‹  All issues

Image: Anthropic

The small tier gets 90% cheaper, and the chips go on credit

Anthropic cut the price of its smallest model by 90% below 100,000 tokens, nine days after three labs put near-frontier work at US$2 and US$10. In the same week the money behind the chips changed shape: Broadcom would lend Anthropic US$42bn, SpaceX is raising US$40bn from Apollo, and Amazon is moving US$8bn of Nvidia inventory off its balance sheet. Inference keeps getting cheaper to buy and more expensive to finance. Those two lines cross on somebody’s books.

The week’s big stories

1. Claude Haiku 5.5 cuts the small-model tier by 90%

Announced 7 Oct. US$0.10 input and US$0.50 output per million tokens below 100,000 tokens, against Haiku 4.5’s US$1 and US$5. Above that line it is US$0.50 and US$2.50, a 50% cut rather than 90%. Cache reads are US$0.01. It is the first Haiku with an effort setting. Anthropic’s figures put it at 39.2% on Terminal-Bench 4.0 against Haiku 4.5’s 0.0%.

Why it matters. Last week’s story was near-frontier work at a fifth of the top price. This puts working tool use and computer use at a tenth of that again, which changes what is worth doing with a model at all, not which model to pick.

For builders. Move high-volume classification and field extraction to Haiku 5.5 and hold prompts under 100,000 tokens, because the price multiplies by five above it. A 20,000-token document then costs US$0.002 to read.

Sources: Announced · Anthropic · AWS · Previously · 1 Oct

2. OpenAI starts selling ads in ChatGPT and ships a visual interface to everyone

Announced 5 and 7 Oct. Visual display ads begin appearing later in October beside images ChatGPT generates, in the United States first, on the free and lower-cost tiers. The named measurement partners are the standard mobile performance set: AppsFlyer, Adjust and Kochava for click attribution, Fospha and INCRMNTAL for full funnel, DoubleVerify for brand suitability. Two days later GPT-6 and a feature OpenAI calls Intelligent UI went live globally, filling answers with interactive charts, calculators and buttons. ChatGPT has 1.2 billion weekly users.

Why it matters. The assistant with the largest audience now has an advertising business and an interface built to show things rather than say them. That partner list means direct-response placements sold on installs and conversions, not brand impressions.

For builders. Anything that answers a question inside a general assistant now competes with a paid placement for the same attention. Price products on owned surfaces.

Sources: Announced · OpenAI on ads and on GPT-6 (403, feed summaries) · Reported by TechCrunch on the ads and on the interface

3. Oracle, Broadcom and SpaceX go to the debt markets to pay for chips

Reported 1 to 7 Oct. Reuters reported Broadcom plans to lend Anthropic up to US$42bn so Anthropic can lease Broadcom’s chips. SpaceX is in talks with Apollo, banks and Pimco to raise US$40bn for Nvidia chips. Amazon is exploring a deal to shift about US$8bn of Nvidia chips off its balance sheet. Lambda took a US$1bn GPU-backed loan and is raising up to US$4bn more. The Wall Street Journal named Apollo, Blackstone and Goldman Sachs among the lenders.

Why it matters. The buildout used to run on cash flow and equity. This week it ran on secured lending against the chips, with chipmakers lending to their own customers. The cost of inference becomes partly a function of credit spreads, and CNBC notes yields on SpaceX’s bonds have risen since its June listing. A price cut funded by a loan is a different promise from one funded by a margin.

Sources: Reported by WSJ (paywall, headline only) · Semafor, citing Reuters · Bloomberg (paywall) · TechCrunch · Previously · 1 Oct

4. Home Affairs has machine learning decide who gets into South Africa

Reported 2 Oct. The Electronic Travel Authorisation system now issues yes-or-no short-stay visa decisions, with a person involved only on exception. “Human intervention is required where the system detects an anomaly that needs further adjudication,” said Carli van Wyk, spokesperson for home affairs minister Leon Schreiber. It checks more than 40 passport security parameters and queries criminal databases, and has cut processing from weeks to 24 hours. It runs at four airports for applicants from 53 countries, against a backlog of over 306,000 paper applications.

Why it matters. This is the largest automated-decision system the South African state has put in front of the public, and it went live with no published framework behind it. POPIA section 71 restricts decisions made solely by automated means that carry legal consequences for a person. Nobody has tested it on a visa refusal.

For builders. The state has set the working precedent here: automate the decision, route the exception to a person, keep the trail. Build credit and claims products to that shape, and make the anomaly-routing rule and its log a named feature, not a compliance annex.

Sources: Reported by ITWeb

5. Sierra, Meta, Shopify, Stripe and Walmart start writing the rules for personal agents

Announced 6 Oct. Sierra co-founders Bret Taylor and Clay Bavor published the Personal Agent Protocol, an open standard for how a consumer’s personal agent deals with a business. Participants are Sierra, Meta, Genesys, Instinct, Rocket, Shopify, Stripe and Walmart. It covers authentication, what access a consumer grants an agent, and the limits a company sets on what agents may do. Version 0.1 is due by end of October. The same week TechCrunch found Yelp, eBay, United and Delta blocking or refusing third-party agents, and Amazon still shutting out Meta’s Muse.

Why it matters. Nine days ago Shopify opened checkout while Amazon locked the door. Now the open side has a specification with the payment and merchant rails behind it. Until it ships, whether an agent can transact is decided site by site, by anti-bot defences that cannot tell a customer’s agent from a scraper.

For builders. Anything that acts for a customer should speak this protocol when v0.1 lands and hand the task back when a site refuses, rather than trying to look human. No South African merchant platform, bank or payment provider is on the list.

Sources: Announced · Reported by PYMNTS, quoting the Sierra post · TechCrunch · Previously · 1 Oct

In the market

  • Standard Bank and Absa. BCG’s Applied AI Index 2026, published 5 Oct from 1,330 executives in over 60 countries, reports 72% of Standard Bank staff as active generative-AI users while its banking IT bill rose 2% in the six months to June, and an Absa multilingual agent resolving about 40% of queries with no human. Banks spend 5.7% of revenue on AI, more than any sector, for a 1.2% revenue uplift. Spending leadership and return leadership are different lists. Reported by TechCentral
  • FNB. Added crypto trading with local exchange VALR on 6 Oct, five coins from R10, ringfenced so clients cannot move coins in or out, which FNB puts down to exchange-control compliance. The FNB app entered the South African top-free 50 at 47 that day on the Digest’s snapshots. Renting a licensed local venue beats building one, which is the pattern to watch for AI too. Announced · TechCentral
  • SAP. Launched Joule Work and Joule Assistants at SAP Connect on 6 Oct and agreed to buy TechWolf. Joule Work is live with 110,000 SAP employees, who SAP says show 20% productivity gains in finance, HR and procurement. No pricing published, which is what buyers ask first. Claims · SAP
  • DNB. Norway’s largest bank said on 6 Oct it will cut 400 technology jobs while raising AI-agent investment. HSBC, reported a day later, plans to cut up to 70% of its UK financial advisers in an AI push. The function being automated has moved from the back office to the advice desk. Reported by Finextra on DNB and on HSBC (403, read from feed summaries)
  • Siena. Raised a US$17m Series A led by York IE on 6 Oct for its Agent of Record, one persistent customer profile shared by a brand’s support, shopping, social and voice agents. SPANX reports half its conversations automated and handling time halved. Memory across channels is the product, not the model. Claims · VentureBeat
  • Vambo AI. The South African startup released MORENA on 7 Oct, open weights at 1.5 billion parameters trained from scratch for 12 African languages including isiZulu, isiXhosa and Setswana, plus English and French. Founded April 2023 by Chido Dzinotyiwei and Isheanesu Misi. No licence or weight host is named, which is the first thing to check. Announced · Disrupt Africa
  • Entelect. Named a Preferred partner in Anthropic’s Claude Partner Network Services Track on 7 Oct, with 2,500 engineers and 30 customers across seven countries. The tier needs a certified delivery team and public customer references, so what is being built locally is a reference list, not a capability gap. Announced · ITWeb

New companies

  • Bujeti (Nigeria). Opened Brain to the public on 8 Oct: four finance agents that pull invoices from email and storage, chase collections by email, SMS and WhatsApp with payment-plan negotiation, and answer staff policy questions. No agent moves money alone. Y Combinator W23, US$2.5m seed, 2,000 finance professionals in Nigeria and Kenya, founded by ex-Paystack engineer Cossi Achille Arouko. Disrupt Africa
  • Tab (US). Out of stealth 7 Oct at a US$300m valuation backed by SV Angel and Valar Ventures, round undisclosed. A personal assistant reached over iMessage and WhatsApp with its own phone number, computer and wallet. Worth watching for the surface: messaging as the agent interface is the South African shape too. TechCrunch
  • lensX (Cameroon). Launched 5 Oct with a cross-border marketplace for African buyers and a vendor listing tool, no funding disclosed. Most African commerce software still competes on getting listings online at all. Disrupt Africa

Opportunities

  • Document work priced per thousand, not per document. Haiku 5.5 landed on 7 Oct at US$0.10 and US$0.50 per million tokens below 100,000 tokens. BCG’s index on 5 Oct has banks spending 5.7% of revenue on AI for a 1.2% revenue uplift, so the buyer is committed and not satisfied. The opening is the high-volume, low-judgement layer those budgets have not reached: statement classification, affordability-document extraction, claim triage, sold on volume at a per-item price that replaces a BPO line. Catch: the price multiplies by five above 100,000 tokens, and cost per task and cost per token move separately. Anthropic · TechCentral

  • The consent layer for South African outbound messaging. The National Consumer Commission launched the opt-out registry on 7 Oct: enforcement from 15 April 2027, penalties to R1m or 10% of turnover, and the marketer stays liable for agencies acting for it. WhatsApp service and in-window utility messages became chargeable here on 1 Oct. Every agent that chases an invoice now needs a checkable consent record per contact per channel, and no local vendor sells that. Catch: DMASA and the existing direct-marketing stack will claim this ground, and the draft guidelines are open for comment until 17 Oct. ITWeb · Previously · 1 Oct

  • A Western open-weight option for in-country hosting, at a price. Mistral released Large 4 on 6 Oct, 1 trillion parameters and 38 on the Artificial Analysis Intelligence Index, the highest from outside the US and China, weights promised by end of October. Reflection published Beam on 5 Oct, 501B parameters under Apache 2.0, weights also due this month. For a South African buyer who cannot send data offshore, this is the first non-Chinese self-hostable option near the frontier. Catch: Artificial Analysis puts Large 4 at US$1.13 per task against GLM-5.3-Flash’s US$0.25, so sovereignty costs four times as much, and neither set of weights has shipped. Mistral · Artificial Analysis · Reflection · Previously · 1 Oct

Model and API releases

Model or product Org What is new Availability and pricing Anchor Link
Claude Haiku 5.5 Anthropic First Haiku with an effort setting; built for subagents and high volume Bedrock, Azure, GovCloud, 7 Oct; US$0.10 / US$0.50 per M under 100k tokens, US$0.50 / US$2.50 above Terminal-Bench 4.0 39.2% against Haiku 4.5’s 0.0% Anthropic
Mistral Large 4 Mistral 1T-parameter multimodal MoE; open weights promised end of October Mistral Studio preview, European datacentres, 6 Oct; US$1.36 / US$4.18 per M, 50% off for two weeks Artificial Analysis Index 38 Mistral
Beam Reflection 501B MoE, 23B active, 1M context, Apache 2.0; weights later in October Waitlist early access, 5 Oct; pricing not published SWE-bench Verified 80.9 Reflection
EmbeddingGemma 2 Google DeepMind 740M params; text, code, image, video and audio in one space; 8K context Apache 2.0 on Hugging Face and Kaggle, 6 Oct MTEB Code 78.68, up from 68.76 Google
MAI-Transcribe-2-Streaming Microsoft AI Streaming transcription, 60 languages, continuous language detection Foundry, Vercel, Azure Voice Live, 1 Oct; US$0.54 per hour of audio to year end Microsoft claims no. 1 for accuracy on Artificial Analysis; first partials in 100ms Microsoft AI
GLM 5.3 Z.ai, on Bedrock 753B MoE, ~40B active, 1M context, prompt caching Bedrock GA for eligible enterprise customers, 5 Oct Selectable effort, reasoning always on AWS
Nova 2.5 Sonic Amazon Speech-to-speech for real-time voice agents, better tool calling Bedrock GA, 5 Oct Expressive voices in seven languages AWS

Traction

  • Claude and ChatGPT in South Africa. Claude rose from 34 to 12 on the South African Apple top-free chart between 1 and 8 Oct, its best local position in the Digest’s records; ChatGPT went 7 to 2. In the US, Claude rose 13 to 8. From the Digest’s App Store snapshots.
  • Microsoft Copilot. Entered both top-50 free charts for the first time in the Digest’s records: the US at 40 on 6 Oct then 25 on 7 Oct, South Africa at 47 on 7 Oct, around the Surface Laptop Ultra event. From the Digest’s App Store snapshots.
  • M-KOPA. Ten million customers and record revenue of US$600m for its 2025 financial year, up 45%, adding 3 million in twelve months across five markets including South Africa. Fifteen years of asset finance, not a round. Claims · TechAfricaNews
  • Repos. DietrichGebert/ponytail gained 8,422 stars in seven days to 158,095, the largest absolute move the Digest tracks; stablyai/orca gained 4,831 to 87,548, the fastest at 5.8%. GitHub star counts, 1 and 8 Oct.

Opinions worth reading

  • Simon Willison. Pay-per-use services need default hard budget caps, not warning emails, because agents have removed the friction from spinning up something that bills. Read it before the next agent gets its own API key. simonwillison.net
  • Nathan Lambert, Interconnects. The cyber-risk case against open weights is unfalsifiable as stated and the evidence points the other way, because documented attacks have mostly come through closed models. The week’s sceptic, aimed at a report last week’s issue acted on. interconnects.ai
  • Gary Marcus. OpenAI’s mathematics release, which The Information counts at over 700 papers, “would never pass peer review”: no failure rate, no method, no word on whether proofs were one-shot or Lean-checked. A template for questioning any result a vendor hands over. garymarcus.substack.com

Builder notes

  • anthropic-sdk-python v1.12.0 on 7 Oct adds claude-haiku-5-5, typed computer and browser toolset calls, lifecycle-stage fields on /v1/models, and display_name on RBAC roles while deprecating name. Pin the SDK, switch RBAC scripts to display_name, and cap Haiku 5.5 prompts at 100,000 tokens. Release
  • The Claude API documented domain restrictions for web search and web fetch on 7 Oct. Put an allowlist on every retrieval agent in production. Docs
  • MCP Python SDK v2.3.0 requires httpx2>=2.10.0, and an invalid x-mcp-header annotation now raises InvalidSignature at registration instead of being dropped silently by clients. Unpin httpx2 and fix the annotations first, because servers that used to start will fail fast. Release
  • GLM 5.3 reached general availability on Bedrock on 5 Oct, eight days after Anthropic’s Frontier Red Team reported its safeguards can be bypassed 64% to 100% of the time. A managed endpoint does not change the weights. Keep it out of production routers and say why when a buyer asks for the cheapest option on the menu. AWS
  • Google limits free Gemini users to Flash Lite from 9 Oct; standard Flash now needs the US$4.99 AI Plus tier. Anything shipped with a free-Gemini fallback needs a paid key this week. The Verge (403, feed summary)
  • Apple said on 2 Oct it will require explicit user action for macOS Full Disk Access, citing autonomous-agent risk, while Microsoft is giving Copilot local file access and the run of Windows. Ask for the narrowest scope desktop tooling can work with. TechCrunch · The Verge (403, feed summary)

Money, markets & policy

Company Amount Valuation Lead Date
SpaceX US$40bn debt, in talks Not disclosed Apollo, banks, Pimco (reported) 7 Oct
Anthropic / Broadcom US$42bn loan to lease chips (reported) Not disclosed Broadcom 1 Oct
Amazon ~US$8bn of chips off balance sheet (reported) Not disclosed Not disclosed 2 Oct
Lambda Up to US$4bn US$14.5bn pre-money Coatue, Blackstone 6 Oct
Manus More than US$500m Not disclosed Not disclosed (reported) 8 Oct
Armadin US$255.5m US$2.5bn Not disclosed 1 Oct
Supabase US$150m, plus Turso acquisition Not disclosed GIC 2 Oct
Nous Research US$90m Series B US$1.5bn Not disclosed 7 Oct
SAP / TechWolf Undisclosed Not disclosed Acquisition 6 Oct

Big tech and structure. The Information reported on 5 and 6 Oct that Microsoft has cut internal Claude spending by a third and that Meta is also moving staff off it, alongside a piece on a Chinese gray market in resold Claude access. None of the three could be read, so treat the direction as reported and the detail as unknown. It is the retention question from three weeks ago in a new form: whether the largest buyers trust the supplier enough to standardise on it. The Information and on its customers (paywall, headlines only) · Previously · 17 Sep

Governance and courts. OpenAI parted ways with three safety researchers on 1 Oct over allegedly sharing confidential information with an AI safety group, and a fourth resigned two days later saying the culture is broken. A federal court dismissed the Chegg and Penske antitrust suits over Google’s AI Overviews on 1 Oct. Executives from Anthropic, OpenAI, Google and Meta will testify at a New York City Council hearing on AI risk. TechCrunch · Ars Technica · CNBC · Previously · 1 Oct

Africa and South Africa. Fifty-eight African startups raised US$583m in the third quarter, up 70% on the same quarter of 2025, taking 2026 to about US$1.39bn against US$1.6bn for all of 2025. Nersa published a draft framework on 6 Oct that would bar algorithmic systems submitting bids without manual review from the new wholesale electricity platform for its first five years, with comment closing 31 Oct; participants may still connect by API. ICASA withdrew its telecom affordability and OTT inquiry notices on 2 Oct. Disrupt Africa · TechCentral · TechAfricaNews · Previously · 1 Oct

Radar

  • 9 Oct. Free Gemini drops to Flash Lite. 15 Oct. Earliest Haiku 4.5 retirement, now with a successor. 16 Oct. Surface Laptop Ultra ships at US$2,599.
  • 17 Oct. Comment closes on the NCC’s draft opt-out guidelines. 23 Oct. OpenAI shuts down gpt-4, gpt-4-turbo, o1 and o3-mini. 31 Oct. Comment closes on Nersa’s framework and its five-year bot ban.
  • Late October. Mistral Large 4 and Reflection Beam weights, and Personal Agent Protocol v0.1. Previously · 1 Oct
  • November. The Financial Stability Board’s AI principles, due in October, which the FSCA has tied its timetable to. Anthropic’s IPO window and investor day. Nutun takes Zoey to the UK. 3 to 5 Nov. Africa AI Summit, Cape Town. 24 Nov. Earliest Opus 4.5 retirement.
  • 30 Nov to 4 Dec. AWS re:Invent. AWS put its MCP Server into six more regions this week, none African, so in-region frontier hosting in Africa stays an open question. Previously · 1 Oct
  • 2 Dec. EU watermarking grace ends; OpenAI says it will watermark ChatGPT and Codex text there. 15 Apr 2027. Opt-out registry enforcement. March 2027. South Africa’s National AI Policy goes to Cabinet.
  • Weeks ahead. Whether a South African payment or merchant platform joins the Personal Agent Protocol, and whether anyone challenges an ETA visa refusal under POPIA section 71.

Get the Digest by email

The new issue in your inbox every week. Unsubscribe any time.