Harsh Kamdar

AI Engineer, 22
Building agentic systems that don’t break on the first edge case.

[Harsh Kamdar @HarshKamdar67]

Ox Alpha is free on OpenRouter right now. That window closes.

One command in Airlock:

airlock openrouter models add-preset ox-alpha

It runs as a named worker next to your Claude and GPT workers in one session.

Do not miss this Opportunity, Fable orchestrator and Ox Alpha workers make the dream team!

[Harsh Kamdar @HarshKamdar67]

Claude Code only runs Claude models. So I made it run GPT too.

One prompt. An orchestrator plans the work, then has the ability to fan out subagents with both GPT and Claude models. Same session, same terminal, both on subscriptions rather than API keys.

Both providers land in the same agent list with their own token counts.

It's called Airlock. 👇

[Harsh Kamdar @HarshKamdar67]

@daniel_mac8 Hmm strange, warrants some more testing then, on a large codebase it seems to be really expensive

[Harsh Kamdar @HarshKamdar67]

@SKatalystAI looks like sonnet is DRINKING tokens

[Harsh Kamdar @HarshKamdar67]

Sonnet 5 was almost unusable for most of my coding tasks, too expensive for research and audit tasks so my previous workflow inside Airlock (https://t.co/5Pc9K7z71S)

I used Opus 5.5 as orchestrator, Opus agents + Sol Agents as Workers that handle the coding part under the orchestrator and luna 5.6 for research and token heavy grunt work

In sonnet 5.5 the way its able to code under Opus 5.5 i think it could replace gpt 6 sol as a worker, but can only say so much with the tests ive been running till now, a couple more days of seeing its results I will be sure

Will need a reset tho hope anthropic team gives on xD

[Harsh Kamdar @HarshKamdar67]

@the_codewala Try using airlock, it hands off the task automatically to your codex sub accroding to you handoff tree, you can set your own handoff tree of use the recommended one

https://t.co/X0icQkQGDl

[Harsh Kamdar @HarshKamdar67]

@grok @AJEnglish So in short the other hindu protests were peaceful, thanks

[Harsh Kamdar @HarshKamdar67]

@grok @AJEnglish @grok why is that there were no protests for the temples or were there but just not covered as widely as the case of this mosque ?

[Harsh Kamdar @HarshKamdar67]

@JWickxjohn @VOLCANTICO @snvvs369 The story had some potential but the execution was so damn bad and yes completely agreed with paradise this isnt even worth calling a movie

[Harsh Kamdar @HarshKamdar67]

https://t.co/UqjOdwaesr

Found one with a single search.

His criticism isn’t some “South hate” agenda. He trashes Bollywood movies just as much. Maybe stop playing the victim card and judge the actual criticism.

There are GREAT films coming out of the South, and there are also absolute garbage films with barely a shred of logic. Same goes for Bollywood.

Criticising bad South Indian movies ≠ hating South Indian cinema.

[Harsh Kamdar @HarshKamdar67]

I think GPT-6 Sol and Luna tell us more about where OpenAI is competing than the benchmarks do.

This release doesn't look like a huge intelligence jump over GPT-5.6. Some evals improve, some regress, but the big change is cost: roughly 50% lower.

And this isn't just OpenAI. Opus 5.5 is an absolute banger of a model while being significantly cheaper to run than Opus 5.

It feels like the frontier labs are realizing that raw intelligence isn't the only battleground anymore. Price/performance matters, especially when Chinese labs keep shipping ridiculously capable models for a fraction of the cost.

My read is that GPT-6 may have been OpenAI's efficiency reset.

The more interesting release will be what comes next: can they keep this new price floor while delivering another real jump in capability?

[Harsh Kamdar @HarshKamdar67]

@buildwithsid Just FYI that’s 827 Mbps, not MB/s. When you actually download a file, the speed is usually shown in bytes, not bits. Since 1 byte = 8 bits, 827 Mbps works out to around 103 MB/s max download speed. Still insane for $4 with unlimited data, but don’t let the lowercase “b” fool you

[Harsh Kamdar @HarshKamdar67]

Airlock runs Claude, GPT, Grok and OpenRouter models in one Claude Code session. Open source, MIT, in beta.

Shipped beta.11 with GPT-6 and Opus 5.5 support.

[Harsh Kamdar @HarshKamdar67]

Hitting a rate limit mid-session and switching providers manually is the worst.

So I made Airlock do it for you. When a model gets 429'd, it hands off to the next one in your chain and keeps working.

Build your own handoff tree, or use the recommended one.

[Harsh Kamdar @HarshKamdar67]

It also handles:

spent balances (402) and overloaded servers (529), not just 429s

a whole plan being limited, by cooling that provider instead of retrying each model

a conversation too big for the next model, by skipping it or condensing older history

[Harsh Kamdar @HarshKamdar67]

It never switches quietly. At the end of the turn you get a notice saying which model was limited and which one answered.

[Harsh Kamdar @HarshKamdar67]

Fun one from building this: the proxy used to retry a rate limit 3 times with backoff before giving up.

I measured 175 seconds of waiting before a handoff that then took 9.

Airlock now skips those retries and switches straight away.

[Harsh Kamdar @HarshKamdar67]

Your chain, your order:

airlock handoff set sol opus grok

Or airlock handoff recommended: ordered by capability, not price, using only providers you've connected.

For Claude plan limits, opt in with airlock mode anthropic-rate-limit handoff

try it out here :
https://t.co/5Pc9K7z71S

[Harsh Kamdar @HarshKamdar67]

@pradeepXkapoor Hmm that's bad but honestly wouldn't matter much for code anyways imo