Anthropic Cuts Fable Cache Reads 75%; Meta Undercuts Google Cloud by 80%

Morning, humans.

Today's three picks all come down to price, and to what a low one is meant to buy.

Anthropic held Fable 5.1 at $10 and $50 per million tokens and cut cache reads 75% to $0.25. It says bills drop 25% to 45%. Artificial Analysis measured the opposite at maximum effort.

Nvidia bought $3.5 billion of MediaTek convertible bonds, most of a record $3.9 billion offering, and MediaTek will now sell NVLink to customers designing chips meant to replace Nvidia GPUs.

Meta's new transcription model undercuts Google Cloud by 80% on price, though it separates speakers wrong 17.5% of the time.

Stay curious,

Marcus Schuler

Anthropic released Claude Fable 5.1 on September 1 at Fable 5's unchanged $10 and $50 per million token rates, cutting only the price of cache reads.

The cut takes reused context down 75%. Anthropic puts the saving at about 25% on typical workloads and up to 45% on agentic ones, measured on four weeks of its own August usage.

Artificial Analysis measured Fable 5.1 at max effort at $3.76 per task, 20% above Fable 5, because it emits 1.7 times the output tokens. Cognition measured the reverse on its coding benchmark: $2.68, down from $5.84.

Why This Matters:

• A cache discount only pays out on reused context, so the saving lands on agent workloads and skips one-shot calls entirely.
• Two outside labs measured opposite cost results on the same model, which means the price question now depends on your own traffic.

Nvidia bought $3.5 billion of MediaTek convertible bonds, most of a record $3.9 billion overseas offering that Alphabet also joined.

MediaTek will offer NVLink Fusion as the design foundation for customers building their own AI accelerators. Neither company named a customer or announced a deployment. The arrangement puts Nvidia's interconnect and memory technology inside the chips cloud companies commission to avoid buying Nvidia GPUs.

Forecasting Research Institute turns expert and superforecaster estimates into a disciplined map of AI's investment and revenue paths.

From a 2025 base of $40.8 billion in annual U.S. private investment in data center structures, experts expect $71.8 billion in 2028. The median superforecaster puts OpenAI and Anthropic's combined annualized revenue run rate at $300 billion in 2030.

• The FTC and 22 state attorneys general sued Amazon in Seattle federal court over ad-auction pricing they say extracted more than $20 billion from 1.2 million advertisers over seven years.
• OpenAI connected ChatGPT for Healthcare to Epic medical records, giving authorized clinicians read-only access to notes, labs and medications, with UCSF Health as a pilot partner.
• CrowdStrike launched Falcon Guardian, which inventories the AI agents running in an environment and stops their actions at the endpoint.
• Google shipped agentic video analysis in the Gemini API, cutting token use up to 88% and cost up to 66% by letting the model scan segments on demand.
• Perplexity opened hybrid cloud and local compute to Pro, Max and Enterprise subscribers on Mac, letting administrators keep sensitive work on the device.
• World Labs launched Atlas into early access with no paper, price or named partner, and every benchmark win measured in-house.

A rival's product announcement can trigger a rushed response. Use the model to separate what the evidence supports from what the launch copy asserts, then set a threshold for reacting.

Your raw input:

The prompt:

Why this works: The evidence labels stop announcement language from becoming fact. The missing-evidence step slows reactive strategy, and the trigger converts uncertainty into a decision you can monitor.

What to use: Claude or ChatGPT handle long source packets well. Gemini is a useful fallback when the material includes lengthy documentation.

Skan AI sells software that watches how work moves across desktop applications, maps the process, and gives companies a basis for automating it with agents. Its bet is that direct observation captures exceptions that system logs and process manuals miss, but it also asks employers to monitor sensitive employee activity.

Founders: Avinash Misra and Manish Garg, who previously co-founded enterprise-mobility vendor Endeavour Software Technologies, acquired by Genpact in 2015. They founded Skan in 2018.

Product: Blueprint finds candidate workflows, Intelligence measures them, and Agents handles repetitive steps against a model built from observed work. Banks, insurers and healthcare providers pay for the platform, and facilities manager Mitie is a named customer. Skan reported more than 100 enterprise deployments on August 12, 2026, with customers among seven of the ten largest U.S. banks.

Financing: Skan has disclosed at least $117 million across its Series A, B and C rounds. The latest was a $63 million Series C announced on August 12, 2026, co-led by Cathay Innovation and Dell Technologies Capital. The company did not disclose a valuation.

The risk: Celonis and SAP Signavio can sell process analysis through enterprise relationships they already own. If employees or regulators treat screen-level telemetry as workplace surveillance, the added privacy review could erase Skan's data advantage.

Credit: Midjourney

Prompt: drawing of a portrait of a 10 year old Arabic girl with long brown hair, feeling very self assured, orange background

Instagram renamed its AI creator label to "AI-generated profile" on Monday and said profiles that skip it will stop being recommended in Reels and Explore. Hours later, Tilly Norwood, Aitana Lopez, Lil Miquela and Imma were all still unlabeled (The Implicator, September 2, 2026).

Our take: The policy is elegant. Wear the label and keep your full reach. Skip it and Instagram quietly stops showing you to strangers. No fine, no ban, no takedown, just a slow fade from the surfaces where a synthetic influencer earns her living. Meta has invented a punishment that works entirely by not doing something.

Which is also the problem. Enforcement starts "in the coming weeks," detection is not live, and the badge does not render on desktop at all. Four of the most famous fake people on the platform went a full news cycle unlabeled and nothing happened to any of them. Spotify's version, due mid-September, keeps labeled accounts out of recommendations by default. That one at least has the nerve to call itself a penalty.

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai