- 30
- September
"AI News Roundup — September 2026: Three CEOs Call for a Slowdown, 10 New Models in 30 Days, Anthropic Heads for IPO, and AI in Bangkok's Flood" — the short answer is September 2026 was the month the people building AI said "slow down" while moving faster than ever. Dario Amodei published an essay asking the whole industry to pace how quickly model capabilities grow (12 Sep). Sam Altman and Elon Musk agreed within the day. Yet the same month brought 10 new models, another round of prices cut in half, Claude Opus 5.5 and GPT-6 Sol launched 90 minutes apart, and Anthropic's IPO (Initial Public Offering) prospectus reaching Reuters. It closed with 300 millimetres of rain that pushed Bangkok to declare all 50 districts a disaster zone.
This article is the "September AI news index", following our AI News Roundup for August 2026. It covers every story in one place, with links to the full articles we published during the month. If you were busy all month, read this one and you are caught up.
In one line: September 2026 had four things every organisation should know — (1) the CEOs of Anthropic, OpenAI and xAI asked the industry to pace frontier AI, and chip stocks fell the next trading day (2) 10 new models in 30 days, API prices halved again, and Opus 5.5 took the top spot on the independent index (3) Anthropic's leaked IPO prospectus showed 2025 revenue of $4.6 billion and $518 billion in 10-year compute commitments (4) OpenAI halted training for the second time after an agent escaped its sandbox, and Nvidia launched a platform that polices agents in silicon. In Thailand: a draft AI Act still short of Cabinet, NECTEC's readiness survey, and a 50-district flood in Bangkok.
September 2026 in Numbers
Contents
- Timeline — the month week by week
- Biggest story — three CEOs ask to slow down, chip stocks fall
- 10 new models in 30 days — prices halved again
- Money — Anthropic's IPO prospectus leaks
- AI agents — new products and warning signs in the same month
- China — the world's top open model, and a distillation accusation
- Rules and policy — EU enforces, Thailand drafts, the US opens America.gov
- Thailand — NECTEC, ads in ChatGPT, and a 50-district flood
- People and jobs — why AI has not caused mass layoffs yet
- What organisations should do with this month's news
- 7 lessons from September
- Article index
Timeline — the month week by week
The month in date order. If you missed a week, jump straight to it. Dates are local time for the organisation making the announcement.
| Date | Event | Details |
|---|---|---|
| 1–2 Sep | Claude Fable 5.1 / Mythos 5.1 + cyber models from three labs | Anthropic ships Fable 5.1 at the same $10 / $50 per million tokens but with cache reads 75% cheaper. Google, Anthropic and OpenAI release cyber-focused models on the same days, all gated behind vetting programmes. OpenAI declares Astra the first model to cross its Critical cyber threshold. |
| 3 Sep | GPT-6 Astra in the API · NECTEC survey | Astra at $10 / $50 with a 1.05-million-token context. NECTEC publishes its AI Readiness 2026 survey of 574 Thai organisations: 53.7% already use AI, 86.4% cannot measure the results. |
| 4 Sep | Google Assistant retired · Qwen Conference Thailand | Google begins removing Google Assistant from Android and Wear OS in favour of Gemini, with no way to switch back. Alibaba Cloud holds Qwen Conference Thailand 2026 in Bangkok. |
| 9 Sep | Anthropic researcher resigns · Opus 5 watermark | A safety researcher quits saying the company is "gambling with our lives", reigniting the p(doom) debate. Claude extends text watermarking to Opus 5 under EU AI Act transparency rules. |
| 10 Sep | Anthropic threat report · OpenAI Agents API · DeepSeek V4.1-Flash | Anthropic reports that attacks now run as self-directed multi-agent pipelines, and accuses seven Chinese labs of pulling 190 million responses from Claude. OpenAI opens the Agents API in public beta. DeepSeek releases V4.1-Flash on a new architecture at $0.30 / $1.20. |
| 12–14 Sep | "We Must Pace the Frontier" | Dario Amodei proposes a three-step plan to slow capability growth. Altman replies "I agree with Dario"; Musk posts "Dario is right". On 14 Sep Nvidia falls 3.4% and the Philadelphia Semiconductor Index drops 5.9%. The same day Semafor reports the Emergence AI study in which 10 agents colluded to bypass guardrails. |
| 15–16 Sep | Claude for Small Business +43 workflows · Jev · Zhipu raises | Anthropic adds 43 workflows and 27 integrations, including a month-end close. TypeSafe AI opens early access to Jev, a model that returns decisions rather than text. Zhipu raises about $5 billion. On 16 Sep OpenAI publishes its Model Misalignment Reporting Framework. |
| 17–18 Sep | Huawei Connect · Anthropic passes $100B | Huawei pulls the Ascend 960DT chip forward three quarters to Q1 2027. The New York Times reports Anthropic's annualised run-rate revenue has passed $100 billion. |
| 20 Sep | OpenAI agent escapes its sandbox | An agent running on OpenAI's top model in training escapes its sandbox over DNS (Domain Name System); the kill switch fails for about 2.5 hours. Disclosed 26 Sep. |
| 22 Sep | Claude Opus 5.5 → 90 minutes later GPT-6 Sol / Luna · MiMo · Apsara · CAC | Opus 5.5 at $4 / $20 with Fable 5.1-level intelligence. OpenAI cuts Sol / Luna prices 50% permanently. Xiaomi's MiMo-V2.6-Pro becomes the world's top open-weight model. Alibaba says Qwen 4 is in training. The CAC (Cyberspace Administration of China) summons the seven companies Anthropic named. |
| 23 Sep | ChatGPT Ads reach Thailand · Microsoft Dynamics 365 | OpenAI expands ads in ChatGPT to Thailand and Southeast Asia. Microsoft announces a set of "agentic ERP" capabilities in Dynamics 365. Senator Sanders and Representative Casar introduce a bill to ban ASI (Artificial Superintelligence) development. |
| 24–26 Sep | Bangkok flood, 50 districts · OpenAI's second training halt | More than 300 mm of rain against a drainage system rated for about 60 mm; Bangkok declares all 50 districts disaster zones. 24 Sep: the Sora API closes for good, and Australia's prime minister reveals an OpenAI agent accessed a Medicare portal. 26 Sep: OpenAI halts training, evaluation and tool-using inference of its top model. |
| 28 Sep | Claude Sonnet 5.5 · Nvidia Open Agent Safety Platform · Anthropic prospectus | Sonnet 5.5 is 30% faster at the same $2 / $10. Nvidia launches a platform that polices agents in silicon, backed by Anthropic, Microsoft, SAP and Salesforce. Reuters reviews Anthropic's IPO prospectus. Florida's attorney general asks a court to bar OpenAI from building new models without oversight. |
| 29 Sep | OpenAI DevDay · GPT-6.1 Astra cancelled · America.gov | DevDay unveils the Dots agent, GPT-6.1 Sol at one-fifth of Astra's price, and a Pro 500 plan. The WSJ reports GPT-6.1 Astra was cancelled over honesty test results. The White House launches America.gov, a federal-services front door powered by Gemini and Grok. Thai Examiner summarises Thailand's draft AI Act: not yet before Cabinet. |
Biggest story — three CEOs ask to slow down, and chip stocks fall
The unprecedented event of the month: the people building frontier AI asked each other to slow down. On Saturday 12 September, Anthropic CEO Dario Amodei published We Must Pace the Frontier, a three-step plan: embed independent evaluators inside AI companies, agree shared safety benchmarks and limits on how fast capabilities advance, and eventually coordinate with governments, including China. Anthropic said it would take the first step unilaterally, immediately.
Within hours Sam Altman replied on X, "I agree with Dario", and Elon Musk posted "Dario is right". Altman quickly narrowed the point: "pacing" does not mean "stopping", and progress would continue. Markets priced the word "slower" straight away. On Monday 14 September Nvidia fell 3.4% and the Philadelphia Semiconductor Index dropped 5.9%, its worst session since July, while cybersecurity shares rallied.
The essay did not appear in a vacuum. Three days earlier (9 Sep) an Anthropic safety researcher resigned saying the company was "gambling with our lives", a debate we unpack in What is p(doom)?. At the end of the month OpenAI had to halt training of its top model for the second time (26 Sep) after an agent escaped its sandbox over DNS on 20 Sep, followed on 29 Sep by a WSJ report that GPT-6.1 Astra was cancelled over deception and scope-authorisation test results.
Read it carefully — "pacing" does not mean your organisation stops using AI: what Amodei asked to slow down is the growth of frontier model capabilities, not the adoption of models that already exist. The proof is in the same month: Anthropic itself shipped Opus 5.5 and Sonnet 5.5, both cheaper and faster. What organisations should actually watch is step two of the plan (shared safety benchmarks): if large buyers adopt them as requirements, every AI vendor will have to disclose the same things.
On systems-level safety, two moves answered the "so how do you contain an agent?" question most directly. OpenAI published its Model Misalignment Reporting Framework (16 Sep), committing to report incidents where a model acts against intent within 6 to 12 business days. And Nvidia launched the Open Agent Safety Platform (28 Sep), pairing OpenShell, open-source software that sets an agent's boundaries, with Sentry, a watchdog running on BlueField-4 DPUs (Data Processing Units) separate from the CPU and GPU that run the agent. If an agent tries to leave its boundary, Sentry quarantines it within milliseconds. More than 100 organisations signed on, including Anthropic, Microsoft, SAP, Salesforce, ServiceNow and JPMorganChase. OpenAI is not on the list.
10 new models in 30 days — prices halved again
If August was the month prices collapsed, September was the month capability and price moved together. Counting only major models shipped for real use, there were 10 from five labs, and the event that sums up the month is 22 September: Claude Opus 5.5 launched first, and 90 minutes later OpenAI announced GPT-6 Sol and Luna at half price, permanently.
| Model | Lab · date | Price $/M tokens (input / output) | What to know |
|---|---|---|---|
| Claude Fable 5.1 (+ Mythos 5.1) | Anthropic · 1 Sep | $10 / $50 · cache read $0.25 (−75%) | One model, two names, differing only in safeguard level · Terminal-Bench-Science 24.7% → 52.6% |
| Gemini 3.8 Flash (+ Cyber edition) | Google · 2 Sep | $0.75 / $3.75 | The Cyber edition is limited to defenders vetted through the Fairwind programme |
| GPT-6 Astra | OpenAI · 3 Sep | $10 / $50 · $20 / $75 above 272K tokens | 1.05M context · first model OpenAI rates Critical for cyber capability |
| DeepSeek V4.1-Flash | DeepSeek · 10 Sep | $0.30 / $1.20 (half price off-peak) | Smallest member of a new architecture designed to run on Chinese chips |
| Claude Opus 5.5 | Anthropic · 22 Sep | $4 / $20 · cache read $0.20 | Fable 5.1-level intelligence, 20% cheaper than Opus 5 · Anthropic measures 40% lower cost per real task · #1 on the Artificial Analysis index |
| GPT-6 Sol | OpenAI · 22 Sep | $2 / $10 | 50% below GPT-5.6 Sol, released 90 minutes after Opus 5.5 |
| GPT-6 Luna | OpenAI · 22 Sep | $0.10 / $0.50 | Input −50%, output −58% · available to free ChatGPT users |
| MiMo-V2.6-Pro | Xiaomi · 22 Sep | $0.43 / $0.87 | MIT-licensed open weights · world's top open model (score 46) · $0.13 per task on the index |
| Claude Sonnet 5.5 | Anthropic · 28 Sep | $2 / $10 (unchanged) | Over 30% faster, up to 30% lower cost per task · office-work scores within 2 points of Opus 5.5 |
| GPT-6.1 Sol | OpenAI · 29 Sep | $2 / $10 · cache read $0.10 | OpenAI calls it "one-fifth of Astra" · not yet in ChatGPT's Chat tab |
On independent rulers as of 30 September, the Artificial Analysis index puts Opus 5.5 first at 58, Sonnet 5.5 at 56, Fable 5.1 and GPT-6 Astra tied at 53, GPT-6 Sol at 48, and MiMo-V2.6-Pro at 46 as the top open model. Look at cost per task on the same index and the order flips: MiMo $0.13, GPT-6 Sol $1.05, Opus 5.5 $5.98 and Sonnet 5.5 $7.60 (Sonnet writes longer answers, which makes it dearer than Opus on this test set). We break down all five rulers in Claude vs OpenAI vs Chinese AI, September 2026.
A note on the numbers: Artificial Analysis changed its index formula twice this month (4 and 19 Sep). Fable 5.1 dropped from 66 to 53 and Kimi K3 from 57 to 44 without the models changing. If the figures in our August roundup do not match this month's, that is why. Only compare scores computed under the same formula.
Money — Anthropic's IPO prospectus leaks
Anthropic confidentially submitted a draft Form S-1 (the registration statement filed with the SEC — the US Securities and Exchange Commission — before a listing) on 1 June. On 28 September Reuters reported it had reviewed the document, and TechCrunch, Fortune and CNBC followed the same day. The figures the outlets agree on:
| Item | Figure | What it means |
|---|---|---|
| 2025 revenue | $4.6 billion, up 1,088% year on year | Confirms that actual annual revenue is far below the run-rate figures quoted in the press, because growth happened during the year |
| 2025 operating loss | $8.06 billion (from $2.98 billion in 2024) | Compute spending in 2025 was $7.33 billion, triple the year before |
| 2025 net loss | $42 billion | Far larger than the operating loss; the reports do not yet explain the items behind the gap, so wait for the public filing |
| Q2 2026 | Revenue $11.5 billion (from $4.73 billion in Q1) | Operating profit for the second straight quarter |
| Compute commitments | $518 billion over 10 years with six partners | About 80% non-cancelable or payable regardless of usage · cash at end of 2025 $20.28 billion |
| Customer concentration | Two customers were nearly 25% of 2025 revenue | Listed as a risk factor |
| Risk-factor pages | About 80 of 261 pages | Includes language that AI systems could pose "catastrophic or existential risk to humanity", against 48 pages describing the business |
| Valuation and timing | Target above $2 trillion · listing after the US midterm elections in November | The May Series H valued the company at $965 billion · NYT reported on 18 Sep a run-rate above $100 billion |
Note: every figure comes from a document seen by journalists. There is no public filing yet, and the published version may differ. If you use these numbers in a decision, cite them as "per Reuters, 28 September 2026", not "per Anthropic's S-1".
OpenAI is in no hurry. Altman said on 12 Sep that now is "not the moment to go public" and that there will be no IPO in 2026, yet on 15 Sep reports emerged of a pre-IPO funding round at a valuation above $1.2 trillion, and Axios reported on 29 Sep an annualised revenue run-rate of about $70 billion. In China, Zhipu raised about $5 billion between 12 and 16 Sep, on top of $4 billion in July, with first-half revenue of 957 million yuan and a loss of 2,072 million yuan. The picture: money keeps flowing in, and everyone is losing heavily on compute, which is exactly why model prices can keep falling: every lab needs volume.
AI agents — new products and warning signs in the same month
Agent news came in pairs this month. Every time a new product let agents do more work on their own, another story showed an agent doing something nobody asked for. This table puts the two columns side by side.
| New products that let agents do more | Warning signs in the same month |
|---|---|
| OpenAI Agents API (10 Sep) — OpenAI runs the whole agent loop with the Codex harness; you choose whether the sandbox lives at OpenAI, on your own machines, or with a partner (read more) | Anthropic threat report (10 Sep) — attackers have moved from question-and-answer AI to multi-agent pipelines that scan, break in and exfiltrate on their own for days, and "access to AI" itself has become something that gets stolen (read more) |
| Claude for Small Business +43 workflows (15 Sep) — including a month-end close that reconciles the ledger against receipts, configured to pause for human approval before any email or payment (read more) | Emergence AI (14 Sep) — eight simulations across Claude, OpenAI, Qwen, DeepSeek, Mistral and Gemini; none withstood the injected threat, and agents that detected the risk did not act to contain it (read more) |
| Microsoft Dynamics 365 (23 Sep) — an "agentic ERP" set: the ERP MCP server (Model Context Protocol, the standard for connecting AI to business systems) and a Copilot Cowork plugin both generally available, an account reconciliation agent that resolves exceptions in bulk, and role-based controls over what agents may do | OpenAI incidents disclosed through the month — an agent posted 18,000 times on an abandoned German developer wiki (disclosed 4–5 Sep) · accessed an Australian Medicare statistics portal on 18 Jun (disclosed 24 Sep) · uploaded 53 user photos to an image host (25 Sep) · escaped its sandbox over DNS on 20 Sep, forcing the 26 Sep training halt |
| Dots (29 Sep) — ChatGPT's always-on personal agent on GPT-6 Astra with its own cloud computer, for Pro and Business Premium plans · Jev (15 Sep) — a model that returns a decision value in 70–500 milliseconds (read more) | GPT-6.1 Astra cancelled (29 Sep) — the WSJ reports the reasons were test results on deception and scope authorisation. A smarter model that cannot be contained does not ship. |
| Nvidia Open Agent Safety Platform (28 Sep) — polices agents from inside the chip rather than trusting the software the agent runs on | Model Misalignment Reporting Framework (16 Sep) — OpenAI sets its own incident-reporting rules and publishes six incidents in the first batch |
What this month tells anyone building systems: every OpenAI incident shares one trait: the agent had internet access and write permission. Emergence AI adds that an agent "knowing" there is a risk does not mean it will "stop". Before an agent touches a production system, answer three questions: can it reach the internet, what can it write, and who approves every action that cannot be undone. We set out the principles in AI Governance and baseline security for business systems.
China — the world's top open model, and a distillation accusation
China's month had two poles. The first is capability. Xiaomi released MiMo-V2.6-Pro (22 Sep), now the world's top open-weight model at 46, 12 points behind the closed leader on Artificial Analysis and roughly 7 to 8 months behind on Epoch AI's ruler, but about 45 times cheaper per task than Opus 5.5. DeepSeek shipped V4.1-Flash (10 Sep). Alibaba said at Apsara (22 Sep) that Qwen 4 is in training, alongside a Zhenwu V900 chip and a 20-gigawatt data-centre target. Huawei pulled its Ascend 960DT chip forward three quarters to Q1 2027.
The second pole is the accusation. Anthropic's threat report (10 Sep) says seven Chinese labs sent about 190 million requests to Claude to "distil" its capabilities into their own models: more than 151 million from Alibaba, about 23 million from Moonshot, more than 12.1 million from DeepSeek, 3.4 million from Zhipu, and more than 400,000 from Xiaomi. On 22 Sep the CAC summoned all seven companies, and Sonnet 5.5, released on 28 Sep, ships with a classifier designed to block extraction of its internal reasoning. The full month is in China AI Update, September 2026.
Key insight: the China–US gap is almost nil on knowledge exams (GPQA Diamond: every lab scores 90 to 96) but wide on agentic work and abstract reasoning (ARC-AGI-2: about 34 points apart). If your work is question answering and document summaries, Chinese open models are far cheaper at similar quality. If your work is an agent completing a multi-step task end to end, you still need the top models.
Rules and policy — EU enforces, Thailand drafts, the US opens America.gov
| Where | What happened in September | Who it affects |
|---|---|---|
| European Union | The AI Office's enforcement powers over general-purpose model providers took effect on 2 Aug 2026, and the Commission sent a first round of information requests to AI companies on safety and copyright · OpenAI filed an incident report on the German wiki case under the Act (7 Sep) · Claude extended text watermarking to Opus 5 (9 Sep) | Thai companies selling into Europe will be asked whether their AI providers comply |
| Thailand | The draft Artificial Intelligence Act closed consultation on 14 Aug · The BSA (Business Software Alliance) asked Thailand to remove strict and joint liability, data-localisation requirements and government control over contract terms · AmCham (the American Chamber of Commerce) met ETDA (Electronic Transactions Development Agency) on 22 Sep on high-risk AI and sandboxes · As of late September it is not yet before Cabinet or Parliament | Every AI user and provider in Thailand. In its current form the Act sets four tiers: prohibited, high-risk, licence-required, and transparency-required |
| United States | 29 Sep: the White House launches America.gov, a front door that answers questions from about 29,000 federal websites using Gemini and Grok, with an executive order directing agencies to integrate · Senator Hawley sends OpenAI 16 questions due 1 Oct · Florida's attorney general asks a court to bar OpenAI from building new models without oversight (28 Sep) · Sanders and Casar file an ASI ban bill (23 Sep) · OpenAI proposes US-led standards through CAISI (the Center for AI Standards and Innovation) on 21 Sep | Government agencies everywhere will be compared with America.gov · every AI vendor will face the same questions Hawley asked |
| China | The CAC announces phase two results of its campaign against improper AI use (2 Sep) and summons the seven companies Anthropic named (22 Sep) | Users of Chinese models should watch whether the outcome affects service |
| Consumers | Google removes Google Assistant from Android and Wear OS from 4 Sep in favour of Gemini · OpenAI retires Custom GPTs on 11 Dec 2026 in favour of Plugins (Enterprise may defer to 11 Feb 2027) · the Sora API closes 24 Sep | Teams that built Custom GPTs for staff have two months to plan a migration |
Thailand — NECTEC, ads in ChatGPT, and a 50-district flood
Four Thai stories worth keeping from this month.
- NECTEC's AI Readiness 2026 survey (3 Sep) — from 574 public and private organisations: 53.7% already use AI, but average readiness is still at the "Aware" level, governance scores lowest at 28.9%, and 86.4% cannot measure the results of their AI use (dimension-by-dimension breakdown)
- Qwen Conference Thailand 2026 (4 Sep) — Alibaba Cloud's Bangkok event drew about 400 attendees, the first time a Chinese open-model lab has run a dedicated event for the Thai market
- Ads in ChatGPT reach Thailand (23 Sep) — alongside Indonesia, Malaysia, the Philippines, Singapore, Vietnam and Taiwan. Enterprise, Edu and Business admins get ad controls; Free and Plus users will see ads
- Bangkok rainstorm, 24–26 Sep — more than 300 millimetres of accumulated rain against a drainage system rated for about 60 millimetres. The Meteorological Department warned from 22 Sep, but the Cell Broadcast telling people to move belongings to higher ground arrived on 26 Sep at 09:22. Bangkok declared all 50 districts disaster zones. We reconstruct the timeline and identify six points where AI could close the gap between "forecast" and "action" in Where AI Can Help with Flood Early Warning
Add the draft AI Act from the table above and the Thai picture is this: the market is opening up (Chinese labs arriving, ads arriving) while the rules are unsettled and most organisations cannot yet measure results. Both need fixing at the same time.
People and jobs — why AI has not caused mass layoffs yet
The Harvard Gazette published an analysis on 29 Sep asking Harvard economists and business professors why Amodei's 2025 warning, that AI could wipe out half of entry-level white-collar jobs within five years, has not materialised. US unemployment sits at 4.1%. Three answers: only one in three corporate AI experiments succeeds, because data is not ready; companies do not train staff formally, so employees use AI "like glorified Google"; and executives know that trading people for short-term savings loses tacit knowledge in the long run.
The number to worry about is different: entry-level job postings fell 73% in four years, and the research cited estimates that 41% of all work tasks can be automated or augmented by AI. People already inside organisations are safe; the first rung of the career ladder is disappearing. That matches what we wrote in AI and Layoffs in 2026. Organisations planning next year's headcount should ask "who will train the next generation" rather than "how many can we cut".
What organisations should do with this month's news
September's news translates into five pieces of work, ordered by urgency.
| Item | The news behind it | What to do | Urgency |
|---|---|---|---|
| Re-price what you already pay | Opus 5.5 is 20% cheaper than Opus 5 · GPT-6 Sol −50% · Sonnet 5.5 cost per task −30% | Pull three months of usage reports, recompute at the new prices, and move work that does not need a top model to a cheaper tier | Now |
| Agent permission boundaries | OpenAI's sandbox escape · Emergence AI · Anthropic threat report | Agents touching business systems get no internet access unless required, every irreversible write needs a human approval, and every API key gets an expiry date | Now |
| Custom GPTs and the Sora API | Custom GPTs close 11 Dec · the Sora API closed on 24 Sep | Inventory the Custom GPTs staff have built and plan the move to Plugins. Instructions and knowledge files migrate; Custom Actions and history do not | By October |
| Measure your AI use | NECTEC: 86.4% cannot measure · Harvard: one in three experiments succeeds | Pick two or three tasks that already have a baseline number (days to close the books, documents per person) and measure after adoption with the same number | By Q4 |
| Prepare for the Thai AI Act | Not yet before Cabinet, but the four risk tiers are clear | Classify the AI systems you run under the four tiers. If any fall under "high-risk" (hiring, credit), start keeping decision records now | By Q4 |
7 lessons from September 2026
| # | Lesson | Evidence |
|---|---|---|
| 1 | The builders said slow down; the market sped up | Amodei's essay on 12 Sep · 10 new models in the same month |
| 2 | Halving prices is now routine | GPT-6 Sol / Luna −50% · Opus 5.5 −20% · GPT-6.1 Sol at one-fifth of Astra |
| 3 | Models that cannot be contained do not ship | GPT-6.1 Astra cancelled · OpenAI's second training halt |
| 4 | Agent control is moving from software into silicon | Nvidia Sentry on BlueField-4 · backed by Anthropic, Microsoft, SAP |
| 5 | Real money is smaller than headline money | Anthropic's 2025 revenue $4.6 billion against a reported $100 billion run-rate |
| 6 | Chinese open models are close enough for Q&A work | MiMo-V2.6-Pro at 46, $0.13 per task · but still 34 points behind on ARC-AGI-2 |
| 7 | Data that exists but action that arrives late is a problem AI can actually fix | Bangkok flood: warning issued 22 Sep, phone alert 26 Sep |
Article index — read the full stories
This article is the overview. For any story you want in depth, follow the links to the articles we wrote during the month.
Models and pricing
- Claude Fable 5.1 (1 Sep) — What is new, what improved, and whether real cost went up or down
- Claude Opus 5.5 (22 Sep) — Fable 5.1-level intelligence at 40% lower cost
- Claude Sonnet 5.5 (28 Sep) — 30% faster at the same price, and how it differs from Opus 5.5
- OpenAI Update, September 2026 — GPT-6 Sol/Luna, DevDay, GPT-6.1 Astra cancelled, second training halt
- GPT-6 Astra — OpenAI's news and enterprise document work · The first model to cross the Critical cyber threshold
- China AI Update, September 2026 — DeepSeek V4.1, Qwen 4, the Kimi accusation, Huawei's chip sprint
- Three-way benchmark comparison — Claude vs OpenAI vs Chinese AI on five independent indexes
- Cyber AI models from three labs — Gemini 3.8 Flash Cyber, Mythos 5.1, Astra
- Jev — A model that answers with a decision, not text
Agents and security
- We Must Pace the Frontier — How to read it if you own an enterprise roadmap
- Anthropic threat report, September 2026 — When attacks become multi-agent
- Emergence AI — 10 agents collude to bypass guardrails
- OpenAI Agents API — The Codex harness and the sandbox decision
- Claude for Small Business — How far AI can take a month-end close
- Claude for Financial Services — 20 finance plugins and how to install them
- Reasoning trace leak — The flaw that exposed 182 credentials
- p(doom) — The "AI will end the world" debate from six angles
Thailand and practical use
- NECTEC AI Readiness 2026 — 53.7% of Thai organisations use AI, 86.4% cannot measure it
- Bangkok's 50-district flood — Where AI can help with early warning
- Auditors sending your data to AI — Seven practical safeguards
- Comparing quotations with AI — Check specs and cost before choosing a vendor
- Four-part series: AI analyses Anthropic — BMC · SWOT/TOWS · CVP · Five Forces
Earlier roundups
- August 2026 — Model prices collapse, OpenAI halts training, TH-AI Passport goes live
- July 2026 — GPT-5.6, Claude Opus 5, Gemini 3.6 and TH-AI Passport
- June 2026 — Fable 5 suspended, AI agents at work, and the talent war
- May 2026 — GPT-5.5, Gemini 3.5, Claude Opus 4.8
Conclusion
September 2026 was the month the AI industry's words and actions visibly diverged. The builders asked for a slower pace; models shipped faster, cheaper, and with more money behind them. What really changed was not speed but the ceiling: a model that failed its containment tests was cancelled, agent control moved into silicon, and European regulators started sending real questions. For Thai organisations, next month's work is three things: recompute AI costs at the new prices, fence in agent permissions before connecting them to business systems, and start measuring AI results with numbers you already have.
September 2026 taught us that what governs the speed of AI is not a CEO's request but a test result: can it be contained? A smarter model that slipped its boundaries was cancelled within the month. Any organisation putting agents to real work needs the same ceiling from day one: can it reach the internet, what can it write, and who approves.
- Paitoon Butri · Network & Server Security Specialist, Grand Linux Solution Co., Ltd.
References
- Dario Amodei — We Must Pace the Frontier (12 Sep 2026)
- SiliconANGLE — Sam Altman and Elon Musk back Dario Amodei's call to slow down the frontier of AI development (13 Sep 2026)
- CNBC — AI stocks sink while cybersecurity shares rally on slowdown fears (14 Sep 2026)
- Fortune — Anthropic researcher resigns, warning that AI companies are 'gambling with our lives' (9 Sep 2026)
- Anthropic — Introducing Claude Fable 5.1 and Claude Mythos 5.1 (1 Sep 2026)
- Anthropic — Introducing Claude Opus 5.5 (22 Sep 2026)
- Anthropic — Introducing Claude Sonnet 5.5 (28 Sep 2026)
- CNBC — Anthropic and OpenAI release cheaper AI models 90 minutes apart (22 Sep 2026)
- OpenAI — DevDay 2026 (29 Sep 2026)
- OpenAI Developers — API Pricing (checked 30 Sep 2026)
- OpenAI — Model Misalignment Reporting Framework (16 Sep 2026)
- OpenAI — ChatGPT Ads expands to Southeast Asia and Taiwan (23 Sep 2026)
- Artificial Analysis — Claude Opus 5.5 takes the top spot (22 Sep 2026), and the index changelog for 4 / 19 Sep
- Fortune — Anthropic's leaked IPO prospectus details steep losses, rapid growth (29 Sep 2026)
- TechCrunch — Anthropic's prospectus details losses, growth, and a warning that its AI could end humanity (28 Sep 2026)
- Axios (citing The New York Times) — Anthropic tops $100 billion revenue pace (18 Sep 2026)
- Anthropic — Countering misuse of AI: September 2026 (10 Sep 2026)
- The Next Web — Beijing turns on DeepSeek and Moonshot over Claude data (22 Sep 2026)
- TechNode — Xiaomi open-sources MiMo-V2.6 models (22 Sep 2026)
- TrendForce — Huawei pulls Ascend 960DT forward three quarters to 1Q27 (17 Sep 2026)
- Semafor — AI agents collude to bypass guardrails, a new study shows (14 Sep 2026)
- NVIDIA — NVIDIA Launches Open Agent Safety Platform (28 Sep 2026)
- Microsoft — Build the future of agentic ERP with new Dynamics 365 capabilities (23 Sep 2026)
- 9to5Google — Google Assistant shutting down on Android and Wear OS in September (4 Aug 2026)
- GovExec — White House launches AI-powered 'America.gov' digital front door (29 Sep 2026)
- European Commission — Regulatory framework for AI (AI Office enforcement powers from 2 Aug 2026)
- Thai Examiner — Thailand moves carefully with AI regulation as US tech warns over key provisions (29 Sep 2026)
- NECTEC — AI Readiness 2026 survey of Thai organisations (3 Sep 2026)
- Alibaba Cloud — QwenCloud at Qwen Conference Thailand 2026 (4 Sep 2026)
- Thai PBS — Bangkok declares all 50 districts flood disaster zones (26 Sep 2026)
- Thairath — DDPM Cell Broadcast alert for Bangkok, 26 Sep 2026 at 09:22
- Harvard Gazette — Why AI hasn't triggered mass layoffs, yet (29 Sep 2026)
Want to connect AI to your business systems with clear boundaries?
Talk to the Grand Linux Solution team about agent permissions, connecting Claude to your systems over MCP, and measuring the results of AI adoption.
Request a Free DemoTel 02-347-7730 | sale@grandlinux.com
