FORSMILE
Issue #8Published October 11, 2026 / Covering Oct 5-11日本語で読む

This Week in AI, 11 Items — Small Models Get Cheaper, and the Tools for Checking Where AI Content Came From Arrive

The week in one place

The week has two threads. The first is that models got cheaper and reached further down the ladder. Anthropic released the small Claude Haiku 5.5 ($0.10 input, $0.50 output), which it says costs about 75% less than Haiku 4.5 on average, and OpenAI extended its new GPT-6 across ChatGPT, including the Free and Go tiers. The second is where AI-made content came from. OpenAI announced text watermarking for the EU but says detection breaks down on short or edited text; Google opened SynthID Detector, which checks images, video and audio, worldwide; and OpenAI disrupted two influence operations, rating the Russian one its first Category 5. Anthropic will also apply a usage policy on 12 November that gathers deceptive activity into one section. Tools that make content got cheaper in the same week that tools for telling it apart arrived, and the second set is not finished.

Models & Products

Anthropic releases Claude Haiku 5.5 at $0.10 input and $0.50 output, which it says runs about 75% cheaper than Haiku 4.5 on average

Its cheapest and fastest small model, and Sonnet 5.5 cache reads are halved on the same day.

Released 7 October. Per million tokens, for prompts up to 100K tokens, it costs $0.10 input and $0.50 output ($0.50 / $2.50 above 100K), against $1 / $5 for Haiku 4.5 and $2 / $10 for Sonnet 5.5. Anthropic says it costs “around 75% less to run” on average; a footnote says the figure already counts that the new tokenizer uses slightly more tokens per task. It is the first Haiku with an adjustable effort setting, and the Claude Code release notes give a 1M-token context. On Anthropic’s own evaluations, Terminal-Bench 4.0 is 39.2% (Haiku 4.5: 0.0%, Sonnet 5.5: 70.6%) and OSWorld 2.1, offline subset, is 72.4% (Sonnet 5.5: 83.9%). The same day Anthropic cut Sonnet 5.5 cache reads from $0.20 to $0.10 (about 20% cheaper on most agentic work, it says) and said Max and Team subscribers get a monthly API credit for the Claude Platform this week: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team.

So What

If you run lots of small jobs such as summaries, classification, compaction or subagent work, it is worth re-checking your unit costs. Anthropic itself says Sonnet 5.5 and Opus 5.5 remain the better choice for complex agentic coding. The numbers are Anthropic’s own evaluations, and its cyber safeguards are looser than Sonnet 5.5’s but still block penetration testing.

Models & Products

OpenAI puts a new GPT-6 and “Intelligent UI” into ChatGPT’s Chat tab, reaching Free users the next day

Answers can come as diagrams, buttons, forms and tools built on the spot, and the model starts answering while it still thinks.

Announced 7 October. In the Chat tab of ChatGPT, used by more than 1.2 billion people a week, Plus, Pro, Business and Enterprise run GPT-6 Sol and Free and Go run GPT-6 Luna (paid tiers that day, Free and Go from the next day; Enterprise depends on admin settings). Intelligent UI lets the model combine text, diagrams, tappable buttons, forms, charts and small tools such as a calculator, depending on the question. The other change is answering while thinking: on questions that need web search, GPT-6 Instant starts answering 44% sooner on average than GPT-5.6 Instant, per OpenAI’s internal evaluation. The models behind Work and Codex are not changing in this release.

So What

If you use ChatGPT for research or learning, the shape of the answers changes even on the free tier. Work and Codex are outside this release, so a change you notice there is not explained by this post. The 44% and other figures are OpenAI’s internal evaluations.

Models & Products

Mistral previews the 1-trillion-parameter Mistral Large 4 and promises the weights by the end of the month

A natively multimodal model with 52 billion active parameters and an open-weight release on the way.

On 6 October Mistral opened a public preview through the API on Mistral Studio. It is natively multimodal with 1 trillion total and 52 billion active parameters, and Mistral says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres. The weights are to be released by the end of the month; until then cybersecurity leaders, vetted partners and state authorities red-team a version with reduced moderation. Mistral’s own figures are 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4, and it says the model ranks in the top five on the Artificial Analysis Cyber Index and first among open-weight models outside China. It also claims Claude Opus 5.5 and GPT-6 Astra score near zero on a vulnerability-reproduction test because they refuse the task; that is Mistral’s claim.

So What

Another high-performing model that you can run in your own environment is due at the end of this month. For now the weights are not out, and the benchmarks are Mistral’s presentation (Artificial Analysis and vals.ai are the third parties it names). The licence terms cannot be confirmed from this post, so check the release page before relying on it.

Developer Tools

Claude Code 2.1.290 to 2.1.296 adds Haiku 5.5 and lets hooks block when they fail

Seven releases between 5 and 9 October: Haiku 5.5 support, an effort setting for subagents, fail-closed hooks and more.

Versions 2.1.290 through 2.1.296 shipped between 5 and 9 October. 2.1.293 added Claude Haiku 5.5 (`claude-haiku-5-5`) as the default Haiku on the Anthropic API. 2.1.292 gave the Agent tool an `effort` parameter, so a subagent runs at the effort level you ask for. 2.1.295 added `onFailure: "block"` for command and HTTP hooks: a hook that cannot start, times out or exits with an unexpected code now blocks the action instead of letting it through. The same release added OSC 7501 support, so terminals that implement it can show whether Claude Code is working, waiting or done. 2.1.296 added an `allow_large` option to the Read tool for reading a text file past the usual size limit in one call. 2.1.291 fixes a regression in 2.1.290 where cloud sessions could drop answers to permission prompts, and a regression in 2.1.288 where the last messages of a session could be lost on quitting.

So What

If hooks are what stops dangerous operations in your setup, `onFailure: "block"` turns “a hook that did not run lets the action through” into “it stops”. Cloud sessions on 2.1.290 had a regression that dropped permission answers, so update to 2.1.291 or later. The release notes list changes; they are not test results.

Market & Industry

OpenAI adds ads shown during image generation in ChatGPT, testing in the US on Free and Go from this month

The ads are kept separate from the image, OpenAI says they do not affect answers, and measurement partners expand.

Announced 5 October. The new format is an image ad shown to Free and Go users while ChatGPT is generating an image; it is labelled as an ad and kept separate from the image being made. Testing begins later this month in the US with an initial group of advertisers, and OpenAI says advertising does not influence answers. The post also adds integrations with Hightouch, Tealium and LiveRamp, support for many attribution partners such as AppsFlyer and Adjust, and brand-suitability pilots with DoubleVerify and IAS. Example results it cites: per DV Rockerbox, WeightWatchers’ attributed cost per acquisition was 15.3% lower than its blended paid-search benchmark, and per Triple Whale, 93% of Portland Leather’s visitors from ChatGPT Ads were new.

So What

For people using free ChatGPT, this is a move toward ads appearing while an image is generated (a US test for now). The results are a handful of cases OpenAI chose, not a typical level. The post does not say when it would reach Japan.

Market & Industry

NVIDIA and Microsoft unveil RTX Spark for Windows; laptops go on sale 16 October

Windows machines with up to 128GB unified memory for local AI, plus a Windows version of the DGX Station.

Announced at an event on 7 October. RTX Spark pairs a Blackwell GPU with a Grace CPU and offers one petaflop of FP4 compute and up to 128GB of unified memory. Systems come from Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI and Gigabyte, including the Surface Laptop Ultra. Laptop preorders opened that day with availability on 16 October; compact desktops follow in November. Microsoft also made Microsoft Execution Containers (MXC), which run agents under OS control, generally available. DGX Station for Windows (GB300, 748GB of coherent memory, up to 20 petaflops of FP4) was previewed, with no availability date given.

So What

From 16 October there are Windows machines in the 128GB class for people who want to run models on their own PC. The post gives no prices. The claim that it runs a model matching many cloud models is NVIDIA’s, and performance needs independent testing after release.

Regulation & Safety

Anthropic launches the Cyber Mission, with a critical-infrastructure program and free scans for open-source projects

Eleven security and industrial firms join the infrastructure program; open-source reports arrive without human review.

Announced 8 October. The Critical Infrastructure Defense Program (CIDP) brings Claude models, on-site engineers and threat research to the providers that defend operational technology in power, water and transport. The founding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation. OSS Scanner gives enrolled open-source projects periodic scans from Anthropic’s strongest models at no charge, with a proof of concept, an explanation and a suggested fix where available. The reports are sent without human review, so some will contain errors such as a wrong severity; Anthropic expects a true-positive rate above 90%. Project Glasswing has been merged into the expanded Cyber Verification Program.

So What

If you maintain open-source software, there is a new free route to vulnerability reports. They arrive before any human check, so you should plan to verify severity and reproducibility yourself. The “above 90%” is an expectation, not a measured result.

Regulation & Safety

Anthropic revises its Usage Policy: deceptive activity gets one section, effective 12 November

Mostly clarifications, with new conditions for connecting Claude to physical equipment and a ban on sustained abuse of the models.

Published 8 October, effective 12 November. Anthropic says most changes clarify existing rules. A new section, “Do Not Engage in Deceptive Campaigns or Artificial Activity”, pulls together rules on hiding who is behind a message or amplifying it with fake accounts, whether political or commercial. The elections section is renamed “Do Not Undermine Democratic Processes”, and the blanket ban on personalized vote and campaign targeting is removed (deceptive targeting and misuse of personal data stay prohibited elsewhere). The surveillance section now states that Claude cannot be used to decide or recommend who to investigate, arrest or charge. The high-risk section adds conditions for equipment that takes autonomous physical actions: a qualified operator must be able to observe and stop it, and it must hold a safe state if Claude is disconnected. A ban on sustained, needless cruelty toward the models is added.

So What

If you have built Claude into a product, now is the time to check before 12 November whether your use touches deceptive activity, surveillance, physical equipment or the high-risk categories. Anthropic says many of the changes do not alter what it enforces in practice; how much stays the same is its own judgement.

Regulation & Safety

OpenAI says it banned “false front” influence operations from Russia and Iran, the Russian one its first Category 5

Both used AI for articles and reports, with a “think tank” staffed by unwitting people and fake journalists.

Reported 8 October. The Russia-origin “Dark Clark” ran a “Social Research Center” in Latin America under a fake persona, “Mia Clark”. Its local staff appear not to have known they worked for a Russian group, and OpenAI rated it Category 5 on the Breakout Scale, which runs from 1 to 6; it is the first Category 5 operation OpenAI has disrupted since it began reporting. The Iran-origin “Bogus Bylines” used seven fake journalist personas to pitch long-form articles to small and medium outlets worldwide and reached Category 4. Both landed content in mainstream media, not only on social platforms (not all of it generated by OpenAI’s models). Across the 30 operations it has exposed in two and a half years, OpenAI says those aiming at real outlets tend to have the highest reach.

So What

If you take outside contributions or cite material from a “research institute”, there is now one more reason to confirm that the writer and the organisation exist by another route. AI text cannot be spotted by how it looks; the pattern OpenAI points to is aiming at real outlets. The report covers what OpenAI could observe, and some attribution is inferred.

Regulation & Safety

OpenAI says it will add an invisible watermark to ChatGPT and Codex text in the EU

API use is opt-in, the detector is limited to researchers and expert organisations, and OpenAI itself says detection weakens on short or edited text.

Announced 5 October in response to the EU AI Act’s requirement that generated text be machine-readably identifiable. The technique, textGrain, embeds a statistical signal in word choice. API customers worldwide can opt in for selected models (off by default), and over the coming weeks the watermark is added to ChatGPT and Codex text on all EU plans. Applications for the detector are open, but for now only approved researchers and expert organisations get access. At a 1% false-positive rate, detection was about 80% for 200-token passages and about 95% for 400-token ones. In 400-token passages, replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%; content with little room for word choice, such as mathematics, scores lower.

So What

A way to tell AI text apart with a detector is, for now, not public, and it is unreliable on short or edited text. OpenAI states that not finding a watermark does not prove a human wrote the text. The ChatGPT rollout is EU-only, and the post does not say whether users elsewhere will get it.

Regulation & Safety

Google opens SynthID Detector to everyone worldwide in English, including content from OpenAI and NVIDIA

Anyone can check whether an image, video or audio file was made with AI; it covers media, not text.

Released 7 October. Google says it has watermarked over 180 billion images and videos and 240,000 years of audio since SynthID began in 2023. Last year’s early version was for media professionals; now anyone, worldwide, in English can use it. It checks content from Google and partners, including OpenAI, NVIDIA and Kakao, with Apple to follow. The built-in verification in Search, the Gemini app and Chrome already handles over 1 million requests a day. It covers images, video and audio, not text.

So What

There is one more place where ordinary people can check where an image or audio file came from. It can only recognise watermarks from supported providers, and Google’s post does not say what a “not detected” result means. It does not mention availability in Japanese.

Watching (no confirmed primary source yet)

  • Mistral Large 4’s weights are promised “by the end of the month”, and the licence terms cannot be confirmed from the announcement. I will cover it once they are published.
  • The EU Scientific Panel held a special meeting on 9 October after investigating “loss-of-control incidents”, and was to present recommendations to the Commission (per the Commission’s announcement). Neither the recommendations nor the companies’ answers are public; I am waiting for the content.
  • Google Cloud announced the “Gemini agent” on 8 October, but the only primary source is a short post on Google’s blog, so I could not confirm features or pricing. I will cover it when the Google Cloud details are confirmed.
  • On 6 October OpenAI published new mathematical results from an internal frontier model and said it is working to release that model responsibly. The model is unreleased and there is little to check against, so it is not an item.
  • I checked Anthropic’s $150 million, three-year Genesis Mission commitment, GitHub’s ReviewBench (5 October) and NVIDIA’s gold-medal-level Nemotron results at the IOI and IMO (7 October), but there was too little to say what changes for readers, so I left them out.
  • No funding round or acquisition this week could be confirmed from a primary source (a party’s own announcement or a filing). New announcements from Meta, Apple and the Chinese labs could not be confirmed from primary sources either.
  • I have not confirmed that GPT-6 actually reached Free and Go users from 8 October, or that the Max and Team API credits (“this week”) were delivered.

"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