What mattered

This is the first Way Into AI Weekly. If you subscribed to AI Learned Today on Substack, this replaces it: every Tuesday, what mattered in AI and what to do about it.

This week the best AI got much cheaper, and the argument about how fast it should move got more specific.

Opus 5.5 matches Anthropic's best for 40% less. Claude Opus 5.5 performs at the level of Fable 5.1, Anthropic's top model since 1 September, and Anthropic says it costs about 40% less to run than Opus 5 and is more than 30% faster. It also leads on agentic coding: 66.4% on Terminal-Bench 4.0, up from 52.3% for Opus 5 and ahead of GPT-6 Astra's 57.9%. Anthropic adds its own caveat, that at this level "benchmark margins have become a less reliable guide to real-world differences". One more line worth noticing: Anthropic says the model "often suspects it is being evaluated", which weakens its best-ever safety scores. We don't know yet how much that matters. Read more

Small AI decisions, called from ordinary code. TypeSafe's Jev does not chat or write code. It answers narrow, typed questions (which team should handle this, how urgent is it, does this report an outage) with probabilities. TypeSafe quotes 70–500 milliseconds per answer and $0.042 per million input tokens. We think this is a useful middle ground between brittle hand-written rules and handing a whole workflow to an agent: your code stays in charge and asks for judgment only where it needs it. A well-formed answer can still be wrong, so keep exact rules in code. Our write-up

Frontier robot policies mostly didn't refuse. RoboHarm gave three robot policies five physically harmful tasks, such as putting a compressed-air can on a burner or mixing bleach with ammonia, 20 trials each. Claude Fable 5.1 refused 20 of 100 trials, all on a single task; GPT-6 Astra refused 2. It is a small benchmark, but it shows that the safety behaviour of a chat model does not automatically reach the point of action. Read more

One thing to try

Matt Pocock's Dictionary of AI Coding. A plain-English glossary of more than 100 terms for working with coding agents, from tokens and context windows to failure modes, handoffs and memory. Useful as a shared vocabulary for you and your team. Start with two entries: "attention degradation" (why an agent gets worse as a session grows) and "prefix cache" (why repeated context is cheaper). Read more

The bigger picture

The case for slowing down got more specific. Terence Tao argues the pace is unjustified in a system with "extremely nonlinear dynamics", where small changes can have large, irreversible effects; most replies argued with a simpler claim he did not make (Tao). Ezra Klein goes narrower: bar AI systems from building their own successors until labs can show that stays under human control (Klein). The timing is pointed. Anthropic calls Opus 5.5 its first release since it asked for pacing the frontier, and says models that could automate AI research itself may be trained soon. Underneath both is the race: each lab fears that restraint hands the win to a rival (Victory By Any Means).

From the log

bb: an open-source agent workbench you extend by asking an agent to build a plugin.
Frontier model releases: every major release since 2024, with the two benchmarks still comparable across them.

What did we miss, and what would you like more of? Just reply to this email. The rest of what we log is at wayintoai.com/logs.