Shekelphile

That's crazy. You obviously don't care much about your children's lives if you let them roll around in 60+ year old vehicles with methhead level repairs done to them.

right now it looks like only anthropic since openai made the staggering business decision to release their midrange model (terra) rebranded as sol 6.1 and quantized astra into the ground.

everyone else mostly seems to have just decided to cheat on benchmarks to make their models look better than they actually are, and none seem to want to actually try to make better models.

Nah it's super easy for the most common game system emus, and especially easy for 'modern' emulators like dolphin and rpcs3.

> Cache is cache, you can't get better cache than... cached...

No other provider is getting the cache hit rates deepseek does or the retention. They use a unique setup that no other provider is bothering to replicate.

> Quantization is not trash, there's a year of evidence that shows Q4 provides ~4% degradation and Q8 provides ~1% degradation, and you don't need that severely quantized to gain benefits in inference performance.

Lol. Quanting at all pretty much neuters a model, just look at how astra performs now compared to at launch.

There is no subscription for 1st party deepseek models, the ones that exist are all fronting as middlemen for discounted openrouter providers that serve quantized models with worse cache retention. The only real option for deepseek has been API for a few months now.

oai and anthropic also subsidize the hell out of their subs compared to what you will find in smaller competitors, a $20 codex sub gets you like $100-150 usage/wk which goes way further than 2x opencode go (which would only net out to $120 of deepseek 4.1 usage a month, on top of being low performing quantized trash).

Luna and Haiku 5.5 are just as cheap and much better.

Don't really understand people who say DS4 or 4.1 have frontier level performance. Anyone who has used it will tell you that it's a hallucination factory. The only thing it has going for it is deepseek's unique infrastructure that allows better cache retention, but the cost savings from that obviously come nowhere near how much subsidized usage you get out of even a $20 subscription with openai or anthropic.

They're cutting prices because they want to cannabalize the market for people using models like deepseek via API as well as people paying for anthropic subs.

When they cut prices on luna the first time around they took (literally) millions of users from anthropic.

Footnote on their pricing page says:

> Cache hits and refreshes on Claude Opus 5.5 are priced at 0.05x the base input price.

If they do the same for Haiku and Sonnet 5.5 then we should also see 5c/mtok and 10c/mtok cache read for those models, respectively. Still too high for Haiku IMO, Luna is 2c/mtok.

I mean, why would you? Chinese models through an API are more expensive than any subsidized subscription and the only one offering a transparent subsidy (opencode) only gives you $15 of credit on any relevant models, and still has a 5 hour limit.

The answer is just a second $20 subscription.

Changing pricing buried in docs (which they didn't actually do when they started switching every model to $15 limit 2 months ago BTW, they only did this once people noticed and called it out) without notifying users is the definition of a rugpull.

All of the chinese labs have been overfitting on benchmark data to game the results for a while now - MiMo and Deepseek are not anywhere near frontier and mostly compete with models like Luna - which they are still worse than.

There isn't much compelling reason to use these unless you are just averse to giving money to openai/altman. A $20 codex sub gives you ~$150 of luna use per weekly limit, while there isn't any good subsidized options for chinese models at all (and the few who were subsidizing, like opencode, rugpulled by reducing monthly limit to $60 to $15 with no notice to users).

No chinese lab has caught up yet. They've tried to fake it by distilling and overfitting on benchmarks to make their models look better than they are, the 'best' models available from chinese labs right now (GLM 5.3 and Kimi K3) fall apart completely when you try to do real work with them. K3 is especially embarrassing because it is larger than Mythos yet performs worse than opus 5 and 5.6 sol in benchmarks they haven't been able to fake yet.

> it's also token efficient and cheaper than gpt 5.6.

Deepswe results show that grok 4.6 is more expensive per-task and consistently scores worse than: luna xhigh, glm 5.3, astra low, sol high/xhigh, opus 5 medium.

Grok also used almost 3x as many tokens/turns to complete tasks than all of those models (besides luna), so it takes way more time to complete a task.

There isn't much reason to use Grok at all, it's gotten better but it's still worse than every other player in the field, which shouldn't be a surprise considering until about a year ago they were just buying tokens from other providers and pretending it was their own model.

With gpt-6 luna and sol coming tomorrow it's going to look even worse too, especially if new luna retains the same dirt cheap pricing that 5.6 luna has.

Yes.