Technology news, read properly.

Phones, chips, games, cars, space, security and AI. Everyone reports what shipped. Astraas reads the announcement twice and writes up the part that actually decides something: who pays for it, what it breaks, and who the deal was really for.

13 sections Written, not rewritten Every source linked Free

Privacy

GrapheneOS is escaping Pixel exclusivity. The price is the catch

GrapheneOS will support Motorola flagship phones from 2027, priced above the Pixel lineup. Cheaper devices will take longer because Qualcomm's security features and update commitments are weaker outside flagships.

23 August 2026 · 4 min read

The Feed

What happened, and what it actually means

Daily cadence starts with the newsletter. Subscribe below to get it from issue one.

Research

Longer pieces, with the working shown

Pricing

The subscription breakeven is 154 million tokens a month.

At published API rates, $100 buys roughly 154 million tokens on a cache-heavy agentic workload. That is the point where a Claude subscription starts winning — and it is much lower than people expect.

4 August 2026 · 7 min read

Prompt caching

A cache entry pays for itself on the second read.

Cache writes cost 1.25x and cache reads cost 0.1x, which puts the break-even at 1.39 reads. Measured across 9.75 billion cache-read tokens, caching cut an agentic bill by 87%.

4 August 2026 · 8 min read

Model routing

Routing 7% of calls to cheap models saved 0.89%.

Routing 6.9% of calls to Haiku and Sonnet cut 0.89% of spend. Measured across 23,157 calls, cheap-model routing is a rounding error unless it moves the work that actually runs.

4 August 2026 · 7 min read

Cost floors

A backtest said Sharpe 2.74. Real data said 0.35.

Nineteen backtested strategies, and the headline finding is not the fourteen failures. It is that the best proxy result overstated the real number by 87% — Sharpe 2.74 modelled versus 0.35 measured.

4 August 2026 · 12 min read

Inference economics

Inference got 95% cheaper. Your bill went up anyway.

Per-token inference prices have collapsed — 95% in two years by some measures — while enterprise LLM spend has doubled. The gap is explained by agentic workloads consuming 5 to 30 times more tokens per task.

4 August 2026 · 9 min read

AI ROI

95% of enterprise AI pilots return nothing. The failure isn't the model.

MIT found 95% of enterprise GenAI pilots produce no measurable P&L impact against $2.52 trillion of AI spending. The pattern in the failures is organisational, not technical.