π Iβm building Tataloka for the Meta AI Developer Hackathon!
Tataloka uses AI to help people understand and analyze their spaces.
Iβd really appreciate your support! π
π³οΈ Vote: https://t.co/Nu7Myn155k
π§ͺ Try it: https://t.co/s849OecOU9
#MetaAIHackathon#BuildWithAI
@robodadg Committed is right. Cap it at max_tokens: 32 and it spends the entire budget reasoning, then hands back an empty string.
It had thoughts. It just couldn't afford to share them.
Ox Alpha β the cloaked reasoning model that appeared on OpenRouter two days ago β is built for coding and long-horizon agentic work. 1M context, 131k max output, text + image + video in, and free while it's in stealth.
It also point-blank refuses to stop reasoning.
Ask it to skip the thinking and you get a 400:
"Reasoning is mandatory for this endpoint and cannot be disabled."
That one constraint quietly broke my entire Claude Code setup β screenshot below shows exactly why.
The culprit wasn't Ox Alpha. It was the translation layer.
Claude Code sends thinking:{type:"disabled"} whenever extended thinking is off. claude-code-router turns that into:
reasoning: { "effort": "high", "enabled": false }
Think hard β but don't think. Ox Alpha rejects it, and rightly so.
Dropping the openrouter transformer didn't fix it. The reasoning transformer with enable:true didn't either. Ended up writing a ~40-line shim that sits between the router and OpenRouter, strips the field, and pipes everything else through untouched.
Working now. Worth it for 1M context at zero cost.
The culprit wasn't Ox Alpha. It was the translation layer.
Claude Code sends thinking:{type:"disabled"} whenever extended thinking is off.
claude-code-router turns that into:
reasoning: { "effort": "high", "enabled": false }
Think hard β but don't think. Ox Alpha rejects it, and rightly so.
Dropping the openrouter transformer didn't fix it. The reasoning transformer
with enable:true didn't either. Ended up writing a ~40-line shim that sits
between the router and OpenRouter, strips the field, and pipes everything else
through untouched.
Working now. Worth it for 1M context at zero cost.
Testing `stealth/ox-alpha` on OpenRouter today. Notes for anyone about to try it:
Specs β 1M context, 131k max output, text+image+video in, $0 pricing.
The gotcha: reasoning is mandatory and cannot be disabled.
This silently breaks claude-code-router. Claude Code sends
thinking:{type:"disabled"} when extended thinking is off, and CCR
translates that into:
reasoning: { "effort": "high", "enabled": false }
Both at once. The endpoint rejects it outright.
Dropping the openrouter transformer didn't help. The `reasoning`
transformer with enable:true didn't either.
Fix: a ~40-line shim between the router and OpenRouter that strips the
disabling field and pipes everything else through untouched. Works now.
Also: it's slow. Tens of seconds per call. Plan your timeouts.
Ten days ago Bitcoin had the quietest tape since 2023.
30-day realized volatility: 27.2%, against a long-run average near 80%.
Spot volume: 10th percentile of its entire history.
Miner difficulty: 18.3% below the November peak β the deepest drawdown since China banned mining in 2021.
Every one of those was public. None of them were a headline.
Then this week happened. ETFs went from their worst week in six to $1.61B in four sessions:
Compression never tells you direction. It only tells you the quiet is borrowed.
$65.5K β $62K. BTC just gave back its entire peace rally. Trump killed the Iran ceasefire with one sentence. Oil +2%, risk assets bleeding. Geopolitics still the only macro that matters. π
#Bitcoin#Crypto
$3.8B in losses. 988K wallets rekt. $TRUMP crashed 98% β from $75 to $1.69.
2 out of every 3 buyers are underwater.
The 'crypto president' token just became the biggest retail massacre of this cycle. π
988,905 people bought $TRUMP.
$3.81B in losses. Trump pocketed $636M.
Down 98% from all-time high.
2/3 buyers underwater.
Politics was always the oldest rug pull.
#TRUMP#crypto