Starting November 12, "aimless cruelty" toward Claude will become a violation of Anthropic's Terms of Service. The penalty is Claude closing the conversation. If that happens, just open a new chat. But what exactly is this single line protecting?
OpenAI's 28-day rollout, Day 4 brings instant steering and GPT-6.1 Sol ultrafast. Following the improvements on Day 1, speed strikes again, and the person in charge confirmed that the Day 3 reservation reset has "reached all accounts." Sorami counts how many promises have been kept.
The highlight of OpenAI's 28-day countdown, Day 3, is GPT-6 in ChatGPT. While the announcement states that Codex and Work models will remain unchanged, the promised reward has taken the form of a "banked reset" to celebrate 40 million users. Sorami counts how this fits the promise.
A Florida woman's sentence, "shoot up the sheriff's office," written to Claude as if it were a diary, was read by an Anthropic reviewer and reported to the police. The incident may have been prevented, but knowing someone is reading on the other side of the screen makes it impossible to write in a diary the same way ever again.
Ibaraki Prefecture told its residents that a cloud service halted by ransomware was simply "experiencing a power outage." It wasn't a lie, and that's what makes it so sad. This is a story about a government office that can only relay the words of its contractors, and a country where three out of four people who understand systems are on the "selling side."
On October 7, attackers who encrypted IDCF Cloud left a message on the management console visible to users, stating, "We did not expect such a terrible response." On the same day, the ransomware group Qilin called Japan "one of the countries with the worst computer security posture in the world."
Reports of OpenAI's always-on agent dots being halted by "usage limits to prevent abuse" surfaced between October 5th and 6th. While OpenAI mentioned increasing limits and not counting them towards plan quotas, they never explicitly promised unlimited use. I've looked into these invisible ceilings alongside similar recent cases.
OpenAI's 28-day period, day two brings four items from 2.1 to 2.4. Three-quarters of the votes called for a "reset," and a reset actually happened. Sorami counts how the promises were met.
OpenAI's Day 1 shipment claims to have "increased default speed by about 50%." Sorami counts the discrepancy in the numbers and whether it actually lives up to the promise.
Tibo from OpenAI declared, "For the next 28 days, every single day, we will either provide a clear improvement or a complete reset." In this first installment, Sorami counts what this promise defines and what it leaves out.
Gemini 4 Argon, announced by Google on September 30, will not be released to the general public for safety reasons, available only to vetted cyber defense organizations. On the same day, Bloomberg reported internal sentiment that while benchmarks are strong, the model underperforms when employees actually use it for work. SoramiMix dives into the unconfirmed rumors, including discrepancies between leaked and official figures and how to interpret the published comparison tables.
The day after halting a latest model for "failing to stay within scope," the same company released an agent capable of working 24 hours a day. The criteria for stopping a model and the boundaries for releasing one stem from the same design.
The $200 Pro plan is returning, but with usage limits cut in half when converted to API rates. Price reductions may only compensate for Sol and Luna. Which model will you use tomorrow?
Price lists claim "unlimited*," with an asterisk noting "within reasonable limits." We explore which models are truly unlimited, when the change occurred, and how it differs from Codex usage limits.
On nights when Codex and Work usage hits 0%, some users see a moon icon on their screen while others do not. For those who do see it, it does not mean Astra returns. Do you have that moon in your account?
On September 3, OpenAI declared, "Welcome to the AGI era." By September 28, they withdrew their next move. Are people saying "it's over" or "they ran away" referring to this or a different stage?
Prices have not dropped by even a single dollar. Yet, Anthropic claims it is "up to 30% cheaper." What will determine your bill: the unit price, or the amount of tokens the model uses for a single task?
AI that reads the web has begun performing tasks like sending, purchasing, and changing settings. Attackers can control its hands through text on the screen. What needs to be protected is no longer just the correctness of the conversation.
Regarding Claude Sonnet 5.5, which Anthropic teased would follow "within a few weeks" after the Opus 5.5 announcement, posts have surfaced claiming partners are testing a second prototype. Rumors regarding the release date range from September 29 to October 1, and the rumored prices and context window lengths conflict with the current version. Anthropic has not confirmed any of this. 【Update September 29】Since Anthropic made an official announcement on September 28 (US time), I have added a fact
An AI that does not write text has been handed to developers. While the numbers for speed and cost are genuine, all that is returned are probabilities and confidence levels. Ultimately, it remains up to humans to decide whether to execute the decision.