I collect roman coins, with latin legends, so the sun/earth/moon references jumped out at me, and partly based on the opus/sonnet/haiku precedent I assumed that these names were referring to different model sizes/prices in a way that mapped to the names (Sun > Earth > Moon).
I'll admit though that until recently I never really thought about Anthropic's naming scheme as having meaning (an Opus being longer than a Sonnet, being longer than a Haiku).
I'm not concerned about the code it's working on, but rather anything else - modifying files outside of the project dir (e.g. incorrect tool call), modifying system configuration, doing something bad on the internet, etc.
Interesting - I'd never heard of this Claude Code VM option.
Does it auto install all the dev/test tools it needs, maybe including things like web server & browser? Does your code live in the VM, or in some external repository? Is the lifetime of the VM the same as the agent, or does it persist until you remove it?
Not specific to OpenAI / Codex, but I'm curious what people are doing to protect themselves from any destructive actions by their coding agents? Just install and pray? Explicity approve all actions? Reconfigure for safety? Run in a sandbox (Docker) ?
I guess it does increase their cost, or rather your share of their hardware depreciation.
AI serving cost is apparently mostly hardware depreciation rather than operating cost (electricity etc), and if your large context request is occupying VRAM for some fraction of a second then you are paying for the depreciation that occurs in that time!
e. H100 costs $20-40K to buy, with a lifetime of maybe 3 years, and will only consume maybe $2K in electricity if run 24x7 for those 3 years.
3D printing is a good comparison - it allows almost anyone to make things, but in the end very few do.
Another example is when the WWW first became available, and suddenly everyone COULD be a publisher (browsers even included built-in HTML editors), and for a while MySpace pages proliferated until the excitement died down and people went back to being media consumers.
I expect we'll see the same thing with consumer use of generative AI. Suddendly everyone is generating 3-D worlds/games with Fable because they can, but I expect that just as with the web the novelty will wear off and they'll leave it up to the pros.
Professional use of GenAI, and coding in particular, is certainly here to stay, but it seems we're still in the early experimental/hype phase. At least tokenmaxxing has passed, and it seems most companies are now paying attention to, and limiting, how much they are spending, but it doesn't seem we've yet progressed to the stage where companies are paying attention to what they are actually getting out of it - is the money spent showing up on the bottom line in the form of increased revenues.
I've integrated AI into my daily life by installing the Gemini voice app onto my phone, which I typically may use once a week, and adding Gemini and Claude bookmarks to my browser which I use all the time - but mostly Gemini since free usage is effectively unlimited.
That's all the integration I need. I don't need OpenClaw running 24x7 trying to hack it's way into my gmail.
In contrast I read that Ukraine is approving 4+ new weapons systems PER DAY !!
Even when it comes to more expensive things like cruise missiles it seems the planning has to be that some high percentage of them may be shot down (and much higher for slower moving drones), so you really want them cheap and in high volume, with reliability somewhat of a secondary concern.
Companies are building chips specialized for inference, so dual use for training isn't necessarily a consideration, but there are other considerations such as:
Weights need to be loaded into the accelerator's processor fast, which means they need to be physically adjacent to it, but there is limited physical space for that - not enough to fit the all the weights of a 1T+ param model, so weights get loaded into VRAM dynamically according to what part of the model is being run.
ROM (I guess we're talking Flash memory) can be dense, since it is built vertically - many hundreds of layers, but this comes at the cost of poor performance, so even if you could fit enough ROM next to the processor it would not be fast enough.
The recent Databricks comparison has GLM 5.2 performing identically to Opus 4.8 on high effort, and some early Twitter reports (e.g. from the OpenCode developers) strongly favor GPT 5.6 Sol over Fable.
As always it depends on what you are using them for, and how you are using them.
Different incentives. Claude Code makes more money for Anthropic by generating larger contexts. Anthropic also recently changed their tokenizer so the exact same code input creates 30% more tokens, so there's a pattern there.
Just to add context (no pun intended), OpenAI also charge differently based on context usage with GPT 5.5 being $5/30 below 200K, and $10/45 above.
Anthropic have a fixed price regardless of context usage.
These per-token pricing schemes aren't directly comparable though since these models all use different numbers of tokens, even for input (Anthropic's recent tokenizer change generates 30% more tokens for exact same input), as well as for reasoning, and context/token usage also varies wildly by harness with Claude Code using 3x the context/tokens of Pi.
It's great to see a large-scale real-world benchmark from a user of these tools, as opposed to the the benchmaxxed results from the vendors themselves. Also great to see different harnesses being tested, with considerably different results.
Definitely a few surprises here:
1) GLM 5.2 using Pi performs identically in terms of pass rate (~87.5%) to Opus 4.8 high using Claude Code, but significantly cheaper ($1.25 per task vs $2)
2) Absolute best pass rate (90%) was from Opus 4.8 x-high using Pi, beating out Opus 4.8 using Claude Code
3) Pareto frontier performance from any of the models (Opus 4.8, GPT 5.5, GLM 2.5) was using Pi rather than native harnesses
Apparently Pi used 3x less context than Claude Code, and one takeaway is to use Pi regardless of what model you are using. The other takeaway is that in real-world performance GLM 5.2 is the equal of Opus 4.8 unless you run Opus 4.8 on x-high in which case you can eke out a 2.5% increase in pass rate at the expense of doubling your cost over GLM 5.2