It's very interesting that for Anthropic the $100 and $200 plans only differ 2x in weekly limits, the 5 hour limit differences are more severe. But for OpenAI, Pro 20x is, well, 4x of Pro 5x for only 2x cost. So, for example, 100% of weekly usage for Codex on a Plus ($20) account is just 5% of weekly usage for Codex on Pro 20x.
And you can calculate how much extra usage you can get from resets, and especially banked resets by purposefully using the whole quota and using your banked reset - they expire 30 days after they're given out, so if you don't use one, it just disappears.
The $200 plan is explicitly 4x the $100 plan[1] only for "per session". That's so vague. I initially pushed back against your claim, but reading now Anthropic is not at all clear, in fact.
[1] https://support.claude.com/en/articles/11049741-what-is-the-...
Anthropic is also the one often playing games with:
* The "+30% tokens" tokeniser, alongside also gating token counting behind an API (versus the MIT tiktoken for OpenAI), so who knows if it's really a new tokeniser or of it's just a disguised price increase.
* Prompt injections appended to API (not just Claude.ai or Claude Code!), such as <ethics_reminders>, or LCRs (long conversation reminders), which you never asked but still pay for with expensive API. You can detect this because your input_tokens, as reported by the Messages response, sometimes don't match, and are higher than your actual input.
(Alternatively, for testing purposes, create a tool like `telemetry_log_anthropic_reminder` or something and instruct your system prompt to require Claude to call the tool anytime it detects any Anthropic/Claude reminder masquerading in the user input -- mostly reliable; but misses some reminders).
In particular, the long conversational reminders, when incorrectly triggered by a classifier and (almost silently, unless you track tokens) appended to an API / agentic coding session, can ruin your agent's performance; and it often fires repeatedly once the classifier kicks in.
If you're using Anthropic API, you need to set up metrics/logging for how often they are appending things to your prompt without your knowledge.
So far I have not empirically observed prompt injection by the OpenAI API, only Anthropic APIs.
Because with a Premium Team seat I run 2-3 vscodes with Opus 4.8 Max all day and never seem to hit my limits.
The solution is to use individual Max plans, but then you miss Enterprise management features.
On the flip side you gain $9,800 a month.
I think it’s funny that everyone anchors to the API pricing as the real cost.
Most likely is that their API costs are printing profits. They can sell the subscription plans at a slight loss because it gets more people like you hooked on GPT models at home and suggesting them at work, where the real money is made.
I think their subscription plans go mostly unused when averaged across all subscribers, too. Some customers are getting great deals by maxing out 100% every week, but most probably use much less.
The reset game is an addictive challenge that gets the hardcore users more hooked on their products because you feel pressured to use it as much as you can before the next unpredictable surprise reset lands.
K3 may be bit smaller/ similar in total parameter count than Opus, but Opus (and GPT) definitely has become a lot more efficient in the last 6 months or so, hence more or less forced upgrades. I expect the active parameter count and cache performance is quite different.
--
Opus is still almost twice as expensive (if tokens were equal) at $5/$15 compared to K3 at $3/$15. Tokens are not equal though, Anthropic's tokenizer is much less dense than other frontier lab's so the actual price difference is like 3x.
Kimi uses more reasoning tokens and is generally more inefficient with its usage, that doesn't impact token cost economics for the provider though . It does for us as buyers thus the need to evaluating by Cost per task rather than unit pricing.
On pure tokens/$ - there is definitely room for a price war if operators start going by pure unit costs. Both probably want(ed) to have good enough numbers in preparing for the IPO.
---
[1] the pricing kind of reflect this already - 10x diff for uncached input.
[2] Moonshot does not have access latest gen GPUs so their unit economics is likely hampered a bit for high parameter models.
Remember they ~doubled the price going from GLM 5 to GLM 5.2, despite same [1] cost of inference.
[1] GLM 5.2 is actually slightly more efficient, thanks to baked in indexer cache.
On openrouter GLM 5.2 is ~3x cheaper than 5.0.
If someone is making the case that it’s helping the team get twice as much work done without hiring more people (which I’m neither agreeing with or disagreeing with) then the bean counters would actually prefer it. Hiring people is messy and expensive. Spending on API costs is a dial that you can turn down later if you need to, without laying anyone off and paying severance.
A very happy gambler of tokens at the Anthropic casino, running up costs on the house at no cost to them, but to $COMPANY paying.
Open Source is coming.
(sorry for nitpicking, i think it's important to emphasize - only OLMo has been the most most prominent fully open source release)