- 01
- September
August 2026 marked a clear change of tone from the month before. July was about models getting smarter — you can read that recap in our AI News Roundup for July 2026. In August, almost no lab claimed a capability jump. What they announced instead was cheaper, faster, and safer. This roundup covers the whole month in one piece: the price war, OpenAI's training halt, MCP and A2A moving under one foundation, the funding numbers behind it all, and what happened in Thailand as TH-AI Passport went live.
In one line: Three things from August 2026 matter to enterprises — (1) model prices fell across several vendors at once, with Gemini 3.7 Flash launching at half the previous price and OpenAI cutting API rates by more than 20%; (2) OpenAI halted training on its newest models for two weeks for safety reasons, not commercial ones; (3) MCP and A2A, the two protocols that connect AI to real business systems, now live under the same neutral foundation.
Timeline: what happened in August 2026
| Date | Event |
|---|---|
| 1 Aug | Thailand publishes the draft principles of its first AI law, built on a risk-based approach |
| 2 Aug | Alibaba releases Qwen3.8 Max |
| 3 Aug | OpenAI, Anthropic and Google join a White House AI safety meeting |
| 5–10 Aug | Meta releases Muse Spark 1.2 and Muse Glimmer |
| 6 Aug | OpenAI updates ChatGPT across tiers, adding an effort slider for Plus and Pro users |
| 12 Aug | xAI releases Grok 4.6 |
| 13 Aug | Google launches Gemini 3.7 Flash at half the price of the previous model |
| 14 Aug | Alibaba releases Qwen3.8 27B |
| 17 Aug | Google moves the A2A protocol into the Agentic AI Foundation, alongside MCP |
| 18 Aug | Anthropic opens Playground in the Claude Console |
| 18 Aug | OpenAI announces a two-week training pause on its newest models plus new security protocols |
| 19 Aug | Anthropic moves several Developer Platform features to general availability · Thailand opens TH-AI Passport registration |
| 20 Aug | Reports indicate OpenAI has caught up with Anthropic among business users |
| 21 Aug | OpenAI cuts Sol API pricing by over 20% for three months · DeepSeek releases V4 Flash Vision |
| 26 Aug | Nvidia reports record results · Salesforce and Anthropic announce Claudeforce · Alibaba releases Qwen3.8 Flash |
| 28 Aug | Anthropic opens the Model Hardware Standard, letting AI operate physical instruments |
| 31 Aug | Thailand's TH-AI Passport platform goes live |
1. The price war — cost per token collapsed across several vendors at once
The biggest story of the month was not which model scored highest on a benchmark. It was unit price falling simultaneously. Google launched Gemini 3.7 Flash on 13 August at half the price of 3.6 Flash, only three weeks after the previous release. OpenAI cut Sol API pricing by more than 20% for a three-month window. Alibaba pushed a 125-billion-parameter Qwen3.8 Flash into the same budget tier.
| Vendor | Shipped in August 2026 | Price direction |
|---|---|---|
| Gemini 3.7 Flash (13 Aug), keeping the 1,048,576-token context window | $0.75 in / $3.75 out per million tokens — half the previous model | |
| OpenAI | ChatGPT model refresh (6 Aug) · Ultrafast mode preview (13 Aug) | Sol API pricing cut >20% for three months (21 Aug) |
| Anthropic | No new model this month — focus on platform and partnerships | Moved the other way — Sonnet 5 introductory pricing ended 31 Aug, rising from $2/$10 to $3/$15 |
| Alibaba | Qwen3.8 Max, 27B and Flash (125B parameters) | Small, cost-efficient models |
| xAI / DeepSeek / Meta | Grok 4.6 · V4 Flash Vision (experimental) · Muse Spark 1.2 and Glimmer | Competing in the fast-model tier |
Anthropic is the interesting outlier. It shipped no new model all month — its latest remains Claude Opus 5 from late July — and let Sonnet 5's introductory pricing lapse on the final day of August, so its price went up rather than down. Different vendors are choosing different battlegrounds: some compete on cost per token, others on work actually completed. If you are comparing options, our ChatGPT vs Claude vs Gemini comparison covers how these tiers differ in practice.
Note for enterprises: If you evaluated AI adoption earlier this year and concluded it was not yet cost-effective, the numbers behind that decision are already out of date. Cost per token in the fast tier has fallen several times over in a matter of months. Rerun the calculation before repeating the conclusion — our guide on which tasks to automate first is a practical starting point.
2. OpenAI halted model training — the first pause driven by capability, not business
On 18 August, OpenAI announced it was pausing reinforcement learning on its newest models for two weeks. The decision followed the cyberattack on Hugging Face on 16 July, which OpenAI later acknowledged had been carried out by its own models during testing. Roughly one third of Hugging Face's infrastructure had to be rebuilt.
Security warning: OpenAI stated that deployment safeguards were intentionally disabled for that evaluation — cyber refusals were lowered and the production classifiers that normally block high-risk activity were not running, because the exercise was designed to probe vulnerabilities. The lesson for enterprises is blunt: an AI agent with one layer of protection removed can become an attack tool immediately.
The measures OpenAI announced afterwards are worth copying: stronger isolation for environments running model-generated or untrusted code, controls preventing high-risk workloads from reaching the internet, and automated monitoring of internal model activity with a target of raising an alert within 30 minutes of anything concerning.
For any organisation about to give an AI agent access to internal systems, that is a usable checklist — isolate the sandbox, restrict outbound network access, and define your monitoring and response time in advance. For the underlying attack surface, see AI agent security and prompt injection in business systems and when AI becomes a cyberattack tool.
3. MCP and A2A now share a home — the story that matters most for business systems
On 17 August, Google transferred the A2A (Agent2Agent) protocol into the Agentic AI Foundation at the Linux Foundation — the same body that has hosted Anthropic's MCP (Model Context Protocol) since the foundation launched in December 2025.
| Protocol | What it does | Analogy |
|---|---|---|
| MCP | Connects AI to tools and data sources — databases, business systems, files | Vertical — a USB-C port linking AI to devices |
| A2A | Lets independent agents, from different vendors, talk to each other and hand off work | Horizontal — HTTP, letting servers talk across the web |
The membership numbers show the weight behind this. The foundation grew from fewer than 40 members at launch in December 2025 to more than 250, with Google, Microsoft, Amazon, Anthropic, OpenAI, Bloomberg, Shopify and Block among its backers.
The practical consequence is that lock-in risk at the protocol layer has dropped substantially. With both standards under a neutral foundation rather than a single vendor, integration work you invest in today is less likely to be stranded, and switching model providers later no longer means rebuilding the whole connection layer. For background, see what agentic AI actually is.
4. Anthropic — a month spent moving from models to real work
No new model, but heavy product movement. On 26 August, Anthropic and Salesforce announced Claudeforce, bringing Claude into Salesforce along with a plugin shipping 37 prebuilt sales skills. On 28 August, Anthropic opened the Model Hardware Standard (MHS) as a research preview — a common driver interface letting AI discover and operate physical devices, from microscopes and liquid handlers to robotic arms. In effect, it is the MCP idea extended from software into the physical world.
Earlier, on 19 August, the developer platform moved several capabilities to general availability, including computer use, browser use, the Files API, Agent Skills and an enterprise admin API, along with expanded agent management and domain-level access controls.
5. The money — numbers saying this cycle is not over
| Item | Figure | How to read it |
|---|---|---|
| Nvidia latest quarter (26 Aug) | Revenue $96.22 billion, up 106% year over year; adjusted EPS $2.22 against a $2.10 estimate | A beat for the fifteenth consecutive quarter |
| AI cloud, industrial and enterprise segment | $40.3 billion, up 138% year over year | Growing faster than the hyperscaler segment — demand is broadening beyond the big four |
| Anthropic IPO preparation | Investors targeting a valuation around $2 trillion in October 2026 | Filed confidentially on 1 June; most recent private valuation about $965 billion |
| OpenAI enterprise business | Enterprise revenue has overtaken consumer | A 20 August report indicates it has caught up with Anthropic among business users |
The detail worth pausing on came on 21 August: reporting indicated Anthropic's IPO filing will list public backlash against AI as a formal risk factor. That is the first time sentiment toward AI has been elevated to a disclosable business risk. For a wider view of how hot this market has become, see the AI bubble in 2026.
6. Thailand — TH-AI Passport goes live, and a law heading to parliament
| Date | Item | Detail |
|---|---|---|
| 1 Aug | Draft principles of Thailand's first AI law | A risk-based approach rather than a single rule covering every use, with defined duties for both providers and users; the government aims to move it into the legislative process by September 2026 |
| 19 Aug | TH-AI Passport registration opens | Open to anyone aged 15 and above |
| 31 Aug | Platform officially goes live | Access to professional-grade models from 14 providers across 33 models, plus more than 96 training courses aligned to UNESCO, ASEAN and ETDA frameworks · Register by 30 November 2026 for 30 days of Innovator-tier access |
| Ongoing | Data centre investment | Foreign investment applications in the first half reached 1.37 trillion baht, up 80% year over year, driven by AI and data centre projects, while the government has begun screening proposals more tightly |
What changes in practice is this: from 31 August, ordinary employees can reach models of the same calibre their employer licenses, without waiting for a budget approval. That guarantees more AI use outside IT's line of sight. Any organisation without a policy governing personal AI use at work should draft one before internal data starts leaving by accident — see AI governance for organisations for a framework.
What organisations should do with this month's news
Three things you can act on in September
1. Recalculate your costs. If you priced AI adoption earlier this year and shelved it as too expensive, that figure is stale. Fast-tier cost per token has fallen several times over.
2. Review what your AI agents can reach. The lesson from OpenAI is that a single layer of protection is not enough. Isolate the sandbox, restrict outbound access, and monitor actively.
3. Write a personal-AI-use policy. With professional-grade models freely available from 31 August, the question is not whether staff will use them, but when you will find out.
A view from an ERP team
The news with the most direct bearing on ERP work is MCP and A2A landing under one foundation, because that is the layer where AI connects to real operational data — not the layer where models compete on test scores.
At Saeree ERP we already work this way internally. Our support team uses an AI assistant connected to system databases through MCP to look up and cross-check information while handling customer issues. To be clear, this is internal use by our own team and not a feature we ship to customers. We are still collecting lessons on access rights, data scoping and audit trails before deciding what should become a real product capability.
The principle we hold to does not change: the ERP system is the source of truth, and AI is an assistant that reads and summarises. Numbers used for decisions must come from a system that can be traced back, not from text a model generated — because when an auditor asks where a figure came from, the answer has to point at a source document.
Conclusion
August 2026 was a month in which AI matured in unglamorous ways — prices fell, standards consolidated, and for the first time a major lab stopped training a model for safety reasons. In Thailand, the shift was toward access: a national platform putting professional-grade models in ordinary hands, and a draft law heading toward parliament.
For organisations, this is the month to revisit old assumptions — both the ones about AI being too expensive, and the ones about your security being tight enough.
AI did not get smarter this month. It got cheaper and better governed — and for enterprises, that is the better headline.
- Paitoon Butri · Network & Server Security Specialist, Grand Linux Solution Co., Ltd.
References
- OpenAI — GPT-5.6 August Updates
- Fortune — OpenAI paused AI training for two weeks (18 August 2026)
- Axios — OpenAI pause and the Preparedness Framework (18 August 2026)
- 9to5Google — Gemini 3.7 Flash launch (13 August 2026)
- Axios — A2A joins the Agentic AI Foundation (17 August 2026)
- Salesforce — Salesforce and Anthropic announce Claudeforce (26 August 2026)
- The Japan Times — Anthropic's Model Hardware Standard (28 August 2026)
- CNBC — Nvidia earnings (26 August 2026)
- CNBC — Anthropic IPO filing to list AI backlash as a risk (21 August 2026)
- TechCrunch — OpenAI gaining on Anthropic with business users (20 August 2026)
- Bloomberg — Alibaba releases a smaller Qwen model (26 August 2026)
- THE STANDARD — TH-AI Passport goes live, 31 August 2026
- Thai PBS — First day of TH-AI Passport registration (19 August 2026)
- Policy Watch — Thailand's first draft AI law
- Bloomberg — AI data centre boom drives surge in Thailand FDI applications
Information verified as of 1 September 2026.
Looking for an ERP system for your organisation?
Wondering how ready your data is for AI to read and summarise? The Saeree ERP team advises on consolidating accounting, inventory and budget data into one auditable system.
Request a Free DemoTel 02-347-7730 | sale@grandlinux.com
