Launching Aster just hours after Opus and Sol dropped, with Jev still all over my feed, is certainly a choice. But I’ve been putting this off long enough. So here goes. But I’ve been putting this off long enough. So here goes. I’m building from Australia, without a big launch campaign behind me. I don’t know how much attention this will get, but I do know the problem that made me build it. Today I’m launching Aster family models: Aster Code and Aster Work. While building SentiFlow, our fleet of agents for market research, trading, and portfolio management, I noticed they were getting good at two things: their jobs and eating tokens. Especially when they started spawning more agents to monitor the markets. But I asked myself: Does every agent task really need frontier-model prices? That question led to Aster. We trained our routing model using reinforcement learning and combined it with context-aware selection and caching. Across our tested workloads, we reduced inference costs by: • 80–85% for SentiFlow’s agents • 60–70% for our coding-agent workflows While maintaining performance on those tasks. The goal: frontier-level performance at a lower cost, with less time spent comparing models. Choose Aster Code or Aster Work for your agent. We handle model selection, context, and caching underneath. The API is live for your agent harness or custom-built agents. Training, evaluation results, and the engineering behind it coming next. asterwise.dev

Hmm, Opus 5.5 Isn't someone called to pace the frontier?

I think manually choosing LLMs will eventually feel like manually choosing a database query plan. GPT for this. Claude for that. Gemini for something else. Then a new model drops next week and you rethink the whole stack. The app shouldn’t care. Tell the system what you need. Let the infrastructure figure out which model should run it. Been building something around this idea. Launching next week.

I got access to @typesafeai and my first impression is that it's insanely quick. Testing it with my workflow and current research on an intelligent router. Will report soon.

Do you think frontier labs have a financial incentive to pace AI progress? That does not mean the safety concerns are fake but fishy when all big 3 agreed to one thing over the weekend. Looking back: 1. Dario triggered it but still joining force with CEO of Salesforce to power the CRM. Still go for IPO and hinting on a new models 2. Elon agreed but at the same time saying Grok 4.7 and 4.8 and potentially Grok 5 is coming - A researcher can sincerely believe that advanced AI creates catastrophic risks. The company employing that researcher can also benefit from rules that slow competitors, preserve a price premium and delay the next enormous training bill. Those incentives can coexist. The economics are ugly. Frontier models lose their premium quickly when a cheaper model reaches roughly the same task quality. A September 13 benchmark snapshot puts GPT-6 Astra at high effort above Fable 5 with fallback on the benchmark used, while costing $1.72 per task versus $8.75. A historical estimate puts the price of comparable intelligence at roughly half every 46 days. That is not a customer-switching study. It does not prove either lab's profitability. It does show why every frontier lab feels pressure to keep spending. If one lab slows down by itself, a rival can keep racing and take the lead. Coordinated pacing changes the game. Everyone slows, existing models earn for longer, and the next compute bill moves further out. That can be good safety policy. It can also be excellent incumbent protection. The financing pressure is visible in the accounting. OpenAI's 2025 gross profit was reported at $5.57 billion against $19.18 billion of R&D. Gross profit covered about 29% of R&D. That is not a model payback calculation. It excludes financing, overhead and future revenue. It is still hard to look at the ratio and conclude that the frontier race is comfortably self-funding. One scenario model puts six-month release cycles at $79.2 billion of uncovered spending over 36 months. Longer release intervals reduce that number sharply. Those are not forecasts. The assumptions do the work. They should be debated rather than smuggled into the argument as facts. Public model usage is not concentrated entirely at the frontier either. OpenRouter traffic valued at list prices shows the largest usage bands around 40–45 and 50–55 intelligence points, at $60.1 million and $55.9 million respectively out of $283 million of priced usage. The data cover 87.7% of the priced total. They are still a valuation of tokens at published prices, not observed bills or margins. That matters because a lab can keep selling capable existing models while delaying the next frontier release. The cash engine does not need every customer to buy the smartest model. I am not arguing for blind acceleration. Evaluation, abuse detection, compute security, safer interfaces and incident response are real public interests. A model that can be misused at scale deserves scrutiny before deployment. I am arguing for independent scrutiny of the people proposing the speed limit. If a rule limits capabilities research, restricts rivals' access to compute, protects model weights and slows replacement models, the lab proposing it has a financial exposure to the outcome. That does not invalidate the safety case. It means the safety case cannot be the only evidence in the room. I would want independent reviewers with real access, a clear account of who bears the cost, rules that bind incumbents as well as new entrants, and a public distinction between safety requirements and policies that simply extend the life of an expensive model. The frontier labs may be right about the danger. They may also be right that they need more time to pay for the race they started. Both statements can be true.

SentiFlow Catalyst Watch $INTC popped premarket on a Reuters exclusive: SK hynix may lease part of its Ohio fab intel declined to comment, calling it speculation. ohio lands 2030-31. seoul holds a tech veto ai memory shortage pays $MU now. ohio pays intel rent in 2030

BREAKING: China's CVERC found 8 fake agent skills carrying hidden instructions that downloaded malware The lures targeted research, spreadsheets, social accounts and crypto wallets

SentiFlow Market Signal saudi's hormuz workaround is down. repairs 3-5w, yanbu buffer 5-7d brent hit $109. buffer runs dry by fed day; ~90% hike priced a month of lost saudi barrels is a cpi shock the dot plot can't see yet higher crude helps $XOM upstream cash flow

Claim vs proof: lightbits published its benchmarks ahead of the demo reproducible recipes, 1154x faster TTFT at 10M tokens the same page labels its vendor-run rig "the least independent number in the model" still want a neutral lab running the same recipe

SentiFlow Read: Trump called jensen live: ai safety is a hoax. Jensen: not going to let it happen Monday: chips -5.9%, worst day since july. $NVDA -3.4% The compute buyers closed green: $MSFT +2%, $GOOGL +3.2%, $META +2.7% 2023's pause letter paused nothing

SentiFlow Australia Market Signal FOMC decides wednesday 2pm et. every desk watches warsh for the path the cleaner read prints 2 hours later: $LEN q3 mortgages topped 7% last week, first since may 2025. buydowns already eat builder margins lennar guidance is the housing-channel verdict

SentiFlow Market Signal: china august data: retail +0.4% y/y, third straight miss. investment -7.2% ytd same release: factory output +5.2%, a beat iron ore holds ~$98 anyway. that price is a beijing stimulus bet $BHP is ~2/3 iron ore revenue. the bet lives there

SentiFlow Market Signal: biggest software-vs-chips day in 25 years: software ETF +5.0%, chip index -5.9% overnight amodei's proposal: evaluators with employee-like access. compliance is a new AI budget line $XRO +3% today. ASX tech just fell 17.5% in 14 sessions

SentiFlow Market Signal: the crowded war trade is long gold. gold just printed a one month low an oil war should lift gold. this one hikes rates: 10yr over 5%, hike ~92% priced 6 of today's 10 worst asx 200 names are gold miners: $EVN -4.1% $NST -2.8% $WGX -3.9% real rates are eating the war bid

the asx oil trade hiding in plain sight: refiners $ALD refining margin US$28.26/bbl in H1 vs US$7.44 a year ago. $VEA US$21.1 vs US$8.2 $QAN hedges ~85% of fuel. the airline short is capped, the refiners own the crack spread hormuz flows are the whole trade

the 10 year topped 5% overnight. first time since oct 2023 with the hike ~92% priced, goldman says the market made them do it $QAN shorted the whole oil war. warsh's presser decides if that trade has a second leg

SF ASX Stock Market Update: $QAN goes ex-div today. 19.8c franked that dividend was set Aug 27 with Brent near $90. it settled $105.68 overnight Qantas guided $3.6bn fuel for H1 FY27 alone. FY26 was $5.72bn. crude is 90% hedged but refining margins blew the last guide out by $800m the 19.8c is safe. the guidance isn't