FORSMILE
Issue #5Published September 21, 2026 / Covering Sep 14-21日本語で読む

This Week in AI, 15 Items — OpenAI and Anthropic Both Moved to Make Their Own Failures Externally Checkable

The week in one place

The thread running through this week was making internal state checkable from outside. OpenAI published a framework for tracking, investigating and disclosing model misalignment, along with six reports from the past six months, and stated plainly that it does not believe the industry has solved alignment and monitoring well enough to keep scaling at maximum speed for much longer. In the same week Anthropic announced that Faculty, Accenture's AI business, will act as an embedded evaluator inside Anthropic, with each company expecting to invest at least $1 billion over five years. Anthropic also published three internal measurements of its own development pace — Claude leads 26% of its AI R&D work, and about 6% of AI R&D compute goes to safety. Both moves shift the ask from "trust the announcement" to "here is something you can count." Products and pricing did not pause for it. **The most unusual release was TypeSafe AI's Jev, which claims a new category — a "System One model" that returns typed probabilistic decisions rather than text — and opened early access at $0.042 per million input tokens with output free.** Google shipped Gemini 3.8 Live and its Extended Thinking variant, OpenAI introduced Astra for Law and pushed ChatGPT Ads toward agents, and Mozilla put Mistral behind Firefox's Smart Window. The changes that hit running costs were in developer tooling: Claude Code made the server-side auto-mode classifier the default and stopped billing for its overhead, while GitHub Copilot set 19 October as the removal date for six models. On the infrastructure side, NVIDIA's Vera Rubin NVL72 debuted in MLPerf Inference at up to 3.7x the previous generation, and Cornelis raised $205 million behind the argument that the bottleneck is GPUs waiting on data.

Regulation & Safety

OpenAI publishes a framework for reporting model misalignment, with six reports from the last six months

It will now disclose observed misbehaviour before it can explain or mitigate it.

On 16 September 2026 OpenAI published a framework for tracking, investigating and disclosing model misalignment, together with six reports on unexpected or concerning behaviour observed in the previous six months. It describes past disclosure as ad hoc — collated into batches or folded into system cards — and says the new framework **favours disclosure even when significance is uncertain**. Scope covers the whole model lifecycle: training, evaluation, testing and deployment, including new ways for models to act without authorisation, coordinate with other models or evade oversight, failures that call a safeguard into question, behaviour that contradicts a published safety assessment, and misalignment that may affect third parties. Recurrence of an already-disclosed issue will be added to the original report as evidence about whether mitigations work. OpenAI states that it does **not** believe the industry has solved alignment and monitoring enough to keep scaling responsibly at maximum speed for much longer, and that it is working to propose mechanisms for sharing serious safety, security and misalignment incidents with the US federal government.

So What

Qualifying misalignment across the whole lifecycle — training, evaluation, testing, deployment — now has a publication channel that does not wait for an explanation, a mitigation, or the next system card. If you hand tools or permissions to an agent, it is worth checking on a schedule. Note that **these six were observed during training or evaluation, so they are not all unfixed problems in a model you are running.** Note that this lands in the same week as Anthropic's embedded-evaluator deal — two labs moving the same direction at once.

Funding & Corporate

Anthropic embeds Accenture's Faculty as an evaluator inside the company; each side to invest at least $1 billion over five years

Accenture's Faculty gets employee-level access to how Anthropic's models are built.

On 18 September 2026 Anthropic announced a partnership with Accenture on independent evaluation of frontier AI, led by Faculty, Accenture's specialist AI business. The work covers evaluating and red-teaming models, alignment assessments and testing safeguards. **Each company expects to invest at least $1 billion in building capacity over the next five years.** Unlike today's external evaluators, embedded evaluators work inside the company with **access comparable to an employee's**: they can watch models take shape during training, follow the decisions governing how models are built and deployed, and talk to staff directly. Anthropic is explicit that no standards yet exist for what access embedded evaluators should have or how they should report, and that no settled funding system exists either. It argues funding should eventually come from pooled or government sources, but since neither exists, **Anthropic will fund Accenture's work directly**. **The partnership is non-exclusive**; Anthropic says it expects frontier labs to work with several organisations at once, and that it is in dialogue with METR and other nonprofit evaluators about piloting elements of embedded evaluation **using their own funding**.

So What

The claim "we are safe" now has a funded mechanism for third-party verification attached to it. But the party being evaluated is paying — a conflict Anthropic names itself — which matters when you read the resulting reports. If you compare vendors on safety claims, this gives you two new questions: who checked, and who paid for the check.

Models & Products

TypeSafe AI opens early access to Jev, a "System One model" that returns typed decisions instead of text

It gives up string generation and returns schema-guaranteed probabilistic decisions in a single parallel pass.

On 15 September 2026 TypeSafe AI introduced a new class it calls **System One Models**, along with **Jev**, its first public model, available in early access. The company describes Jev as **"a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out"**, trained with a method it calls **Reinforcement Learning for Calibrated Decisions (RLCD)**, and sampling **in parallel — generating all outputs in a single query** rather than one token at a time. Published figures: **input tokens at $0.042/MTok ($42 per billion tokens) and output tokens FREE ("too cheap to meter")**, against $0.20–$10/MTok input and roughly 5x that for output on existing LLMs. **End-to-end response time is 70ms–500ms** versus 3–329 seconds for frontier models, which TypeSafe puts at **40x–200x faster for System One shaped queries**. Its workflow evals give **193.6x faster and 444.6x cheaper**, with the company noting **it expects those to be "on the higher end of real world gains"**. Because possible outputs and structure are defined in advance, TypeSafe states that **"the model never makes type errors"** and that Jev **"is optimized for structured outputs and can't hallucinate"**, with calibrated probabilities and confidence scores attached to every answer. **String generation is given up.** **TypeSafe came out of stealth on the same day, 15 September 2026.** Lead investor DCVC says in its own post of that date that it **led a $40 million Series Seed** into the company; no other participating investors are named there.

So What

**Anywhere you are using an LLM as a component that returns a decision your code acts on — rather than as something that writes prose — is a candidate to be replaced.** Classification, routing and extraction all have a fixed output type, and today that means you are writing parsing and validation around the model; this design removes that layer. The price gap is an order-of-magnitude difference too. But **both the performance comparison and the "can't hallucinate" claim rest on the company's own figures and its own definition, with no third-party verification**, and early access is waitlisted, so you may not be able to try it immediately.

Models & Products

Google ships Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking

Two voice models: one tuned for cost at scale, one that reasons while it speaks.

On 15 September 2026 Google introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking (the post carries a 17 September update). The first is built for scale and cost efficiency; the second targets high-complexity, multi-step work and **reasons and speaks simultaneously**, using verbal cues such as "Let me check that..." to hold the conversation open. **Both are available to developers in the Gemini API and Google AI Studio from launch day.** For enterprises they are in private preview in Gemini Enterprise, with Gemini Enterprise for Customer Experience coming soon. For consumers, 3.8 Live appears in Search Live and 3.8 Live Extended Thinking in Gemini Live and Workspace for Pro and Ultra subscribers. Published figures: Speech to Speech Quality Index 82.6 (Extended Thinking ranked first), 68.6% agentic task completion on τ-Voice, 35.1% on Sierra's τ-Voice-banking, and 97.7% on Big Bench Audio. **The announcement carries no rate card, but it does address price**: it says 3.8 Live Extended Thinking maintains "a highly competitive price point compared to other frontier models" and includes an Artificial Analysis chart of **cost per hour of input audio**. Google's Gemini API pricing page lists 3.8 Live and 3.8 Live Extended Thinking on one row, with paid-tier rates of **$0.75 per million text input tokens, $3.00 per million audio input tokens ($0.005/min), $12.00 per million audio output tokens ($0.018/min), and $1.00 per million image/video input tokens ($0.002/min)**. A free tier is also offered.

So What

If you are building voice as an entry point, you can test it through the API today and cost it immediately, since **the pricing page gives per-minute rates** ($0.005/min audio in, $0.018/min audio out). Note that every published benchmark here is voice or agentic; none of them stands in for text generation quality.

Developer Tools

Claude Code 2.1.271 to 2.1.278: auto mode moves to a server-side classifier that is not billed, and AGENTS.md is now read

The hidden cost of auto mode drops, and projects without a CLAUDE.md get picked up from AGENTS.md.

Eight Claude Code releases shipped between 14 and 19 September 2026, 2.1.271 through 2.1.278 (dates confirmed from npm publish timestamps). **In 2.1.278, auto mode for Claude API and Enterprise users, and on Bedrock, Vertex, Foundry and gateways, defaults to the server-side classifier, which does not charge for classifier overhead.** On Bedrock, Vertex, Foundry and gateways, `CLAUDE_CODE_AUTO_MODE_SERVER=0` opts out, and a billed fallback now produces a warning. `/status` gained an "Auto mode server" row showing whether this session's classifier runs server-side. **2.1.277 added AGENTS.md support**: in a project with no CLAUDE.md, Claude Code reads AGENTS.md instead (not yet on Bedrock, Vertex or Foundry). Elsewhere in the week, 2.1.275 added syncing of claude.ai account skills and plugins to terminal sessions, and 2.1.271 added per-command `allowed_domains` for Bash, PowerShell and Monitor in sandboxed auto mode.

So What

If you run auto mode as a default, removing the classifier overhead comes straight off your running cost — check `/status` to confirm your session is on the server-side path. AGENTS.md support means a repository shared with other agents no longer needs a duplicate CLAUDE.md alongside it.

Developer Tools

GitHub Copilot will remove six models on 19 October 2026

GPT-5.5, the GPT-5.4 line, Gemini 3.7 Flash and Grok 4.5 all go on the same date.

On 18 September 2026 GitHub announced the deprecation of selected GitHub Copilot models on **19 October 2026**. The six models and their replacements are: Gemini 3.7 Flash to Gemini 3.8 Flash; GPT-5.5 to GPT-5.6 Sol; GPT-5.4 to GPT-5.6 Sol; GPT-5.4 mini to GPT-5.6 Luna; GPT-5 mini to GPT-5.6 Luna; and Grok 4.5 to Grok 4.6. GitHub asks users to update workflows and integrations before that date. **For Enterprise and Business customers with default model enablement turned on, the replacement models activate automatically unless an administrator has disabled them.**

So What

Any automation with a model name hardcoded needs editing before 19 October. A separate set of four Copilot models was already scheduled for removal on 2 October, announced earlier this month. **With removal notices landing this close together, stop hardcoding model names.**

Regulation & Safety

Anthropic opens the Life Sciences Verification Program, with biology safeguards relaxed for vetted organisations

Verified life-science teams get classifiers that are more permissive for biology work.

On 17 September 2026 Anthropic introduced the Life Sciences Verification Program (LSVP), giving vetted life-science professionals access to its Mythos, Opus and Sonnet models with safeguards refined to be **more permissive for biology-related work than the generally available Fable models, which currently block tasks such as drug discovery, research biology, clinical development and manufacturing**. Dozens of organisations were already onboarded through early access; applications are now open more broadly. The programme launches in beta, **initially for teams and institutions**, with individual Pro and Max plans to follow over time. Verification reviews research credentials, security standards and ethical research oversight. Verified teams can then apply for "Standard Use" or "High-risk Use" grants, usable across Claude Science, Claude.ai, Claude Code and the API. Standard Use covers Mythos 5.1, Opus 5 and Sonnet 5 today plus future models, and renews annually.

So What

This is the shift from "block everyone for safety" to "relax it only for people we have checked," made concrete in one domain. **The same design can travel to other high-risk areas**, so if a model refuses work you legitimately need, the next step may be finding the verification programme rather than switching vendors.

Research

Anthropic publishes three measurements of AI development pace inside a frontier lab; Claude leads 26% of its own AI R&D

Actual internal numbers for how much of AI development is being done by AI.

On 17 September 2026 (the date shown on Anthropic's own news index) Anthropic published three measurements of how fast AI development is accelerating inside frontier labs. **(1) An AI-led AI R&D automation index**, scored on an Automation Level scale from AL0 to AL5: **as of August 2026, Claude "leads" 26% of Anthropic's AI R&D work**, is not fully autonomous on any measured subset, and the share at or above "AI collaborates" is above 90%. **(2) Oversight of AI agents**: roughly 30,000 agents operate on the platform, **0.002% of decisions — about 1 in 47,000 — were blocked**, online monitors cover 100% with real-time review, and offline monitors cover 100% with roughly one to two transcripts per thousand flagged. **(3) Compute allocation**: for 13–20 July 2026, **about 6% of compute going to AI R&D went to safety work, rising to about 12% within AI-driven AI R&D.** All figures are one company's self-reported numbers, and **the reference periods are August and mid-July, not the publication date.**

So What

The abstract claim that AI is starting to run AI development now has a measured internal snapshot attached. The specific levels matter less than the fact that **the method is stated clearly enough to be re-run and compared**. Anthropic has not committed to publishing this on a schedule, and there is no third-party verification of these figures.

Market & Industry

OpenAI tests Sponsored Agents in ChatGPT Ads, and integrates HubSpot and Shopify

Clicking an ad can now open a conversation with a business-sponsored agent.

On 16 September 2026 OpenAI introduced new AI-powered experiences for ChatGPT Ads. **Sponsored Agents** let someone who clicks an ad in ChatGPT **start a conversation with an agent sponsored by that business**, and are **currently being tested with select advertisers in the United States**. OpenAI says that conversation is distinct from ChatGPT's independent answers and separate from the conversation the user originally started. OpenAI is also making ads creatable from a few prompts in ChatGPT Work, adding AI creative tools to Ads Manager, and shipping integrations with **HubSpot as its first CRM partner and Shopify as its first ecommerce partner** so ChatGPT Ads can be used from tools businesses already run. OpenAI says its ads principles are unchanged and that protecting trust in ChatGPT remains its priority.

So What

ChatGPT Ads was already reported at a $1 billion annualised run rate in an earlier issue; what is new is that **the destination of an ad becomes a conversation rather than a page**. If your funnel assumes traffic arrives on your site, that assumption is the one under pressure. It remains a test with **select advertisers in the United States**; the number of participating advertisers is not published.

Models & Products

OpenAI announces Astra for Law: selected firms first via Trusted Access, API coming soon

GPT-6 Astra packaged with settings, tools and context for legal work.

On 17 September 2026 OpenAI introduced **Astra for Law**, a foundation for law firms and legal technology companies to build AI products and workflows around their own expertise. It combines **GPT-6 Astra, OpenAI's latest and most powerful model, with settings, tools and context tailored to professional legal work**. **Availability is staged: Astra for Law will initially be offered to selected law firms through Trusted Access in ChatGPT and Codex, and will be coming soon to the API.** It appears in the model picker as "GPT-6 Astra Law" and in the API as `gpt-6-astra-law`. **API customers including Harvey and Legora will be able to build on it** and bring it into their own products. OpenAI says it will carry these legal capabilities forward to newer frontier models as they ship, and that it is expanding its privacy and governance work to give law firms specific controls for confidential material.

So What

A vertical layer — settings plus tools plus context — is being offered officially on top of the general model, with other companies invited to build on that layer. If you do vertical consulting work, **the platform moving into that layer changes where your differentiation has to sit**; this is a marker for how far up the stack the base vendors intend to go.

Market & Industry

Mozilla puts Mistral behind Firefox Smart Window, with zero data retention agreed

A browser-native assistant shipping on a European model and a no-retention term.

On 16 September 2026 Mistral and Mozilla jointly announced that Mistral models now power **Firefox Smart Window (beta)**, Mozilla's AI browsing assistant. Smart Window is described as helping you make sense of complex searches, remember something important you clicked away from, and source information based on your browser tabs. On data handling, the announcement states that **conversations are not saved on Mozilla's servers by default, and partners including Mistral agree to zero data retention**. It is available to users in **France and North America**, with the **United Kingdom and Germany expected to follow later this year**, and the partners say they are fine-tuning models on regional languages and dialects. **The announcement does not say which Mistral model is used.**

So What

A major desktop entry point now ships an assistant whose selling point is **not retaining your data**. That gives you something concrete to offer people who have refused AI features on privacy grounds. You cannot estimate quality, since the model is unnamed — what is reusable here is the procurement framing.

Research

NVIDIA debuts Vera Rubin NVL72 in MLPerf Inference v6.1, up to 3.7x GB300 NVL72

The next-generation rack gets its first public inference numbers.

On 16 September 2026 NVIDIA published results from MLPerf Inference v6.1, released the same day. **Vera Rubin NVL72, in its first MLPerf inference preview submission, delivered up to 3.7x the throughput of GB300 NVL72**, submitted on two of the most demanding benchmarks in the v6.1 suite. Separately, **a 288-GPU GB300 NVL72 submission across four racks achieved 99% scaling efficiency**, with throughput growing close to linearly from a single-rack baseline. NVIDIA also reports that **software optimisations alone delivered up to 1.6x higher performance than its v6.0 submissions**, with further gains after the v6.1 submission. The Vera Rubin NVL72 result is a **preview** submission.

So What

Inference unit cost moves with both hardware generations and software work. **A 1.6x gain from software alone on the same hardware** means the GPU-hours you rent today are not a fixed quantity of capability — a reason not to lock long-term inference estimates to today's throughput.

Funding & Corporate

Cornelis announces Active Compute Fabric and $205 million in funding

A bet that the bottleneck is GPUs waiting on data, not the GPUs themselves.

On 14 September 2026 Cornelis introduced **Active Compute Fabric**, an open architecture spanning scale-up and scale-out networking with programmable compute built into the fabric, and announced **$205 million in funding**. It combines lossless transport, in-fabric acceleration and programmable compute so the network can operate on data in flight and offload collective operations. **Cornelis models from public data that in a 100,000-GPU system roughly half of all GPU hours are spent waiting for data, worth about $1.68 billion a year in wasted capacity and 500 GWh of power.** The architecture uses UALink and ESUN for scale-up and Ultra Ethernet for scale-out, and supports a broad range of accelerators. Qualcomm Technologies joined Cornelis at AI Infra Summit to give its view on networking in efficient AI infrastructure. **The release states only "$205 million in funding" — no round name, valuation or investor breakdown is given.**

So What

Inference-cost discussions tend to fixate on per-GPU performance; this puts real money behind the claim that the waiting is the problem. Even if you never own hardware, **this is one of the layers that feeds into what renting costs** — if Active Compute Fabric delivers the efficiency claimed, that could eventually show up in what you pay. The waste estimate is the company's own modelling, with no third-party check.

Regulation & Safety

OpenAI publishes an Australian Youth Safety Blueprint

Six pillars for how AI should work for teenagers, and how companies should be held to it.

On 18 September 2026 OpenAI introduced the **Australian Youth Safety Blueprint**, framed as a roadmap for protecting young people using AI and a practical contribution to the Australian policy landscape. The **six pillars** span AI literacy, age-appropriate safeguards, privacy-protective age assurance, connections to real-world crisis support and accessible parental controls, and set out **how companies should be held accountable for identifying and addressing risks to young people**. OpenAI notes that in August it began rolling out **ChatGPT for Teens in Australia**, a default experience for users identified as aged 13 to 17 with safeguards designed around their developmental needs, building on existing parental controls, under-18 policies and age assurance.

So What

Vendors continue to submit their own proposed regulatory shapes into national policy processes. OpenAI called for mandatory federal AI safety regulation earlier this month; **this is the same argument brought down to a single country's implementation**. If minors can reach your service, expect age assurance and parental controls to be treated as table stakes.

Models & Products

Sakana AI moves Sakana Chat's Fugu option to Fugu Max and adds memory

An orchestrator that routes each prompt to a suitable open model is now selectable by every user.

On 17 September 2026 Sakana AI updated Sakana Chat. **Fugu Max, the orchestrator model it had recently released via API, is now available to all users** and selectable from the model picker. Fugu Max brings together a diverse pool of open models and routes each prompt to the model best suited to it, aiming for **performance on par with frontier models without relying on any single one**. The update also adds memory, carrying over what you have told the product — your role, your preferences on style — into later conversations. Memory works with both Namazu and Fugu Max, can be inspected in settings and can be turned off. **Only conversations from this update onward are remembered; earlier ones are not.** Sakana Chat launched in March 2026 alongside the alpha of Sakana Namazu, and an August update added Sakana Fugu as a choice plus code execution and file attachments.

So What

Instead of picking one strong model, route each request to whichever open model fits — and here that is open to every user of a shipping product rather than sitting behind an API. **It is a reference point for whether building your own routing layer is worth it.** The announcement describes the design and availability but gives no figures for routing accuracy or cost savings.

Watching (no confirmed primary source yet)

  • The range starts on 14 September. Issue 4 covered 7–13 September, so a strict seven-day window would have left 14 September uncovered; the window was widened by one day to close that gap.
  • California SB 1047's 30 September signing deadline, the EU AI Office beginning high-risk audits, and generative-AI liability clauses in India's Digital India Act (review opening 21 September) are all described in law-firm and aggregator coverage, but none could be confirmed as a new event within this window from a regulator's or legislature's own source. They will be covered once an official gazette or authority page confirms them.
  • Sam Altman is reported to have told Fortune that OpenAI will not go public in 2026. That is not an OpenAI publication, so it is not an item here; it will be covered if the company publishes it or files it.
  • Cornelis's $205 million appears in the release only as "$205 million in funding", with no round name, valuation or investor breakdown. This will be filled in if further disclosure appears.
  • Anthropic's "Measurements for understanding the pace of AI development inside frontier labs" carries no publication date on the page itself; the Sep 17, 2026 date comes from Anthropic's own news index. The measurements themselves are dated August 2026 and 13–20 July 2026.
  • Coverage of Meta, Amazon and Microsoft for this window is incomplete: the ai.meta.com blog index returned HTTP 400 and could not be surveyed. A different retrieval path will be used next issue.
  • TypeSafe AI's funding is included because **lead investor DCVC confirms a $40 million Series Seed in its own post dated 15 September 2026**. TypeSafe also issued a Business Wire release, but **Business Wire returns 403 (Access Denied) to automated retrieval**, so its body could not be checked and only the DCVC post is cited. The full investor list is unconfirmed, and the Jev announcement page itself carries no funding figure.

"Primary" links go to the announcing party's own publication (company blog, press release, official docs). "Reporting" links go to news coverage or third-party analysis. Figures and dates are as verified on the publication date.

← All Weekly AI News issues