minimaxir

From a financial analyst on Bluesky: https://bsky.app/profile/peark.es/post/3mzdkyxzzns2w

> Okay so Q3 ARR at $67.8bn versus $46.2bn for Q2, +47% QoQ. Turns out tokenmaxxing's end over the summer slowed revenue growth versus an absolutely torrid 100%+ growth rate in Q1/Q2 but 46% sequential is still eye-popping

> Operating margins (GAAP) went from 8% to -3$, with a -28.5% QoQ incremental operating margin. $SPCX compute coming online is the reason.

> Bottom line after seeing this Q3 update: the basic takeaways from the original S-1 don't change, but fears that they were excluding Q3 because it was a disaster were definitely wrong. Do with this information what you will.

It's also worth noting that Q3 only has one week of Opus 5.5 (9/22 - 9/30) which may have changed the economics for Q4.

Don't complain about downvotes because that just gets you downvoted even harder.

Opus 5.5 is an order of magnitude better at design and design nuances than GPT 6.1 Sol, surprisingly.

If there is a design that LLMs do that you don't like, tell it that. You're allowed to be picky.

Not surprising: Claude has behaved like this for awhile and it was only a matter of time.

At the least, it deters people from spinning up Astra Pro requests via the ChatGPT interface willy-nilly.

You could make the same argument for these LLMs years ago but there's a reason that they've become faster and more accurate at beating these games. GPT 5.5 took 218 hours to beat Pokemon Fire Red; Astra took 21.

Only months ago did these LLMs make it past the Victory Road boulder puzzle without harness assistance.

Essentially yes. For games with some real-time component (FFVII), the game pauses after actions are executed to take a screenshot and let the agent think.

The LLMs no longer brute force a starter, and they have to play competently, including the use of setup. Here's a screenshot of the completion of the GPT-6 Astra run in 21 hours with comically underleveled Pokemon (and it actually savescummed RNG!): https://www.reddit.com/r/ClaudePlaysPokemon/comments/1w79vim...

Another run with Ironlocke rulesets for Crystal (hard mode, level caps, permadeath) has even more diverse teams: https://www.reddit.com/r/ClaudePlaysPokemon/comments/1wf6zx0...

The channel is here, with past broadcasts: https://www.twitch.tv/gpt_plays_games

If you haven't been keeping an eye on LLMs playing games, GPT 6.1 Sol can new beat games 100% vision-only (without harnesses that give the agent extra information about the game). Pokemon is no longer a difficult challenge for it, even with additional difficulty restrictions such as hard mode and level caps. Recently, GPT 6.1 Sol beat Final Fantasy VII and its bonus bosses vision-only surprisingly quickly.

I suspect Opus 5.5 can do even better but there aren't as many people testing it.

Show Hacks:

I've been holding off because I will need to do experimentation with local models at some point...but not yet and won't be for a bit but I am afraid prices will rise given component cost trajectories.

I've been tempted to just get the M5 Max instead and just innovate with smaller models myself.

I've had my eye on the normal Mac Studio M5 Ultra (normal plebeian 96GB RAM) and the release (edit: ship) date there keeps slipping, now into January. I'm surprised Apple is pushing the backlog even more.

> Sidenote: I should also add that you can see this is an anomaly when you use Anthropic’s own “dollar of revenue per compute” number. For Q2 2026, it brought in $2 of revenue per $1 of compute spend, up from $1.39 in Q1 2026. This number dropped to $1.70 in Q3. I think it’s a stupid number as it doesn’t really say very much about the business, but even based on Amodeiconomics, this quarter feels like the exception rather than the rule.

...making money is a pretty important part of the business?

Is this the first time VendingBench was used as a serious benchmark result in a model release? That's kinda funny.

There is apparently an in-person event that is not livestreamed. https://www.tomsguide.com/news/live/apple-touchscreen-macboo...

This bullet point is funny:

> Draw. Sketch and annotate directly onscreen.

Their demo example for the touch interface is piano keys? That's a 2010 Scott Forstall level of skeuomorphism that doesn't translate well for a laptop screen because it's not convenient.

Similar 75% price cut trajectory as GPT-5.6 Sol -> GPT-6.1 Sol. Overall, Terra is still half the price of Sol.

But like the original release of Terra, it's still a weird middle child between the small-model-that-got-substantially-better (Luna) and the large-model-that-got-significiantly-cheaper (Sol). The TerminalBench 4.0 cost-per-task results are a bit confusing and I'm not sure why you would pick Terra over Sol in that case. Overall this release seems like a flex on Sonnet 5.5: a model which is having its own middle child problems between the godly Opus 5.5 and the surprisingly capable Haiku 5.5.

For me, I used Zed Agent with Codex as it has native integration with Zed's ACP so it's pretty convenient.

However, Anthropic is pickier about the harnesses you can use with Claude Code and I noticed discrepancies when using Zed's External Agent feature with Claude Code, so I just use the Claude app. It's fine.

That's included in my estimate. Take for example Opus 5.5: https://openrouter.ai/anthropic/claude-opus-5.5

Currently for today, it's 1.19T input tokens, 5.41B reasoning tokens, and 4.29B completion.

Therefore, reasoning + completion is 0.8% of the total usage. Actually below my estimate!

Calling your reply aggressive isn't against the rules, bad faith or not (multiple people on HN clearly think it is aggressive). What is in the Guidelines, however:

> Be kind. Don't be snarky. Converse curiously; don't cross-examine. Edit out swipes.

Simon's comment apparently got flagkilled; I vouched for it.

If enough people are flagging it, we may be at a point where the pelican benchmark is truly overdone.

In an agentic pipeline, the output text is 2-5% of the total tokens used, so verbosity is a nonissue costwise. Although, higher verbosity could cause input to balloon faster and potentially cost more, but that would be captured by the cost-per-task metric.

But as noted by other replies, it's still lower than Opus/Sonnet.

Artificial Analysis index results are up, since multiple people are asking: https://artificialanalysis.ai/models/kimi-k3-1-pro

It's 55, at 4th overall, beating both GPT-6 Astra and Fable 5.1. It's also faster than both.

> When evaluating the Intelligence Index, it generated 150M tokens, which is very verbose in comparison to the median of 87M.

Sonic Adventure 1 is infamous for its lighting/textures successively degrading after each port from Dreamcast -> GameCube -> PC. There's a famous video showcasing all the changes: https://www.youtube.com/watch?v=SORYL5J1Heg

Page isn't up but it's a 55 (4th overall) https://x.com/ArtificialAnlys/status/2113688083145314416