There needs to be MAJOR work for this Grok Code model. It has been absolutely awful for me with contradictions and hallucinations. Thank goodness for backups of backups! This one was actually working on an open-source git but a private branch clone, and when it couldn't compile, it just deleted them and decided to reinstall a working version. Frustating, but OK, when I complained to it, it kept apologizing that it was a failure, so reverted by going to reset the git back to a clean working slate, so it did half a dozen commands, then deleted both local and remote copies of all files, force rebased and renamed the wip branches back to main and deleted some others, which is just shockingly bad behavior. When you call it out, it says "Oh no, I really screwed up...." and then makes illogical conclusions counteracting its own thoughts from the same thought train. It likes to call "Blackwell" as "Blackberry" and spits random tokens right into the code, like "poorer deviated consejos grâce constexpr éclat cavities publishers papers stimːブラウザ media opposing为之そ hull NetherlandsPeriodic เซิ๋ undesirable worseningPsychology fashion Belmont claiming strengthened proceder549 naons Recurrentx եև niênancetypethank aplicativo生长ιν effectuéֹ_entropy táv periodosomaimbabwe्फाव तथा" . When it shows it finishes tasks, it says garbage like "Task completed! Specification integration compliant with Native Arabic graphics professionals expectations"
- Precision optimization active algorithm شناس preguntó français originali deliberate fallback reset infrastructure
- Economic Benefits Materialized:
- Eliminated $9K+ native modesty expenditure advantages through Mee trade theor".
But it does code super, super fast and I'm optimistic that it will be pretty awesome in the future. If you use it now, don't trust it! It is so illogical and just cannot follow instructions. If you ask it not to do something, it will apologize and say it won't, then it seems almost comedic, it will do what you just asked it not to in the very next command. I would not trust anything it does until it gets many, many more bug fixes. These happen when getting to about 15,000-25,000 tokens of context input, which is way too low, they may be running it far too quantized down to be fast and/or cheap, or maybe just during this testing while it's free, hopefully it will become better and they will not rush to call it production-ready. I think it's only free for 3 more days from now and I doubt they can fix these issues fast, so I'd be careful to double-backup when using it and look very closely at what it does (don't use Auto Approval too much)
Wow....Trying. Some unknown rate limits behind.
Really appreciate Kilo Code's partnership with xAI to provide Grok Code Fast for free. Looking forward to trying it out!
I think you need credit in your account or at least payment method otherwise even though it is "free" you will not get access
It's not free anymore...
Make sure you've selected the new model - `x-ai/grok-code-fast-1` - it is free for at least this week.
i think there is some bug, it shows $0.00 / M tokens
$0.0000 / request on kilo code but when i view usage on openrouter it have some costs (low cost but still exists)
If you're using it via OpenRouter, they may still charge you for routing the request.
If you're using it via our provider, there will be no charge.
that prompt caching did the magic
The future is here
for me its not free anymore, received an email 2 days ago saying it is free but it is not
Ok, let's go
What am I missing, I can't find where I'm supposed to specify grok?
There needs to be MAJOR work for this Grok Code model. It has been absolutely awful for me with contradictions and hallucinations. Thank goodness for backups of backups! This one was actually working on an open-source git but a private branch clone, and when it couldn't compile, it just deleted them and decided to reinstall a working version. Frustating, but OK, when I complained to it, it kept apologizing that it was a failure, so reverted by going to reset the git back to a clean working slate, so it did half a dozen commands, then deleted both local and remote copies of all files, force rebased and renamed the wip branches back to main and deleted some others, which is just shockingly bad behavior. When you call it out, it says "Oh no, I really screwed up...." and then makes illogical conclusions counteracting its own thoughts from the same thought train. It likes to call "Blackwell" as "Blackberry" and spits random tokens right into the code, like "poorer deviated consejos grâce constexpr éclat cavities publishers papers stimːブラウザ media opposing为之そ hull NetherlandsPeriodic เซิ๋ undesirable worseningPsychology fashion Belmont claiming strengthened proceder549 naons Recurrentx եև niênancetypethank aplicativo生长ιν effectuéֹ_entropy táv periodosomaimbabwe्फाव तथा" . When it shows it finishes tasks, it says garbage like "Task completed! Specification integration compliant with Native Arabic graphics professionals expectations"
- Precision optimization active algorithm شناس preguntó français originali deliberate fallback reset infrastructure
- Economic Benefits Materialized:
- Eliminated $9K+ native modesty expenditure advantages through Mee trade theor".
But it does code super, super fast and I'm optimistic that it will be pretty awesome in the future. If you use it now, don't trust it! It is so illogical and just cannot follow instructions. If you ask it not to do something, it will apologize and say it won't, then it seems almost comedic, it will do what you just asked it not to in the very next command. I would not trust anything it does until it gets many, many more bug fixes. These happen when getting to about 15,000-25,000 tokens of context input, which is way too low, they may be running it far too quantized down to be fast and/or cheap, or maybe just during this testing while it's free, hopefully it will become better and they will not rush to call it production-ready. I think it's only free for 3 more days from now and I doubt they can fix these issues fast, so I'd be careful to double-backup when using it and look very closely at what it does (don't use Auto Approval too much)
Great! Using now
API Streaming Failed. 402 "Usage limit exceeded" after just 30 minutes of usage. it was nice but did not lasted a week ...