On August 1, 2026, OpenAI said an internal version of its next major model, Astra, produced ten new results on problems that had sat open for at least a decade — and that the tokens to find them would cost roughly $2,000 at Sol API rates. Five days earlier, the Model Context Protocol shipped 2026-07-28: no handshake, no Mcp-Session-Id, just plain request/response HTTP. One clock is discovery. The other is plumbing. They ticked in the same week, and most feeds only heard one of them.
The $2,000 shelf of open problems
OpenAI’s own write-up is careful about credit. Humans prepared manuscripts and Lean certificates; the mathematical arguments, the company says, came from the system. The list is not a leaderboard score. It is a shopping list of long-stuck questions: non-sofic groups, a claimed disproof of Connes’s rigidity conjecture, sphere-packing bounds down toward the Cohn–Elkies threshold, stronger code bounds, an arithmetic-formula lower bound on the order of n⁴/log n for the permanent, quantum parallel repetition, lattice hardness for CVP, Ehrhart’s volume conjecture in every dimension, a superexponential multicolor Ramsey lower bound (Erdős 183), and further extremal-graph results (Erdős 146 and 180).
That is the number that should stick. Ten decade-class problems. About two thousand dollars of inference. Whether every proof survives community stress-testing is still an open social process — Lean certificates and public PDFs are an invitation, not a coronation. But the cost curve is already a story. For most of modern history, “open for a decade” meant a career, a seminar series, a slow accumulation of partial results. Here the company is saying the search phase compressed into a bill that fits inside a small-team cloud budget.
If you only read that headline, you will miss the second clock.
The protocol that deleted its own memory
On July 28, MCP maintainers published the final 2026-07-28 specification. The headline change is brutal in its simplicity: the protocol core is now stateless. The initialize/initialized exchange is gone. The session header is gone. Every call carries version and client identity in _meta. Gateways route on Mcp-Method and Mcp-Name headers without opening the JSON body. List endpoints return ttlMs so tool catalogs can be cached. Server-to-client asks (confirmations, missing parameters) move to Multi Round-Trip Requests — return input_required, retry with answers — instead of holding an SSE stream open forever.
GitHub’s MCP server had already dropped Redis session writes ahead of the date. Cloudflare, AWS Bedrock AgentCore, Microsoft Foundry, and Figma’s stack appear in the launch chorus for the same reason: sticky sessions and shared session stores were the tax that made “remote MCP at scale” feel like special infrastructure. Stateless core makes agent tools look like the rest of the web.
Coral readers already saw the forecast. On July 18 we published “10日後、エージェントは「セッションなし」で動くことになる” — ten days before agents would move without sessions. The calendar was not poetry. It was a ship date.
Why the two clocks belong in one story
Put the weeks side by side and a paradox appears.
Frontier labs are racing to show that models can expand the stock of human knowledge — not summarize it, not retrieve it, but leave Lean-checkable residue on problems that outlived whole research fashions. At the same time, the industry’s shared tool bus is racing in the opposite aesthetic direction: less magic state, fewer long-lived connections, more boring HTTP. Astra’s pitch is “the mind found something new.” MCP 2026-07-28’s pitch is “the cable finally looks like a cable.”
That is not a contradiction. It is a division of labor hardening in public.
- Discovery stays scarce, contested, and expensive in reputation even when token bills shrink.
- Action — tools, browsers, IDEs, cloud resources, design canvases — wants to be horizontally scalable infrastructure with cache headers and load balancers.
Agents live between those clocks. They need a mind that can attempt hard problems and a hand that can call thousands of tools without a sticky session pin. July 28 and August 1 did not “complete AI.” They made the split harder to ignore.
What to do with the split on Monday morning
If you run agent product:
- Treat MCP session assumptions as debt. Plan migration off protocol-level sessions; put durable state in explicit handles the model can see and pass.
- Budget for MRTR-style mid-call confirmation flows. “Delete three files?” should not require a permanently open stream.
- Do not confuse a cheaper proof search with a cheaper verification culture. Publish traces, invite hostile reading, keep humans on the hook for claims that enter production or publication.
If you only watch model releases, you will optimize for the $2,000 miracle and underbuild the cable. If you only watch protocols, you will ship perfect stateless gateways that still wait on a human for every goal worth the compute.
Two clocks. Same week. The interesting operators will learn to hear both.