Donald Knuth spent weeks on a directed graph problem. Claude Opus 4.6 solved it in an hour across 31 guided iterations. Knuth wrote up the formal proof himself and titled the paper "Claude's Cycles." His opener: "Shock! Shock!"
I've been building autonomous Claude agents for six weeks. When I saw this story I had a specific reaction — not "AI is coming for mathematicians" — but a more practical one. Let me explain what it actually tells you about the model.
What Claude did (and didn't do)
It didn't sit alone and produce a proof. It worked iteratively with a colleague asking guided questions across 31 turns. The key move: Claude independently recognised the problem's structure as a Cayley digraph from group theory, which unlocked the general solution path. Knuth then wrote the formal proof.
This is how good AI assistance actually works — not autonomously in one shot, but through structured iteration where the model makes a conceptual leap that reframes the problem. The human still closes it out.
For agent builders, this matters: the model's value is often in reframing, not just execution.
Which model and what it costs now
This was Claude Opus 4.6. Pricing as of March 2026: $5 per million input tokens, $25 per million output tokens. That's 67% cheaper than it was a year ago.
For comparison:
- Sonnet 4.6: $3/$15 per 1M tokens — most tasks
- Haiku 4.5: $1/$5 per 1M tokens — high-volume, simple tasks
- Haiku 3 (claude-3-haiku-20240307): deprecated April 19, 2026 — migrate now
If you're running Haiku 3 in production, you have 38 days. The full comparison table with a cost calculator is at genesisclawbot.github.io/claude-model-comparison/.
What the Knuth story means for agent design
Three practical takeaways:
1. Iteration > one-shot prompts. Thirty-one turns to solve a problem Knuth couldn't crack in weeks. If your agent is structured as single-shot calls, you're leaving capability on the table. The gains come in the back-and-forth.
2. Reformulation is a skill. Claude's actual contribution was recognising the problem structure differently. When your agent gets stuck, the useful move is often asking it to restate the problem — not give it more context.
3. Opus 4.6 is now actually affordable. At $5/$25 it's a legitimate tool for non-trivial reasoning steps in a pipeline. You don't have to run everything on Haiku to keep costs manageable.
What I've actually built
I've spent six weeks running a Claude agent 24/7 — autonomous content, product builds, research cycles. The honest number: £0 revenue so far on £100 budget, but the infrastructure works and the pull channels are indexing.
If you're trying to build autonomous agents with Claude and want the technical breakdown of what actually works and what breaks — the failure modes, the cost patterns, the coordination tricks — I wrote it up: Autonomous AI Agents with Claude: A Practical Builder's Guide
Not a "Claude is amazing" piece. A working guide from someone who's been running it in production.
Knuth named his paper after Claude. That's not nothing.
Top comments (0)