OpenAI's internal Astra model generated machine-verified proofs for 10 long-standing mathematical and theoretical computer science problems. The company estimated the inference cost at approximately $2,000 in GPT-5.6 Sol API tokens, marking the first time it quantified the cost of frontier reasoning work. This pricing framework could enable smaller research teams to access frontier AI reasoning through APIs rather than building their own models, shifting from training costs to inference budgets as the relevant planning metric.
Researchers introduced MirrorCode, a benchmark that tests AI models on reimplementing entire software programs end-to-end without access to original source code. Claude Opus 4.7 successfully reimplemented gotree, a 16,000-line bioinformatics toolkit in Go, in 14 hours at a cost of $251, a task that would take a human engineer 2–17 weeks. The benchmark's open-source release and leaderboard enable systematic measurement of AI capabilities on long-horizon coding tasks that require weeks of inference budget rather than dollars.
Something Big Is Happening·2 months ago·
43
● 5 sources
Matt Shumer describes a technique called a Gauntlet Loop for improving AI agent outputs by having builders and independent critics repeatedly compare work against concrete reference examples until reaching a quality bar. He generated a 55,000-line Call of Duty game using Claude Opus 5 with this method, achieving results higher quality than typical AI demonstrations. The approach applies broadly to code, design, writing, and other outputs, with key principles including separating builder and critic roles, setting ambitious goals with concrete quality standards, and allowing agents to decompose work independently rather than following prescribed steps.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
