Subscribe

Five minutes.
AI, made clearer.

Less hype. More perspective.

Read the briefing

Members read every story in the archive, re-verified. Sources and corrections stay open to everyone.

What the pen marks mean
  • rememberThe takeaway: what to remember
  • figureThe number that matters
  • evidenceThe evidence line
  • rulingThe ruling
  • claimThe claim that fails

Each mark is drawn once, as you reach it. Numbers are only circled when they come from the story’s own figures.

GPT-6 Sol costs less. Its benchmark score barely moved.

OpenAI released GPT-6 Sol and Luna priced at $2 and $10 per million input and output tokens for Sol, and $0.10 and $0.50 for Luna, about half of GPT-5.6 pricing. The launch leaned on a cost-and-mistakes pitch the same week Anthropic and Google both cut prices on their own models.

Members · 30 days free

See what the evidence actually shows.

Members read the full check on every story: what the evidence shows, why it matters to you and the one line to take with you. Every past edition, re-verified, and the Receipts Pack come with it. A$89 a year, about A$0.24 a day.

First membership: 30 days free, then A$89 a year. One introductory trial per customer. Card required; renews annually until cancelled. Cancel before the trial ends to avoid the first charge.

Open the evidence2 source pages

The claim we checked

GPT-6 Sol and Luna offer lower cost and fewer mistakes than GPT-5.6 Sol and Luna

These are quoted receipts, not a count of independent investigations. Several reports may rely on the same original source.

OfficeChai ↗
GPT-6 Sol’s hallucination rate falls from 92% to 60%, and GPT-6 Luna’s from 93% to 77%.
OfficeChai ↗
running GPT-6 Sol at max effort through the full Intelligence Index costs $1.06 per task, about 50% less than GPT-5.6 Sol’s $1.99
OfficeChai ↗
GPT-6 Sol at maximum effort scores 48, essentially level with GPT-5.6 Sol’s 47
THE DECODER ↗
GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens, while Luna comes in at $0.10 for input and $0.50 for output.
Open this check in the full collection →
Correction · 24 Sept 2026

One benchmark cannot establish a universal intelligence claim, and hallucination is not evidence of deliberate deception.
Previously: The headline said the model was not smarter, and the takeaway said it lies less.
Corrected: The headline now describes the small reported benchmark change, and the takeaway limits the conclusion to the tested measures.

The idea, illustrated01

Go inside the check.

    The full explanation appears when your reading access is confirmed.
    Next: The phishing trick behind the AI headline.

    EvilTokens used an old sign-in trick, amplified by AI.

    Microsoft's Digital Crimes Unit disrupted EvilTokens, a subscription phishing kit sold on Telegram that stole Microsoft sign-in tokens through device-code phishing. Since launching in February, the service had compromised more than 12,000 email inboxes across over 10,000 organizations, and Microsoft seized 50 websites tied to the operation.

    Members · 30 days free

    See what the evidence actually shows.

    Members read the full check on every story: what the evidence shows, why it matters to you and the one line to take with you. Every past edition, re-verified, and the Receipts Pack come with it. A$89 a year, about A$0.24 a day.

    First membership: 30 days free, then A$89 a year. One introductory trial per customer. Card required; renews annually until cancelled. Cancel before the trial ends to avoid the first charge.

    Open the evidence3 source pages

    The claim we checked

    Microsoft disrupts AI-assisted platform that compromised 12,000 accounts

    These are quoted receipts, not a count of independent investigations. Several reports may rely on the same original source.

    Microsoft Security Blog ↗
    providing cybercriminals with AI capabilities for tailoring phishing lures and analyzing compromised inboxes to identify high-value targets.
    Microsoft Security Blog ↗
    This AI-powered cybercrime platform facilitated sophisticated business email compromise (BEC) campaigns that compromised more than 12,000 inboxes in over 10,000 organizations worldwide.
    Microsoft On the Issues ↗
    In short, AI was not simply helping attackers write more convincing messages. It helped them decide who to target, who to impersonate, and how to most effectively exploit the relationship to extract as much money as possible.
    Microsoft On the Issues ↗
    Microsoft seized 50 websites used to operate the service and disabled more than 150 additional domains tied to its supporting infrastructure.
    Cryovex ↗
    Microsoft disrupted EvilTokens, an AI-powered platform used to compromise 12,000 Microsoft accounts through automated phishing and credential theft.
    Open this check in the full collection →
    Correction · 24 Sept 2026

    We drew an unsupported boundary between the phishing stage and later fraud. The actual claim says AI-assisted, which Microsoft's account supports.
    Previously: The ruling was True, but. The headline and check said AI ran the fraud but did not assist the break-in.
    Corrected: The ruling is Holds up. The check now includes Microsoft's description of AI tailoring the phishing lures as well as analyzing stolen inboxes.

    The idea, illustrated02

    Go inside the check.

      The full explanation appears when your reading access is confirmed.
      Next: An AI discovery with an open question.

      Claude flagged an enzyme system. Its function is still unknown.

      Anthropic opened a life sciences lab and said roughly 950 Claude agents spent 21 hours combing DNA databases before one flagged a repeat pattern beside a reverse transcriptase gene. The company calls the find a new enzyme system it named ART, with CRISPR-like repeats, and released it as a pre-print rather than a peer-reviewed paper.

      Members · 30 days free

      See what the evidence actually shows.

      Members read the full check on every story: what the evidence shows, why it matters to you and the one line to take with you. Every past edition, re-verified, and the Receipts Pack come with it. A$89 a year, about A$0.24 a day.

      First membership: 30 days free, then A$89 a year. One introductory trial per customer. Card required; renews annually until cancelled. Cancel before the trial ends to avoid the first charge.

      Open the evidence2 source pages

      The claim we checked

      Claude autonomously discovered a novel enzyme system that is associated with an array of DNA repeats, a pattern reminiscent of CRISPR

      These are quoted receipts, not a count of independent investigations. Several reports may rely on the same original source.

      Anthropic ↗
      After 21 hours spent searching this data by roughly 950 agents using 210 million tokens, one of the agents spotted something remarkable
      Anthropic ↗
      Although we don't yet know its function, the system that Claude discovered has a set of characteristics that have only ever been found together in a handful of other systems
      Anthropic ↗
      This is an exciting example of how AI agents can contribute to biological discovery. The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation
      Unite.AI ↗
      The system's function is not yet known
      Open this check in the full collection →
      Correction · 24 Sept 2026

      Corrected the publication date and kept the explanation within what the primary source establishes.
      Previously: The Anthropic source receipts were dated September 22, and the copy was unclear about the unknown function and the status of Feng Zhang's comment.
      Corrected: The source receipts are dated September 23. The check distinguishes an unknown function, a preprint and a researcher's positive comment from an established application.

      The idea, illustrated03

      Go inside the check.

        The full explanation appears when your reading access is confirmed.
        Next: AI glasses without the camera.

        Meta's new audio glasses have no camera. Other privacy questions remain.

        Meta unveiled the Ray-Ban Meta Audio Glasses at Connect 2026, audio-only smart glasses with Meta AI, calls and music but no camera, weighing 43 grams and starting at $349. The launch responds to criticism that camera-equipped AI glasses let wearers secretly record people, a habit that earned them the nickname "pervert glasses."

        Members · 30 days free

        See what the evidence actually shows.

        Members read the full check on every story: what the evidence shows, why it matters to you and the one line to take with you. Every past edition, re-verified, and the Receipts Pack come with it. A$89 a year, about A$0.24 a day.

        First membership: 30 days free, then A$89 a year. One introductory trial per customer. Card required; renews annually until cancelled. Cancel before the trial ends to avoid the first charge.

        Open the evidence3 source pages

        The claim we checked

        the Ray-Ban Meta Audio Glasses ship with no camera, weigh 43 grams and start at $349, addressing the backlash against camera-equipped AI glasses

        These are quoted receipts, not a count of independent investigations. Several reports may rely on the same original source.

        The Verge ↗
        Plus, the camera-less glasses are lighter.
        TechCrunch ↗
        Because they don't have to house cameras, the glasses are slimmer than Meta's other models and weigh only 43 grams.
        TechCrunch ↗
        The glasses are designed by Meta's partner in its AI hardware efforts, EssilorLuxottica, and will start at $349.
        Business Standard ↗
        That is $100 cheaper than the latest version of its Ray-Ban glasses, which have cameras and will come in two new styles
        Open this check in the full collection →
        Correction · 24 Sept 2026

        The cited hardware specifications support a camera claim, not a blanket privacy guarantee.
        Previously: The takeaway described removing the camera as the fix for the privacy complaint.
        Corrected: The takeaway distinguishes camera recording from the separate questions around audio and data handling.

        The idea, illustrated04

        Go inside the check.

          The full explanation appears when your reading access is confirmed.
          Next: A training bill with a missing subtotal.

          Xiaomi's '$3M' top open model traces to one hedged tweet at $2.6M.

          Xiaomi released MiMo-V2.6-Pro, a 1.02-trillion-parameter open-weights model that Artificial Analysis scored 46 on its Intelligence Index, the top mark among open models, alongside a smaller Flash version. A widely shared newsletter headlined the release as "trained for $3M," but the only sourced cost figure in that same report is $2.6 million, and it covers the reinforcement-learning run alone.

          Members · 30 days free

          See what the evidence actually shows.

          Members read the full check on every story: what the evidence shows, why it matters to you and the one line to take with you. Every past edition, re-verified, and the Receipts Pack come with it. A$89 a year, about A$0.24 a day.

          First membership: 30 days free, then A$89 a year. One introductory trial per customer. Card required; renews annually until cancelled. Cancel before the trial ends to avoid the first charge.

          Open the evidence2 source pages

          The claim we checked

          MiMo-V2.6-Pro is Xiaomi's new top open-weights model, trained for $3M

          These are quoted receipts, not a count of independent investigations. Several reports may rely on the same original source.

          Latent Space ↗
          Artificial Analysis says MiMo-V2.6-Pro debuts as the top open-weights model on its Intelligence Index (46), with 1.02T total / 42B active parameters and strong cost efficiency at $0.435/M input and $0.87/M output tokens.
          Latent Space ↗
          @zephyr_z9 cites 130 hours, 75B tokens, and $2.6M for the RL run behind the result
          Latent Space ↗
          If these numbers hold up, the implication is that post-training/RL is becoming a far cheaper route to frontier-adjacent gains than many assumed.
          i-SCOOP ↗
          Pro and Flash each completed 30 large RL steps covering roughly 750,000 trajectories in under six days, at reported costs of about $2.62 million for Pro and $850,000 for Flash.
          Open this check in the full collection →
          Correction · 24 Sept 2026

          Different websites do not necessarily provide independent validation. Our source set cannot prove that no confirmation exists anywhere.
          Previously: The check called the second write-up independent reporting, and the takeaway said nobody had confirmed the narrower cost.
          Corrected: The check identifies two reports and states what these particular receipts do not establish.

          The idea, illustrated05

          Go inside the check.

            The full explanation appears when your reading access is confirmed.
            The finish line

            Edition complete

            You’re up to speed.

            That’s the 24 Sept 2026 briefing. Keep the useful bits. Leave the noise.

            Reading estimate: 880 words at 200 words per minute. Source quotes and the optional sections below add reading time.