Arena news: Opus 5.5 is #1 on Code Arena, above Astra, and the HTML and 3D stuff it generates is “completely insane.” His one caveat: on his hardest max-effort prompts, a single generation cost $60 to $80 in API terms, so the long tail can still get pricey.And then Nisten, who (like all of us) has specific opinions about Anthropic’s politics, said it writes the best code he’s seen, it’s crazy good at WebGPU and kernels, and it’s “an absolute banger.” It’s his default now. Folks, do you understand what it takes for Nisten’s default to NOT be some obscure open source model he runs himself on 17 GPUs?! Anthropic won over Nisten. I don’t think you guys get what just happened here.Wolfram is the holdout, he left Anthropic over the OpenClaw bans and hasn’t touched it since. Wolfram, how do I say this gently... you’re an evaluator, this is not allowed 😂An Anthropic employee basically said “sorry for Opus 5, we hope this makes up for it”, and honestly it does. Opus 5 was slow, incoherent and full of Claude-isms. Nobody wanted to talk to it. The Claude-isms are gone, the jargon is gone, this model talks like a human. I used to love Opus back in the Opus 3 days, and it feels like that again.Want to squeeze even more out of it? Read the great Addy Osmani’s guide to Opus 5.5 (@addyosmani). 2 things blew my mind: stop writing “think carefully” in your prompts (”You don’t need to ask it to think”, it picks its own depth, and replies start sooner without it), and one early tester found Opus 5.5 on its LOWEST effort caught more bugs than Opus 5 on high, with fewer false alarms. Try low effort before you go max!Tip from me to you: if it ever gives you something confusing, the pstack “bro” skill (/bro) makes it say it again in human words. I use it all the time.(Full disclosure: Opus 5.5 produced the show AND helped with this writeup, so it might be a tiny bit biased 😅 but the quota thing is 100% me.) Anthropic, please, please don’t nerf this one.GPT-6 Sol and Luna - half the price, and honestly better than Astra for me (X, Blog, Caching)Hours later, OpenAI answered with GPT-6 Sol ($2 in / $10 out) and Luna (10 cents in / 50 cents out, that’s basically free!), at half the GPT-5.6 price. The lineup is now Astra, Sol and Luna (bye Terra). On DeepSWE, Sol gets 68.8 and Luna 66.6. The sneaky big one for builders: 90% off cached input, and changing reasoning effort or tools no longer busts your cache 👏Ok, hot take... Astra has been kind of dumb for me day to day, burning tokens for ...
adbl_web_anon_alc_button_suppression_t1
まだレビューはありません