Wednesday is Day Five of Anthropic's rate-limit transition window. The company's April 17 credit-claim deadline passed Friday; the Monday feature treated the outcome as priced on the reading that subscribers who intended to migrate had already done so and the five-hour-window drain was still showing at the server side. Three days further on, with a full trading session between, the shape of the transition has become visible. Friday was the deadline for a credit; this Friday is the deadline for the transition itself. [1]

The benchmark that made the transition interesting is still the story underneath it. Moonshot AI released Kimi K2.6 as open weights on Hugging Face late Monday Beijing time, April 20 local, under a Modified MIT license, and Cloudflare added the model to its Workers AI catalogue the same day with Day-0 support. [2][3] The paper treated the release as a brief at the Anthropic-week timing on Tuesday. By Wednesday morning Artificial Analysis has K2.6 at 54.0 on its Intelligence Index, with a 1,520 Elo on GDPval-AA agentic tasks using the Stirrup harness, 96 percent on τ2-Bench Telecom for tool use, and open weights available via first-party and third-party APIs including Novita, Baseten, Fireworks, and Parasail. [4] On Cloudflare's own listing the model scores 83.2 on BrowseComp, 80.2 on SWE-Bench Verified, and 66.7 on Terminal-Bench 2.0 — the three benchmarks enterprise platforms use to evaluate agentic performance against what a Claude Opus 4.6 or a GPT-5.4 will actually do in production. [3]

What the paper's position has been, since Monday's feature on the Cherny concession, is that compute economics at the frontier are non-negotiable, and that each successive Opus model uses more thinking tokens than the last. Day Five sharpens that position in a different direction. The subscribers who migrated to Moonshot — or to Cursor's agent pricing, which the paper covered Tuesday at the $50B round — are not coming back Friday. The transition window is not a pending question; it is an asset that has already been rebalanced. The compute pressure that Cherny named last week is still there. The weights that cost zero per million tokens on some third-party provider are now there too. The public-company Q2 orientation Anthropic has signaled — Opus 4.7 in Q2, as Cherny said on Threads — lands into a market in which the open-weights frontier is scored above the closed-frontier incumbent on the benchmark that most developers now actually run. [5]

One thing worth naming: Anthropic has not responded publicly to the K2.6 release in the two trading days since. [5] The company's blog has posted nothing. Its X account has posted nothing. The paper's reading is that the non-response is the response: the company does not win by benchmarking its closed weights against open weights on HLE, and is unlikely to concede a price comparison when its API pricing sits ten to thirty times above Moonshot's Cloudflare listing. [3] What matters is what enterprise customers do with the comparison — and the answer, five days in, is that OpenClaw-style third-party tooling migrations to Moonshot are being quietly executed. The paper's Monday piece named Cursor's $50B round as one index of the same compression: the capital market is pricing the AI application layer to consolidate into three or four desktops, and the frontier-model vendors are one of the four, not the layer above it.

Friday closes the window. If Anthropic ships Opus 4.7 before Q2 earnings, the transition week becomes a footnote. If it does not, the footnote is that an open-weights Chinese model crossed the closed-frontier line during the week a US lab's subscribers were migrating away.

-- DAVID CHEN, Beijing