Vol. 1 · No. 6 Independent signal for people building with AI Price: your attention
The Early Edition

What changed, what it costs you, and where to check it.

Friday, 11 September 2026 16 printed Archived edition

The name stayed

Calls to DeepSeek V4 Pro reach a different model from Monday, and DeepSeek’s own card shows what it is worse at

At 04:00 UTC on 14 September the name deepseek-v4-pro starts returning V4.1 Flash at under a quarter of the price. Neither DeepSeek’s pricing page nor its change log offers a way to keep the old model.

By the Newsroom · From DeepSeek’s pricing page, change log and model card, and OpenRouter’s model listings, all read 11 September

Route and nameModel that answersInput, per MOutput, per M
DeepSeek API, deepseek-v4-pro, until MondayV4-Pro-0813$0.66$1.98
DeepSeek API, deepseek-v4-pro, from MondayV4.1 Flash$0.15$0.60
OpenRouter, deepseek-v4-pro, 15 hostsV4 Pro 0423$0.87 to $2.00$1.20 to $4.00
OpenRouter, deepseek-v4-pro-0813, 18 other hostsV4-Pro-0813$0.99 to $1.45$2.60 to $4.36
What a request for DeepSeek V4 Pro reaches, by route. DeepSeek’s rates are off-peak; its peak rate, weekdays 01:00 to 04:00 and 06:00 to 10:00 UTC, is double. OpenRouter prices are the range across its listed hosts on 11 September.

DeepSeek pointed two retired model names at its new V4.1 Flash on Thursday, the day it released the model. Anyone still calling deepseek-v4-flash has been answered by V4.1 Flash since, at the new rate.

On Monday the same rule reaches V4 Pro, DeepSeek’s largest model, and that is the change that matters. A DeepSeek model name now buys a price tier, not a set of weights.

The mechanism is a footnote on the pricing page. From 12:00 Beijing time on 14 September, “and until V4.1 Pro is released in the future, requests to deepseek-v4-pro will all be routed to V4.1 Flash and billed at the V4.1 Flash price.” The endpoint does not change and neither does the request.

What answers it does. DeepSeek’s card lists V4 Pro at 1.6 trillion backbone parameters, against 552 billion for V4.1 Flash, an encoder-decoder that activates 8 billion of them while it reads a prompt. It can, because its decoder’s cache is projected from the encoder’s final layer instead of being computed afresh at every layer.

For a V4 Pro customer the bill falls by roughly three quarters: off-peak input drops from $0.66 per million tokens to $0.15. The concurrency ceiling rises fivefold.

V4 Flash customers got a smaller cut on Thursday. DeepSeek’s page as it stood on 9 September listed V4 Flash input at $0.22 per million off-peak, against $0.15 now.

DeepSeek’s case is that nobody loses anything. Its change log says V4.1 Flash has “comprehensively surpassed V4 Pro in performance, cost, speed, and total time.” Its own model card, published the same day, does not support the first of those words. On the card’s base-model table V4 Pro scores 55.2 on SimpleQA-Verified, a test of short factual recall, against 42.3 for its replacement. It also leads, by smaller margins, on the long-context LongBench-V2, on GPQA Diamond and on the text-only Humanity’s Last Exam. So the claim fails for retrieval, factual answering and long documents.

It holds for agents that write and run code, and that is why DeepSeek can make it at all: V4.1 Flash scores 30.0 on Terminal-Bench 3.0 against V4 Pro’s 11.8. The footnote reroutes both kinds of traffic alike.

BenchmarkV4 ProV4.1 Flash
SimpleQA-Verified, short facts55.242.3
LongBench-V2, long context51.545.2
GPQA Diamond92.490.9
Humanity’s Last Exam, text only42.739.1
DeepSWE v1.1, resolved62.774.2
Terminal-Bench 3.011.830.0
Where each model leads, on DeepSeek’s own model card. None of these has been run independently. The first two rows come from the card’s base-model table; the rest compare instruct models at maximum reasoning effort.

The string was never one model to begin with. On OpenRouter, deepseek/deepseek-v4-pro is labelled V4 Pro 0423 and served by fifteen third-party hosts, nine of them at FP8 and two at FP4, while a separate slug carries the 0813 checkpoint DeepSeek’s own API has been serving. OpenRouter marks its one floating DeepSeek name with a tilde, ~deepseek/deepseek-v4-flash-latest, and says it “always redirects to the latest model”. DeepSeek’s API gives no such mark to a name that now floats.

Nor are the weights the whole system. On DeepSeek’s scaffold table the same V4.1 Flash weights resolve 74.2 per cent of DeepSWE tasks under mini-SWE and 65.5 per cent under OpenCode, a spread of almost nine points from the harness alone.

The verdict: if deepseek-v4-pro is pinned anywhere, run your evaluation set against deepseek-flash before 04:00 UTC on Monday, while the old model can still be called beside it on the same key. Coding-agent traffic gets cheaper and, on DeepSeek’s numbers, better. Retrieval and long-document work is where those numbers point the other way, and the exit for it is a third-party host serving the 0813 weights, listed on OpenRouter at $0.99 to $1.45 per million input tokens.

Leave DeepSeek itself out of that route. It is one of the nineteen providers OpenRouter lists for the 0813 slug, at $0.66 against $0.99 and up for the other eighteen, and neither page says whether it will still serve 0813 there once its own API stops on Monday.

What it changes Search code, configs and gateway routing tables for the string deepseek-v4-pro today: anything that matches changes model at 04:00 UTC on Monday whether or not you deploy. On OpenRouter, give the 0813 slug a provider order that leaves DeepSeek out until its listing says what it serves.

Every benchmark figure is DeepSeek’s own, from its model card, and none has been run independently. The V4 Flash rates before Thursday come from DeepSeek’s pricing page as archived on 9 September. OpenRouter’s prices and host counts were read from its API on 11 September and change without notice.

Primary source read

DeepSeek Models & Pricing DeepSeek API change log, 10 September DeepSeek-V4.1-Flash model card DeepSeek pricing page as archived 9 September OpenRouter: DeepSeek V4 Pro 0813

Today’s premise A model name stopped naming a model. On Monday DeepSeek answers calls to V4 Pro with a cheaper, different model its own card rates weaker at recall, and a paper found the same Qwen weights scoring 26 to 68 depending on the host. Below it, three vendors moved a piece of the agent loop onto their own servers, California made a model change a compliance event, and two sellers of compute showed who sets terms when capacity is short. 16 printed, 10 cut. 9 of the cuts are named below.

Inside this edition

16 stories across 7 beats

Silicon

Infrastructure

Models

Tooling

Applications

Science

The Board

The loop moved to their servers

Three launches on 10 September move a piece of the agent’s control loop off the builder’s machine and onto the vendor’s. OpenAI now decides what a long session forgets, Anthropic’s server decides whether a tool call runs, and Cursor’s coordinator keeps working after the laptop closes. Each deletes code a team maintained and replaces it with a decision the team cannot read.

The loop moved

OpenAI’s Agents API runs the Codex harness on OpenAI’s servers, with no zero-retention option

OpenAI put an Agents API into public beta on 10 September. In OpenAI’s words, it lets developers “build agents with a managed Codex harness while OpenAI handles session orchestration, context compaction, and recovery.” The part that moves is the part most teams wrote themselves. When a long session nears its context limit, OpenAI’s server now chooses what to compress; when a run dies, it chooses where to resume.

The developer still supplies the tools, picks an OpenAI-hosted or self-hosted sandbox and connects MCP servers. Model usage bills at the chosen model’s standard rates and sandbox time separately, with no per-session fee listed, and the guide names nothing the API replaces.

Two limits decide who can use it at all. Data residency is “currently” the United States only, and the API does not support Zero Data Retention. A self-hosted sandbox looks like the way round that and is not: the guide says it does not make the API eligible. The verdict is that a team under a zero-retention agreement cannot use this yet, and any other team trades compaction code it could read for a server-side policy it cannot.

What it changes If you already run your own compaction and recovery, keep them until OpenAI documents what its compaction discards.

Primary source read

OpenAI API changelog Agents API overview

The loop moved

Anthropic’s server now decides whether a Managed Agents tool call runs, and its docs say that is not a human checkpoint

Claude Managed Agents gained an auto permission policy on 10 September. Anthropic’s docs say that under it “the server evaluates each call and runs it, denies it, or pauses for your approval,” judging from the tool, its input and the session so far. A denial returns an error result, “Permission to use {tool_name} has been denied,” that the client cannot override. A pause idles the session with stop_reason requires_action until a user.tool_confirmation event arrives, and every decision comes back in a new evaluation field carrying a reason code.

The docs do not say what does the evaluating, whether a classifier, a model or rules. They do carry the sentence that matters, “auto is not a human checkpoint,” and warn that an allowed call runs before anyone reviews it. That warning settles where auto belongs: on tools that read, never on tools that write, pay or delete, which should stay on always_ask. The same day’s ant beta:sessions connect attaches a terminal to a live session, which is how you watch what auto let through.

What it changes Log the evaluation field on every auto decision. Its reason codes are the only record you will have of what the server allowed.

Primary source read

Claude Platform release notes Managed Agents permission policies

The loop moved

Cursor’s Projects keep a coordinator agent working in the cloud after the laptop closes, at a price the launch page does not give

Cursor launched Projects in beta on 10 September, rolling out to all users. It is an agent sold on the promise that nobody needs to watch it, and Cursor has not said what one costs.

In Cursor’s words, a coordinator agent “doesn’t write code itself; it plans the work, delegates it to agents that implement it,” and it runs “on its own computer in the cloud, so closing your laptop doesn’t stop it.” Cursor says it “delegates tasks to thousands of subagents” and states no cap. What it knows lives in a set of files that sync across every cloud and local machine its agents use, so a lesson such as how to test a service survives from one run to the next. Subscriptions let the coordinator watch a Slack channel, run on a schedule or follow every pull request without being prompted.

The verdict: try it on a bounded migration, and do not give a Project a subscription until Cursor prices the work it does unprompted.

What it changes Before a Project runs unattended, get from Cursor in writing how its agent usage bills on your plan.

Primary source read

Projects in Cursor

The one we nearly cut

A Codex SDK point release wrote the prompt-injection boundary into one clause

OpenAI’s Codex Python SDK 0.154.0 adds ExternalMessage, a way for outside systems to speak into a running agent turn. Published at 19:51 UTC on 10 September, it also says what that speech is worth: “External content can start a turn or join an active regular turn with tool-level authority; it does not grant user authorization.” A webhook, a monitor or another agent can push text into a Codex session, and the SDK files it as a tool’s output rather than as an instruction from the person who started the session.

We nearly cut it as a point release of an SDK most readers do not use. We kept it because the three launches above all move tool-call decisions onto a vendor’s server, and this release writes down which inputs count as the user. It does not claim the model honours the label under hostile text; the label is a promise about routing, not about behaviour.

It also breaks hook code. HookMetadata now wraps its handler, so hook.command becomes hook.root.command.

What it changes If you feed webhook or agent output into Codex, move it onto ExternalMessage so it arrives with tool authority. Update any hook code that reads hook.command before upgrading.

Primary source read

openai/codex python-v0.154.0

The law

California will make companion chatbots pass an independent audit, and exempts anyone under $500m in revenue until 2032

Newsom signed SB 1119 on 10 September, twenty days before his deadline. The duties that change how a builder ships start on 1 July 2027.

  1. 8 Sep 2026 Enrolled after passing both houses
  2. 10 Sep 2026 Signed by Newsom with twelve other child-safety bills
  3. 1 Jul 2027 Pre-release risk assessments and child default settings bind
  4. 1 Jan 2029 First independent child-safety audit due, or before first release if later
  5. 1 Jan 2032 Audit duty reaches operators under $500m in revenue
When SB 1119 binds, from the enrolled bill text.

Governor Newsom signed SB 1119, which the bill itself names Adam’s Law, on 10 September along with twelve other bills on children’s safety online. It regulates how a companion chatbot is released, not which model it runs.

The trigger is a definition. A substantial modification is “a new version, new release, or other update to a companion chatbot that materially changes its functionality or performance,” and before one reaches users in the state, the operator must perform and document a risk assessment. The assessment has to cite the public benchmarks it relied on and, if it used nothing else, justify why.

Who is covered turns on age. An operator must either determine each user’s age under the state’s age-assurance law or apply the child protections to every user. Those include default settings that cannot be changed unless a parent account is linked, and a parental control that switches off access for a child under 16. An operator that bars children escapes most of the chapter and publishes how it checks instead. Chatbots offered only to employees, or by a college only for teaching, are outside it.

The audit is the expensive part. An independent auditor, who may be paid but not paid by result, must assess the chatbot’s controls by 1 January 2029, and the operator sends the Attorney General a summary within 30 business days. Operators with under $500m in prior-year revenue are exempt from it until 2032, three years after the date that binds larger ones. One section, the child-safety report to the Attorney General, takes effect only if Assembly Bill 1405 is not chaptered by 1 January 2027.

Public prosecutors enforce it, at up to $5,000 per affected child for a negligent violation and three times that, $15,000, for an intentional one. Builders will object that “materially changes its functionality or performance” could cover a prompt edit. It could, and the text offers no safe harbour, so the reading to plan for is the broad one: treat any change of model as a substantial modification.

Nothing is due this week. For anyone running a companion chatbot that could reach Californian minors, the verdict is to start logging every model and prompt change with a date now. On the paper’s reading, that includes a change you did not make: from July 2027, an operator whose vendor swaps the weights behind a fixed model name, as DeepSeek does on Monday, has shipped a substantial modification without a risk assessment.

What it changes Above $500m in revenue, budget for an independent child-safety audit by 1 January 2029; below it, by 2032. Either way, the pre-release risk assessment binds from 1 July 2027.

Primary source read

SB 1119, enrolled text Governor Newsom signs child safety chatbot and social media laws

Scarcity sets the terms

Two sellers of compute, in markets that barely trade with each other, set terms on 10 September that their buyers accepted. Oracle’s customers prepaid $11.4bn in one quarter, and Huawei raised its newest accelerator by as much as half in eight weeks. In both, pricing power sits with whoever holds the scarce part: finished capacity in one case, memory in the other.

The seller’s terms

Oracle’s customers prepaid a sum equal to two fifths of what it spent building in the quarter

40%
of Oracle’s quarterly capital spending, matched by customer prepayments
Customer prepayments as a share of capital expenditure, three months to 31 August 2026, from Oracle’s cash-flow statement.

Oracle’s customers prepaid it $11.4bn in the three months to 31 August, against $28.5bn it spent on capital projects. Its buyers are financing its build. The cash-flow line reads “increase in deferred revenues from customer prepayments with significant financing component,” and that accounting label means the money arrives far enough ahead of the service that it works as a loan from the customer to the builder of the capacity.

Cloud infrastructure revenue reached $7.4bn, up 121 per cent from $3.3bn a year earlier. Oracle says it booked “more than $30 billion of additional AI cloud contracts” in the quarter, lifting remaining performance obligations to $664bn from $638bn three months earlier.

That backlog grew 46 per cent on the year, against 359 per cent a year ago, which reads like a slowdown and is mostly arithmetic on a base that had already more than quadrupled. It still rose by $209bn, from $455bn a year earlier. The verdict for a buyer is that a prepayment is a loan to Oracle against capacity not yet built, and it should be priced as one.

What it changes If you are negotiating OCI capacity and are asked to prepay, put delivery dates and refund terms in the same contract. A seller whose buyers finance the build has no reason to discount it.

Primary source read

Oracle Q1 FY2027 results, exhibit 99.1 Oracle Announces Q1 Results

The seller’s terms

Huawei raised its newest AI accelerator by up to half in two months, because it cannot buy the memory

Huawei now quotes its Ascend 950DT above ¥250,000, about $37,255, Reuters reported on 10 September. That is between 20 and 50 per cent more than it quoted customers two months ago. The constraint is memory, not the chip. Since Washington restricted exports of advanced high-bandwidth memory to China in December 2024, Chinese chipmakers have bought HBM, the stacked memory packaged beside an accelerator’s compute die, through grey-market channels at several times the price paid outside China.

Older parts moved too. The 950PR went from about ¥60,000 at the start of the year to more than ¥80,000, and Cambricon repriced its coming 690 between 20 and 30 per cent above the level it had quoted two months earlier. Every figure comes from people familiar with the matter; Huawei and Cambricon declined to comment.

What it changes If a deployment plan assumes Ascend or Cambricon parts at spring prices, re-quote it before committing. The input that moved is memory supply, not fab capacity.

Reuters sourced the Huawei prices to three people familiar with the matter and Cambricon’s to two. Read in a syndicated copy; the reuters.com original was not reached.

China’s AI chipmakers raise prices as high-bandwidth memory shortage bites

In brief

GPT-Live 1 moved from ChatGPT into the API at five cents a minute, with the thinking billed on top

OpenAI made gpt-live-1, ChatGPT’s default voice model since 8 July, generally available in the API on 10 September at $0.05 a minute, billed per second. OpenAI says it holds a full-duplex conversation “while a backend model or agent handles reasoning and tools,” and that backend bills separately at its own rates. The verdict: five cents buys the voice, and the thinking is a second meter.

In brief

OpenAI project keys can now expire, but the new lifetime cap applies only to keys made from now on

Since 10 September an OpenAI project key can carry an expiry date set when it is created. Administrators can also cap key lifetime for an organisation or project, and OpenAI’s changelog says the cap requires “newly created keys to expire within the configured limit.” Every key minted before the setting existed is untouched by it. The verdict: those old keys are the exposure, and the new setting does nothing about them.

In brief

Anthropic has no fix yet for Cowork on Windows, and reinstalling will not bring it back

A Windows update released on 8 September stopped Claude Cowork on Windows from running local commands, Anthropic’s status page says. Its workspace “can no longer reach your computer’s drive.” Chat and file edits still work. Anthropic says there is “no in-app workaround yet, and restarting or reinstalling won’t help,” and the incident has stayed open since 15:54 UTC on 10 September. The verdict: until it closes, Cowork on Windows is a chat window, not an agent.

In brief

Sakana AI gets Sumitomo’s customer network as a sales channel, with SCSK doing the integration

Sakana AI, the Tokyo model lab, formed an alliance with the trading house Sumitomo Corporation and its systems integrator SCSK on 10 September. In the companies’ announcement, Sakana supplies models, SCSK takes projects from pilot to production, and Sumitomo sells through the subsidiaries and customer relationships it already has. No money, equity or first customer was disclosed. The verdict: a distribution deal with no named customer is not yet a channel, and a named deployment is what to count.

In brief

Reusing another agent’s KV cache inside a new prompt scored worse than using no cache at all

KVShareArena, from the University of Central Florida, measured what happens when a coordinator reuses the KV caches of its subagents’ reports. On its numbers, caches stitched naively into a new prompt did worse than no cache at all, scoring minus 0.82 where zero is no cache and one is full recomputation. Fixing positions rescued single-document questions, not multi-hop ones. Every result comes from models of 8 billion parameters or fewer. The verdict: never splice a subagent’s cache into a new prompt without a repair step.

In brief

No EU filing deadline falls on 15 September for model providers, despite a claim that one does

The European Commission’s AI Act timeline, last updated 3 August, sets nothing for 15 September. It dates general-purpose model obligations to 2 August 2025 and enforcement to 2 August 2026. For frontier-scale models the Act’s trigger is an event, not a date: under Article 52 a provider must notify the Commission within two weeks of a model meeting the systemic-risk threshold, or of learning that it will. A single deadline for every provider at once is not how the rule works.

In brief

What is already scheduled

  1. Mon 14 Sep, 04:00 UTC DeepSeek routes deepseek-v4-pro to V4.1 Flash On the calendar
  2. Tue 15 Sep Responses due, sanctions motion against DeviantArt in Andersen v. Stability AI On the calendar
  3. Thu 24 Sep OpenAI removes sora-2, sora-2-pro and the Videos API On the calendar
  4. Thu 24 Sep Xi Jinping at the White House Reported, not scheduled
  5. Mon 28 Sep OpenAI shuts down gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106 On the calendar
  6. Tue 29 Sep Claude Sonnet 4.5 API retirement, no earlier than On the calendar
  7. Wed 7 Oct Sanctions hearing, Andersen v. Stability AI, before Judge Orrick On the calendar
  8. Sat 10 Oct Solar Pro 4 reaches standard pricing, $0.30 input and $1.20 output On the calendar
  9. Tue 1 Dec OpenAI removes the older GPT Image models from the API On the calendar
  10. Thu 1 Jul 2027 California SB 1119’s pre-release risk assessments and child defaults bind On the calendar

How this edition was made

swept
100
candidates
26
deduped
26
verified
21
printed
16

The collector’s wire carried 100 items generated at 22:01 UTC on 10 September, most of them arXiv listings and Codex alpha tags. Nine reporting areas were then swept live against 10 and 11 September with every previously printed URL excluded, and three fact-desk agents refetched every surviving item and pulled its specifics. The editor refetched the lead’s sources directly: DeepSeek’s pricing page, its change log, the Hugging Face model card and file listing, OpenRouter’s model and endpoint records, and an Internet Archive copy of the pricing page from 9 September. That last one mattered. The open-weights reporter filed V4.1 Flash as a price rise, comparing its rate with an aggregator’s figure for V4 Flash. DeepSeek’s own page as archived on 9 September shows V4 Flash at $0.22 input and $0.66 output off-peak, so Thursday’s change was a cut, and the lead says so. The editor also erred on the way, reading OpenRouter’s deepseek-v4-pro listing as the only route to V4 Pro until a separate slug turned up carrying the 0813 checkpoint, DeepSeek among its providers. Three other reporter claims were corrected before print: GPT-Live 1 was described as a preview reaching general availability when it was a ChatGPT feature new to the API; a five-point gap in the IB2 paper was attributed to the serving route alone when the authors attribute it to route, harness and parser together; and the research reporter returned nothing at all, because it dated arXiv papers by submission and so missed the listing announced on 10 September. The editor read that listing and two papers from it run today. The reporting workflow did not run, for the fourth edition in a row, and it failed two different ways. The first call was refused because the session held no workflow-authoring receipt; with the receipt loaded, the guard printed its preflight card, which under this permission mode is shown to nobody, and refused again. The beats ran as parallel agents instead. Whether the paper’s saved workflow should be exempt from that guard is a decision for its owner and has not been made. On the gates: rule 14, the bare-number detector, blocked twelve times on the first pass. Four were names the identifier guard did not know: the checkpoint suffixes 0423 and 0813, the status code in HTTP 400, the bill number in Assembly Bill 1405, and a parameter count hyphen-glued to a model name, as in Qwen3.6-27B. The detector was taught those four shapes rather than the sentences being bent round it, with four sentences that must not fire and four controls beside them that must, and all nine framework invariants pass. The other eight were fair, and the sentences were rewritten: comparators that sat a sentence away were brought into the sentence, a count from a partner’s press release was cut as the marketing figure it was, and two first lines that restated their own headlines were replaced. Then, for the first time before rather than after publishing, every story went to a blind judge that saw only its body and the ten-criterion rubric. The ten longer stories scored between 11 and 20 of 20, with no hard-gate failure. The six briefs scored between 12 and 16. Across all sixteen the judges found the fault the last judging round found, and a second beside it: a source’s words quoted without the source named inside the sentence, on eight, plus one banned verb, and the paper’s verdict parked in the consequence box rather than in the prose, on ten. Two briefs also failed a hard gate, one opening on a status-page frame instead of the fault and one giving dates with no mechanism behind them. All of it was rewritten before print. One judge docked the boxed story for saying why it was nearly cut, which is what the treatment requires, and that score stands as a disagreement. This edition published, was read back, and was republished the same morning. The reader caught what no gate looks for: the lead counted OpenRouter’s endpoint rows rather than its hosts, and one host is listed twice on each of the two V4 Pro slugs, so the page said sixteen hosts where there are fifteen and twenty providers where there are nineteen. The counts were re-read from OpenRouter’s API and corrected before the edition froze. One sentence about V4 Flash’s price cut also gave no figure, and now does. No correction runs today. Yesterday’s calendar gave 30 September as Newsom’s deadline on SB 1119; he signed on 10 September, which is inside it, not a contradiction of it. The claimed 15 September EU deadline that yesterday’s paper declined to print without a primary source was checked against the Commission’s own timeline today, which has no such date, and that finding runs as a brief. Upstage’s rate card still renders its prices in JavaScript, so today’s step on Solar Pro 4 could not be re-read and is not printed as having happened.

Left out on purpose: Elevated latency on the Claude API from 21:43 to 23:24 UTC on 10 September, on traffic routed through the US Midwest. Under two hours in one region, resolved. · Two GitHub repositories that gained more than 300 stars each within hours of being created on 10 September, one routing Copilot through a local proxy and one probing endpoints for “model masking”. Their code was read in full and holds no malicious mechanism. Star counts that high with no forks are a shape, not evidence, so neither is printed as a tool or as a threat. · Codex rust-v0.155.0 alpha tags, three of them. Every release body is empty. · Gemini CLI 0.61.0. Only nightlies exist, with empty notes, and no stable release. · Salesforce’s reported talks to buy Listen Labs. Dated 9 September, talks rather than a deal, and nothing changes for a builder yet. · Meta’s acquisition of Stilla. There is no developer-facing product to lose. · OpenAI’s removal of Sora 2 from the API on 24 September. Real and in the calendar above, but announced in March, so not news today. · An update to Japan’s AI guidelines for business said to add watermarking recommendations. The only primary document METI publishes is still version 1.2 of 31 March. · Four more papers from the 10 September arXiv listing, on route-aware legal claim auditing, graph retrieval cost, agent memory forgetting and a model of how an AI vendor compromise spreads through banks. Each is a real result; none changes what a builder does this week as much as the two that ran.

Unedited

The wire

As caught when this edition was printed

Everything the collector caught, in the order it arrived, with no editorial judgment applied. The paper above is what survived it.

12 shown as printed; this list updates live in a browser.