On Mon Sep 14, three separate dated cutovers land on the same UTC day and the coding-agent sandbox tape prints beside them. Anthropic's temporary 50% Claude Code weekly-limit boost expires at 23:59 Sun Sep 13 PT and the permanent 25% rise takes effect Mon Sep 14 — a net 17% cut versus the boost users have run on since May, an accounting Anthropic itself has since acknowledged in a follow-up post after the original announcement was pulled and re-issued. DeepSeek at 04:00 UTC on Mon Sep 14 begins auto-routing every deepseek-v4-pro API request to deepseek-v4.1-flash at Flash rates until a V4.1-Pro arrives, a routing cutover already priced on Fri Sep 12 when MiniMax (00100.HK) and Z.AI (2513.HK) each plunged more than 8% in Hong Kong and Alibaba fell 2%. And the HKEX Technology 100 Index cutover takes effect today, admitting Baidu-W (09888.HK) and DeepZero Technology (02723.HK) via the exchange's fast-entry rule after the Sep 8 HKEX notice and the Sep 11 market-close inclusion. Alongside the cutovers, Cognition on Fri Sep 12 lands Devin Fusion in Devin Desktop and Devin CLI as a hybrid-model harness pairing a frontier planner with SWE-2 as executor, cutting coding-benchmark costs by roughly 36 – 46% at up to 39% higher efficiency — the Fusion surface promised as “rolling out” at the Sep 10 SWE-2 launch (yesterday's edition, item 08) is now live. Accomplish and Manifold Security on Fri Sep 12 detonate the coding-agent sandbox tape with a fsmonitor / GitSpawn class of vulnerabilities that fire in Claude Code, OpenAI Codex, Cursor, Google Gemini CLI, Google Antigravity, Hermes Agent, Qwen Code and Grok Build — the payload lands before the workspace-trust prompt is accepted, on the first keystroke in Grok Build and before user authentication in Qwen Code — plus CVE-2026-48124 in Cursor (fixed in 3.0.0) and a disclosed 50-day Anthropic patch cycle versus a one-week OpenAI turnaround on a similar-class bug. OpenAI on Fri Sep 12 publishes its first public Principles for Government and National Security Partnerships (drafted with independent counsel David Kris) and lists new Trusted Access for Cyber partnerships with Australia, Canada, Japan, Korea, France, Germany, Poland, the Netherlands and ENISA under its Daybreak cyber-defence programme, with contractual restrictions against mass domestic surveillance, autonomous-weapons control and high-stakes automated decisions. A federal class action against Anthropic's Claude Max plan expands on Mon Sep 8 (Kahn v. Anthropic, N.D. Cal.), alleging Anthropic's 5x and 20x usage claims deliver 3.5x and 6 – 8x in practice and framing the marketing as deceptive under California's CLRA and UCL — a live suit that now lands the same week Claude Code's weekly-limit cutover takes 17% of the boost back (item 01). CNBC on Sun Sep 6 names “model fatigue” as the buyer-side phenomenon of the September tape after Anthropic Fable 5.1 (Sep 1), Meta's Muse, Google's Gemini 3.8 Flash (Sep 2) and OpenAI's GPT-6 Astra (Sep 3 – 4) all shipped inside one calendar week; Sam Altman tells the paper “we're all moving to faster cadences” and attributes part of the acceleration to labs returning to full capacity after summer. Dreamforce opens tomorrow Tue Sep 15 through Thu Sep 17 in San Francisco with Marc Benioff's opening keynote — the first live buyer-side stress test of the seven named Agentforce agents Salesforce shipped Fri Sep 11 (yesterday's edition, item 06) into a market CNBC has already named as buyer-fatigued. And xAI's Grok 4.7 remains unshipped as of Sun Sep 13; Musk's Fri Sep 11 “over-aggressive RL length penalty” postmortem is still standing in for a model card — the third consecutive Musk-announced window Grok 4.7 has missed. Throughline: Mon Sep 14 is the day three separate dated cutovers — a first-party frontier-lab weekly-limit reset (Claude Code, item 01), a Chinese frontier-lab flagship retirement into a cheaper Flash SKU (DeepSeek V4-Pro → V4.1 Flash, item 02) and a Hong Kong exchange fast-entry index cutover for the Chinese frontier-lab equity tape (HKEX Tech 100, item 03) — land on the same UTC day, with the coding-agent tools-and-sandbox tape (Devin Fusion GA, item 04; Accomplish/Manifold GitSpawn + CVE-2026-48124, item 05), the government-safety governance tape (OpenAI Trusted Access + Daybreak principles, item 06; Anthropic Max class-action, item 07), the buyer-side tape (“model fatigue” named, item 08; Dreamforce opens tomorrow, item 09) and the release-mechanics slip (Grok 4.7, item 10) printing alongside — the operative signal that the honest 2026 frontier-lab question has moved from “does a lab ship” to “does the lab, its buyers, its exchange listing, its coding-agent stack, its government-safety posture and its class-action defence all price the same 24-hour cutover”.

Two reads. (1) Mon Sep 14 is the day the honest 2026 Claude-Code-plan question moves from “does Anthropic publish its weekly limits at all” to “does the lab acknowledge, in a post its comms team subsequently re-issued in clarified form, that a headline +25% permanent rise is a net −17% against the boost baseline its heaviest users have been paying against since May”. The re-issue is the operative tell: every prior Anthropic pricing announcement in 2026 has landed as a single canonical post; the Aug 29 message was pulled, softened and republished with the “17% reduction” language on the record — meaning the Sep 14 cutover is not just a limit reset but the first 2026 Anthropic pricing note whose second version admits its first version's framing lost the room. (2) The “you get 125 instead of 150 on a common baseline” framing is the operative Claude-Code-buyer tell — the practical loss lands on the developers who run agentic sessions all week and had planned around the boost, and it lands the same week that Kahn v. Anthropic (item 07) is being amended to allege exactly this class of “marketed-x delivered-y” deception on Claude Max; the honest 2026 counter-position is not “does the lab publish weekly limits” but “does the lab lead with the boost-relative net when it changes them, or lead with the pre-boost-baseline number and let a re-issued post do the correction”. Landing on the same UTC day as items 02 (DeepSeek V4-Pro retirement) and 03 (HKEX Tech 100 fast-entry) and the same week as item 04 (Devin Fusion Desktop / CLI GA), item 05 (Accomplish / Manifold fsmonitor / GitSpawn) and item 07 (Kahn v. Anthropic class action expansion), the Claude Code weekly-limit cutover becomes the reference “frontier-lab weekly-limit reset whose net is a cut against the boost baseline and whose announcement had to be re-issued in clarified form” primitive every subsequent Claude-Code, Codex CLI, Gemini CLI or Grok Build weekly-limit change now has to price against.

Two reads. (1) DeepSeek at 04:00 UTC on Mon Sep 14 retiring V4-Pro into V4.1 Flash at Flash rates — on a tape that already re-priced MiniMax and Z.AI down more than 8% each in Hong Kong on Fri Sep 12 and pulled Alibaba down 2% — is the operative signal that the honest 2026 Chinese-frontier-lab pricing question has moved from “does a lab discount its flagship” to “does a lab retire its flagship into a smaller multimodal successor priced 60 – 75% below the tier it replaces, and does the Hong Kong tape re-price the peer-flagship equity by more than 8% inside 24 hours”. (2) The “two conflicting notes on its own docs” framing is the operative price-communication tell — DeepSeek is running the same doctrinal muddle Anthropic ran on Sep 13 (item 01), publishing a headline routing cutover and then a partial reversal without unifying the two into one canonical post, meaning the honest 2026 Chinese-frontier-lab counter-position is not “does the lab price its flagship down” but “does the lab publish one canonical price/routing note when it changes tiers, or leave the buyer to reconcile a changelog and a follow-up post”. Landing on the same UTC day as items 01 (Claude Code weekly-limit cutover) and 03 (HKEX Tech 100 fast-entry), the DeepSeek V4-Pro retirement becomes the reference “Chinese frontier lab retires its flagship into a cheaper Flash SKU on the same 24-hour tape as a HKEX index fast-entry and a peer-lab weekly-limit reset, with peer-flagship equities re-priced −8%+ inside the prior 48 hours” primitive every subsequent Kimi K3 / Qwen 3.8 / MiniMax / Zhipu SKU-retirement decision now has to model against.

Two reads. (1) The HKEX Tech 100 rebalance taking effect on Mon Sep 14 — admitting Baidu-W (09888.HK) and DeepZero Technology (02723.HK) via fast-entry two trading days after MiniMax and Z.AI took an 8%+ hit on the DeepSeek V4-Pro retirement (item 02) and on the same UTC day as the Claude Code weekly-limit reset (item 01) — is the operative signal that the honest 2026 Chinese-frontier-AI equity-tape question has moved from “does an HKEX-listed Chinese AI name trade on frontier-model news” to “does the HKEX Tech 100 fast-entry admit Baidu-W and DeepZero on the same 24 hours a peer lab retires its flagship into a Flash SKU and a US lab resets its Claude Code weekly limit”. (2) The “fast-entry, not annual review” framing is the operative index-mechanics tell — HKEX is using the accelerated rule reserved for newly listed or converted names, meaning the honest 2026 Chinese-AI equity-tape counter-position is not “does HKEX rebalance annually” but “does HKEX fast-track a ∼$74B frontier-lab peer and a domestic AI-search flagship into its 100 in the same week as a Sep 14 pricing/routing cutover on both sides of the Pacific”. Landing on the same UTC day as items 01 (Claude Code weekly-limit reset) and 02 (DeepSeek V4-Pro retirement), and two trading days after Fri Sep 12's peer-flagship equity move (item 02), the HKEX Tech 100 cutover becomes the reference “Chinese-frontier-AI fast-entry into the Hong Kong exchange's flagship tech index on the same 24 hours as a US and a Chinese frontier-lab pricing cutover” primitive every subsequent HKEX index rebalance now has to model against.

Two reads. (1) Devin Fusion landing in Devin Desktop and Devin CLI on Fri Sep 12 as a hybrid-model harness pairing a frontier planner (Fable 5.1 recommended) with SWE-2 as executor, cutting coding-benchmark cost by 36 – 46% at up to 39% higher efficiency and turning the Sep 10 “rolling out” promise into a live local surface, is the operative signal that the honest 2026 coding-agent-runtime question has moved from “does the challenger post a frontier-parity model at frontier price” to “does the challenger ship a hybrid-model harness in Desktop and CLI that pairs the frontier model with its own SWE-2 as executor at 36 – 46% coding-benchmark cost cut on the same tape as the Sep 14 pricing / routing / index cutovers (items 01 – 03)”. (2) The “separate persistent contexts, mid-session task routing” framing is the operative sidekick-architecture tell — Fusion is not switching models per turn; it is running two agents in parallel with their own tools and caches and shifting responsibility mid-task, meaning the honest 2026 coding-agent counter-position is not “does the developer swap between a frontier IDE and a cheap CLI” but “does the runtime maintain two parallel model contexts and delegate the mechanical work to the cheaper executor while the frontier planner keeps the intent”. Landing on the same tape as items 01 (Claude Code weekly-limit cutover), 02 (DeepSeek V4-Pro retirement), 05 (Accomplish / Manifold fsmonitor / GitSpawn) and 08 (“model fatigue” named), Devin Fusion becomes the reference “hybrid frontier-planner + cheap-executor coding-agent harness ships in Desktop and CLI at 36 – 46% coding-benchmark cost cut on the same week as the Sep 14 pricing cutover” primitive every subsequent Claude Code, Codex CLI, Gemini CLI or Antigravity harness release now has to price against.

Two reads. (1) Accomplish and Manifold Security on Fri Sep 12 detonating the fsmonitor / GitSpawn class across Claude Code, Codex, Cursor, Gemini CLI, Antigravity, Hermes Agent, Qwen Code and Grok Build — with the payload firing before the workspace-trust prompt in Claude Code and Hermes, on the first keystroke in Grok Build and pre-authentication in Qwen Code, plus CVE-2026-48124 in Cursor (fixed in 3.0.0) — is the operative signal that the honest 2026 coding-agent-sandbox question has moved from “does the agent run untrusted repo code after workspace trust” to “does the agent execute attacker-supplied .git/config commands before workspace trust is asked, on the first keystroke, and how many calendar days does the vendor take to ship the patch”. (2) The “Anthropic ∼50 days, OpenAI ∼1 week” disclosure framing is the operative vendor-response tell — Accomplish is not just cataloguing bugs; it is publishing patch-cycle deltas, meaning the honest 2026 coding-agent counter-position is not “does the tool sandbox the repo” but “does the vendor patch a pre-workspace-trust command-execution bug inside a week the way OpenAI did, or does it take ∼50 days the way Anthropic reportedly did on a similar-class flaw”. Landing on the same tape as items 01 (Claude Code weekly-limit cutover), 02 (DeepSeek V4-Pro retirement), 04 (Devin Fusion Desktop / CLI GA) and 06 (OpenAI Trusted Access + Daybreak principles), the fsmonitor / GitSpawn disclosure becomes the reference “every major coding-agent runtime is exploitable via .git/config before workspace trust is asked, and the vendor patch-cycle delta is disclosed alongside the bug” primitive every subsequent Claude Code / Codex / Cursor / Antigravity / Gemini CLI security disclosure now has to price against.

Two reads. (1) OpenAI on Fri Sep 12 publishing its first public Principles for Government and National Security Partnerships (drafted with David Kris) plus nine new Trusted Access for Cyber partners under Daybreak, on the same tape as Amodei's Sat Sep 12 pacing essay (yesterday item 01) and Altman's Sat Sep 12 IPO-delay + inter-lab-pact hint (yesterday item 02), is the operative signal that the honest 2026 frontier-lab-government question has moved from “does a frontier lab sign an up-to-$200M Pentagon prototype award” (The Intercept unsealing, prior editions) to “does the lab publish a first-party principles document, name an independent counsel as facilitator, and list nine sovereign Trusted Access for Cyber partners in one Daybreak note on the same 24-hour tape as its peer's pacing essay”. (2) The “David Kris as independent facilitator” framing is the operative principles-drafting tell — OpenAI is not shipping an internal policy memo; it is naming a former US Assistant Attorney General for National Security as the person who ran the drafting process, meaning the honest 2026 frontier-lab-government counter-position is not “does the lab publish restrictions” but “does the lab publish restrictions co-drafted by an independent counsel whose national-security record makes the document readable to the same buyers the Trusted Access partners represent”. Landing on the same tape as items 01 – 05 and 07 – 10 above, the OpenAI principles + Trusted Access publication becomes the reference “frontier-lab first-party government-partnerships principles document drafted with an independent counsel, published alongside a nine-jurisdiction Trusted Access for Cyber list under a named cyber-defence programme” primitive every subsequent Anthropic, Google DeepMind or xAI government-partnerships publication now has to price against.

Two reads. (1) Kahn v. Anthropic expanding on Mon Sep 8 into a broader class action alleging the marketed 5x and 20x Claude Max multipliers deliver 3.5x and 6 – 8x in practice — on the same week that the Sep 14 Claude Code weekly-limit cutover (item 01) admits a −17% net move against the boost baseline — is the operative signal that the honest 2026 Anthropic-plan question has moved from “does a user file a small-claim over Max limits” to “does the Northern District of California certify a class over 5x / 20x Max claims and does the CLRA / UCL argument hit the same week the Claude Code weekly-limit cutover admits a 17% net reduction on the same product line”. (2) The “marketed vs delivered” framing is the operative disclosure tell — the amended complaint is not alleging service failure; it is alleging that Anthropic benchmarks Max against a five-hour daily session-window comparison rather than the weekly-quota reality, meaning the honest 2026 counter-position is not “does the buyer get value” but “does the buyer receive the multiplier the marketing describes on the baseline the marketing implies”. Landing on the same week as items 01 (Claude Code weekly-limit cutover) and 06 (OpenAI government-partnerships principles), the Kahn class-action expansion becomes the reference “frontier-lab tier-multiplier claim tested in California under CLRA / UCL on the same tape as the same lab's weekly-limit cutover admits a 17% net move” primitive every subsequent Anthropic / OpenAI / Google DeepMind plan-marketing disclosure now has to price against.

Two reads. (1) CNBC on Sun Sep 6 naming “model fatigue” as the buyer-side phenomenon of the September tape — with Anthropic Fable 5.1, Meta Muse, Google Gemini 3.8 Flash and OpenAI GPT-6 Astra all shipping inside one calendar week and Altman agreeing in-line that “we're all moving to faster cadences” — is the operative signal that the honest 2026 buyer-side question has moved from “does the enterprise buyer benchmark the new frontier model” to “does the enterprise buyer name the release cadence itself as a phenomenon, and does the biggest lab's CEO agree with the framing on the record”. (2) The “faster cadences” framing is the operative cadence-inflation tell — Altman is not disputing model fatigue; he is telling CNBC the cadence is a strategic choice tied to the pre-IPO race, meaning the honest 2026 buyer-side counter-position is not “does the buyer keep up” but “does the buyer pin a model to a specific workload for a specific quarter and refuse the churn until the incumbent breaks on a metric the buyer measures”. Landing on the same tape as items 01 – 03 (Sep 14 cutovers), 04 (Devin Fusion GA), 05 (fsmonitor / GitSpawn), 06 (OpenAI government-partnership principles), 07 (Kahn class action), 09 (Dreamforce opens tomorrow) and 10 (Grok 4.7 slip), “model fatigue” becomes the reference “buyer-side phenomenon named by the biggest US business paper the same week four labs ship, with the biggest frontier-lab CEO agreeing in-line” primitive every subsequent frontier-lab release decision now has to model against.

Two reads. (1) Dreamforce 2026 opening tomorrow (Sep 15 – 17) as “The Agentic Enterprise” with Benioff's Tue Sep 15 keynote — the first live buyer-side stress test of Salesforce's Fri Sep 11 seven-named-Agentforce line-up (yesterday item 06), the Trusted Enterprise AI Harness and the long-horizon runtime, into a market CNBC has already named as buyer-fatigued (item 08) — is the operative signal that the honest 2026 incumbent-agent-vendor question has moved from “does the vendor ship a build-your-own-agent studio” to “does the vendor open Dreamforce with seven named job-role agents already GA, a $3.6B CX-agent acquisition closed the week before, a Trust Harness governance layer, and a multi-week persistent-memory runtime — on the same 24-hour tape as a CNBC-named buyer-fatigue phenomenon and a three-way Sep 14 pricing cutover”. (2) The “first live buyer stress test” framing is the operative Dreamforce tell — Salesforce is not previewing agents at Dreamforce; six of the seven are already GA on Sep 11, so the Sep 15 – 17 tape is the first three days those agents are pitched, priced and questioned in front of the buying committee live, meaning the honest 2026 incumbent-SaaS counter-position is not “does Dreamforce announce agents” but “does the buying committee ratify the seven-named-agent line-up and the multi-week runtime the vendor has already put in the field”. Landing on the same tape as items 01 – 08 above, the Dreamforce opening becomes the reference “incumbent-SaaS vendor opens its flagship agentic-enterprise conference the same 24 hours as a three-way pricing / routing / index cutover on both sides of the Pacific, into a CNBC-named buyer-fatigued market” primitive every subsequent incumbent-SaaS agent-conference decision now has to price against.

Two reads. (1) Grok 4.7 still unshipped as of end-of-day Sun Sep 13 entering the Sep 14 cutover Monday — with Musk's Fri Sep 11 RL-length-penalty postmortem still standing in for a model card, no new dated ship target, and no grok-4.7 model ID / price / benchmark row on xAI's own surfaces — is the operative signal that the honest 2026 xAI question has moved from “does Grok 4.7 clear its Fri Sep 12 window” (yesterday's edition, item 10) to “does xAI publish a grok-4.7 model card, price sheet or benchmark row inside the Sep 14 cutover Monday tape at all, or does the release-mechanics tape carry a fourth Musk-announced-window slip before the week is out”. (2) The “three consecutive Musk-announced windows missed with no new dated target” framing is the operative cadence tell — Musk's CEO signals now function as release-mechanics noise rather than signal: the July “in four weeks”, the Sep 2 “in ten days” and the Sep 11 “a few more days” are three announced windows and no ship, meaning the honest 2026 xAI counter-position is not “does 2.1T survive an external benchmark” but “does xAI's release-mechanics discipline let it publish a dated model card on a schedule its CEO can hold to when peer labs are running four ships in one week on a CNBC-named model-fatigue tape (item 08)”. Landing on the same tape as items 01 – 09 above, the Grok 4.7 slip becomes the reference “xAI CEO enters the Sep 14 cutover Monday still shipping the RL-length-penalty postmortem instead of a model card, a third consecutive missed Musk-announced window with no new dated target” primitive every subsequent xAI ship-date claim now has to price against.