Vol.45 · August 24, 2026
dera news AI Weekly Vol.45 | 2026-08-24 - This Week's AI News
🤖 dera news AI Weekly Vol.45
2026-08-24
This week's AI world in one sentence? Tools reached everyone, and design became the only thing left that differentiates. Codex passed 20 million users and Japan's national informatics institute released a domestic model for free — while changing only the harness moved a coding model from 30th place to the top 5.
📊 What You Need to Know This Week
The number that matters most this week: Codex passed 20 million users on 21 August. It had crossed 15 million on 13 August — five million in eight days. In early June it was five million. Roughly 33x in eight months.
In the same week, Japan's National Institute of Informatics released LLM-jp-4 33B under Apache 2.0. Commercial use is permitted, and the thinking variant beats every previous model in the series across major Japanese-language benchmarks. Even the DPO training data for the thinking model was published — an unusual decision for a research institute. Alibaba's Qwen3.8-27B was likewise shown running frontier-class coding agents locally with no cloud API.
So as of this week, capable models are available locally, in Japanese, and for free. Having one is no longer a differentiator.
Where, then? This week produced a number for that too. Changing only the harness — the execution environment an agent runs in — moved a coding model from 30th to the top 5 in LangChain's experiment. Inherent, founded by DeepMind alumni, beat Claude Opus 4.8 and GPT-5.5 at reproducing research findings using a model of just 27 billion parameters. The gap came from design, not scale.
The same pattern holds at the organizational level in Japan. According to the 2026 White Paper on Information and Communications, 86.4% of Japanese companies now use generative AI somewhere in their operations, up from 55.2% — while 27.0% report having no organizational approach to it at all. Tools everywhere, design nowhere.
And this week produced a Japanese company showing what design delivers. SBI Holdings announced a company-wide partnership with Anthropic; SBI Securities invested ¥500 million in a customer-facing agent and projects ¥2.7 billion in additional annual revenue, largely through reactivating dormant accounts. Chairman Yoshitaka Kitao put it plainly: "Son-san has OpenAI; I have Anthropic."
On the capital side, meanwhile, the first headwind appeared. Alibaba announced a $10.2 billion share placement and its stock fell 10%. AI spending itself is now being questioned by shareholders. OpenAI, separately, paused reinforcement learning training on its latest model for two weeks — though its products and distribution did not slow at all.
This week also widened where you touch AI from. Codex can be driven by voice and keeps working after you step away, ChatGPT on macOS connects to Apple Messages, and Claude Code sessions talk to each other directly. None of it is a new model — it is where and how you reach the models you already have.
What we're watching is that the same conclusion arrived from both the technical and the organizational side at once. Not what you have, but how you run it.
💡 This Week's Actions
1. Try the free domestic model on your own data (1 hour) NII's LLM-jp-4 33B is Apache 2.0 with commercial use permitted, beats its predecessors on Japanese benchmarks, and ships with DPO training data. For work where Japanese is the primary language — customer response automation, internal document summarization — let's see how it performs. → NII: LLM-jp-4 33B release
2. Decide how you'll run it before deciding which model (1 hour) The same model moves 25 places depending on the harness. Interactive, batch, and asynchronous modes carry different cost structures and suit different providers. Let's work out which one our work actually is, first. → MarkTechPost: three ways to run an agent loop and the economics of each
3. Look at the payback design, not the investment size (30 min) SBI Securities laid out ¥500M in, ¥2.7B projected annual revenue out. What's instructive is not the amount but that the target metric — reactivating dormant accounts — was chosen before the agent was built. → ITmedia: SBI's Kitao on the AI-driven strategy
📰 This Week's Articles (14)
1️⃣ Codex passes 20 million users, up 33x in eight months
🏷️ Adoption, Developer Tools, OpenAI What happened? OpenAI's coding agent Codex passed 20 million users on 21 August, eight days after crossing 15 million. It was at five million in early June and ten million on 21 July. The GPT-5.6 launch and the decision to merge Codex into ChatGPT as a single workspace drove the acceleration. OpenAI credited every subscriber with a banked usage reset to mark the milestone. Our view At 20 million, coding agents are no longer a specialist tool. The question of whether to adopt is settled; only how to run it remains. Complaints about usage limits are also surfacing, so if this is going into production work, decide in advance what happens when a limit is hit. 📎 Memeburn
2️⃣ Japan's NII releases domestic model LLM-jp-4 33B under Apache 2.0
🏷️ Open Weights, Japan, Domestic Models What happened? NII published two models: llm-jp-4-33b-base and the thinking variant llm-jp-4-33b-thinking, both Apache 2.0 with commercial use permitted. The thinking model outperforms every earlier model in the series on major Japanese benchmarks including MT-Bench and AnswerCarefully, trained with SFT and DPO. Unlike April's MoE release it uses a dense architecture. The DPO training data for the thinking model was published as well. Our view For readers in Japan this is the most directly consequential item of the week. A national research institute has released one of the strongest Japanese-language models available under a commercially usable license. Publishing the training data is rare even among research institutions and makes customization on your own data a realistic starting point. If Japanese is your working language and the data cannot leave the building, this is the highest-value thing to test right now. 📎 National Institute of Informatics
3️⃣ Qwen3.8-27B runs frontier-class coding agents locally, no cloud API required
🏷️ Open Weights, Local Execution, China What happened? Following last week's weight release, Qwen3.8-27B has been shown running coding agents and reasoning tasks locally with no cloud API dependency. It is natively multimodal across images and video and handles context in the million-token range, under Apache 2.0. Our view Last week the story was the requirement — 16GB. This week it is the demonstration. Together with NII's release, the number of capable options that run on hardware you already own went up by two in a single week. Work previously blocked by metered APIs and data-transfer constraints becomes more feasible in-house. 📎 VentureBeat
4️⃣ Same model, different harness: 30th place to top 5
🏷️ Agents, Design, Cost What happened? Analysis suggests the design of the harness — the execution environment an agent runs in — matters more than model choice. In LangChain's experiment, changing only the harness moved a coding model from 30th to the top 5. The Python agent Decode proposes three execution modes: interactive online for real-time dialogue, remote offline for parallel server-side processing, and asynchronous online for queued work returned later. Each carries different latency and cost efficiency and suits different providers. Our view This is the center of the week. Most companies spend their evaluation time on which model to choose — but if the same model moves 25 places on how it is run, the order of investigation itself is wrong. Across 1,000 documents, API costs that are steep in interactive mode can drop substantially on offline GPU. Decide which mode the work is, then choose the provider. 📎 MarkTechPost
5️⃣ DeepMind alumni at Inherent beat the majors with a 27B model
🏷️ Agents, Research, Startups What happened? Inherent, a London lab founded by Google DeepMind alumni, emerged from stealth with $50 million in seed funding. Its agent Faraday outperformed Claude Opus 4.8 and GPT-5.5 at independently reproducing findings from scientific papers without being told the answers. Faraday runs on Qwen 3.6 — 27 billion parameters. The co-founder describes using reinforcement learning to instill research taste: which experiments to run and how to design them. Rather than building its own coding tool, Faraday uses OpenAI's GPT-5.5 Codex. Our view The concrete case behind the previous item. The difference came from design, not model size. Declining to build a coding tool and using Codex instead is telling too — the way a human researcher relies on existing software. Small models plus good design can deliver, which makes in-house work a more realistic option than it looked. 📎 TechCrunch
6️⃣ DeepSeek Harness hits developer preview, MIT-licensed with every component swappable
🏷️ Agents, Open Source, Developer Tools What happened? DeepSeek released a developer preview of DeepSeek Harness under an MIT license. Every element an agent needs functions as a plug-in. Built on a meta-framework called Cordis, it allows models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and even the UI to be swapped. It supports both Anthropic and OpenAI. Our view What we described in Vol.44 as an open-source answer to Claude Code has arrived MIT-licensed and provider-agnostic. Not being locked to a single vendor is stated as a design principle. Diversifying dependencies is something we return to often, and having that choice at the harness layer matters. It is still preview, so expect compatibility changes. 📎 GitHub
7️⃣ Codex voice mode updated: the agent keeps working after you step away
🏷️ Interfaces, Developer Tools, OpenAI What happened? OpenAI expanded voice mode across Codex and ChatGPT on the desktop. Powered by GPT-Live, it is full-duplex voice control available from Chat, Work and Codex on macOS and Windows — in the desktop app, not the terminal. The 18 August update lets voice run directly from existing task composers, improves connection reliability, and keeps task actions running in the background after you leave the conversation. You can interrupt mid-answer, and on a Mac "Take a look at this" shares your frontmost window. Our view The substantive change is that leaving the keyboard no longer means stopping the agent. Sending off a long job and then checking progress and issuing the next instruction while moving around becomes realistic. In this week's terms, it adds one more option at the execution-unit layer — the same model, a different way of touching it, and a different shape of work. 📎 VentureBeat
8️⃣ ChatGPT on macOS connects to Apple Messages, with per-send approval
🏷️ Interfaces, Automation, OpenAI What happened? OpenAI released a plugin connecting the macOS ChatGPT app to Apple Messages. From the Codex or ChatGPT Work chat window, you can search past messages, draft replies and send them. To prevent mis-sends, a human approves each actual send. It requires an Apple Silicon Mac, and message triage is handled through the plugin too. Our view One of the first practical cases of AI entering a personal communication channel. The line worth borrowing is AI drafts, human sends — a permission pattern that transfers directly to other workflows. It does grant access to past conversation history, so confirming your data-handling policy before business use is the safer order. 📎 ITmedia
9️⃣ Claude Code sessions can now message each other, and agent teams arrive
🏷️ Agents, Developer Tools, Anthropic What happened? Claude Code sessions can now message each other directly. It ships as a first-party channel rather than a plugin, carrying plain-text messages alongside permission-aware approval dialogs. Alongside it, agent teams arrived: unlike subagents, each teammate works independently in its own context window and communicates directly with the others, and users can talk to individual teammates without going through the lead. Cross-vendor conversation — ChatGPT talking to Claude Code — still runs through community plugins. Our view If you run several agents in parallel, you have been the courier between them. That part moves into the product. In exchange, approval design matters more: once agents can agree among themselves and proceed, deciding in advance where a human gate remains is the safer position. The same question the Messages plugin raises, now on the development side. 📎 Claude Code Docs
🔟 SBI partners with Anthropic company-wide; ¥500M in, ¥2.7B projected out
🏷️ Japan, Enterprise Adoption, ROI What happened? SBI Holdings announced a company-wide partnership with Anthropic, naming AI adoption its top near-term priority and bringing in specialist talent from outside. SBI Securities invested ¥500 million in a customer-facing AI agent and projects ¥2.7 billion in additional annual revenue, largely through reactivating dormant accounts. Chairman Yoshitaka Kitao: "Son-san has OpenAI; I have Anthropic." Our view The most operationally relevant Japanese case this week. What matters is not the size of the number but that the investment and the payback are tied to a specific business metric. The thing to move — dormant account reactivation — was chosen first, and the agent was pointed at it. Hiring design talent from outside also fits this week's theme: someone has to do the designing. Projections remain projections, so the results report is worth waiting for. 📎 ITmedia
1️⃣1️⃣ Claude designs protein binders for 14 of 15 targets at roughly double the industry hit rate
🏷️ Applied, Research, Anthropic What happened? Of 1,320 designs produced by Claude Opus 4.8 and an unreleased Mythos Preview, independent synthesis and evaluation by Adaptyv Bio and Twist Biosciences confirmed binding in 354 — a 26.8% hit rate overall, rising to 35.1% when focused on a single target, against an industry norm of 10–15%. Our view What matters is that AI moved past proposing candidates to producing results independently verifiable by outside labs. Early candidate design consumes enormous trial effort in drug discovery, and compressing it changes the cost structure of exploration. Still research rather than product, but a useful read on how contract research competition will shift. 📎 Anthropic coverage
1️⃣2️⃣ Alibaba’s $10.2B share placement sends the stock down 10%
🏷️ Market, Capital, China What happened? Alibaba priced a $10.2 billion share placement in Hong Kong to fund AI investment, and the stock fell 10%. Earlier in August its shares had already declined after AI spending drove a 75% drop in net income. Our view The one story this week where capital met resistance. Across Vol.44 and this issue the theme has been that money does not slow down — but AI spending itself is now being questioned by shareholders. A placement funds future growth while diluting existing holders. If we have entered a phase where the payback path matters more than the investment size, the SBI approach of fixing the metric first becomes more important, not less. 📎 CNBC
1️⃣3️⃣ Nvidia takes $1.5B stake in SB Energy with up to $105B for the Ohio data center
🏷️ Infrastructure, Capital, Japan What happened? Nvidia will invest $1.5 billion in SoftBank subsidiary SB Energy, becoming exclusive supplier of compute infrastructure to OpenAI's Ports-Pike data center in Ohio, alongside a credit facility of up to $105 billion. The site scales from 4.25GW to an eventual 8GW, with a 9.2GW gas plant on site at an estimated $33 billion. Our view Last week we reported Nvidia halving its OpenAI guarantee from $250B to under $120B. This week the same envelope reappeared as equity plus credit line — restructured rather than withdrawn. For companies in Japan it marks the point where AI outcomes depend on securing power and data-center capacity long-term, not only on model quality. Separately, the Anthropic acquisition of decart we covered last week is now reported at $7 billion, up from around $6 billion. 📎 TechCrunch
1️⃣4️⃣ Tracked rare books turn up destroyed inside an Amazon AI training facility
🏷️ Data, Copyright, Ethics What happened? 404 Media planted tracking devices in rare books and followed the shipment to an Amazon warehouse in Las Vegas, Nevada. Employees there say their main task is receiving large book shipments and cutting the bindings to speed scanning — destroying the books in the process. The team, designated VGT3, uses a logo of a dinosaur baring its teeth while clutching a book. Our view In Vol.42 we covered reports of AI companies buying and destroying rare books; this week brings physical evidence. How training data is actually acquired has been shown by field investigation rather than inference. We do not take a side on the copyright question — we process and rewrite information with AI ourselves — but the tension is worth holding in view rather than setting aside. 📎 404 Media
📚 Editor's Note
This week the conversation moved decisively from what you have to how you run it. Codex reached 20 million users, NII gave away a domestic model, and Qwen ran frontier-class agents on hardware you already own. Owning a good model is no longer a differentiator.
Design is. Change only the harness and a model moves 25 places. A 27-billion-parameter model beats the frontier when the task is designed well. And per the government white paper, 86.4% of Japanese companies already use generative AI while 27% have no organizational approach to it. The same conclusion arrived from the technical layer and the organizational layer in the same week.
In that light the SBI case was the telling one. Not the ¥500 million, but that the metric to move — dormant account reactivation — was fixed before the agent was built. If you know what you want to improve, the model can be chosen afterwards. The reverse does not work.
Alibaba's placement and the 10% drop are the same point inverted. Investment size alone no longer persuades the market; the payback design is being asked for.
See you next week, with useful information and something to think about. The dera news team