The Bleeding Edge — Episode Briefing W41

Date range: 2026-10-02 to 2026-10-09 (Europe/Madrid)

Headline of the Week

AI models got cheaper to run this week, and the companies building the data centres behind them got harder to sell. Anthropic shipped Claude Haiku 5.5 at $0.10 per million input tokens with a 1M-token context window. In Australia, Nvidia-backed AI-infrastructure builder Firmus closed its IPO books short of demand. Its share-price talk slid from $9 to $8.25, and its prospectus was pulled from the data room shown to fund managers. These sit on top of a crowded late-September launch cycle (GPT-6.1 Sol, Gemini 4 Argon, Claude Sonnet 5.5). There's also a report that OpenAI is holding a finished model, Astra 6.1, off the market. Inference The pressure comes from both ends. Buyers get more model per dollar almost every month. Public investors are now asking who keeps the margin when the product keeps getting cheaper. This week's answer was not the company pouring the concrete for the data centre.

Top 5

1. Firmus's Nvidia-backed mega-IPO stumbles, a public-market "no" to AI-infrastructure pricing. Bloomberg reports the Australian AI-infrastructure company closed its IPO books after the bookbuild fell short of the demand it needed. AFR's Street Talk reports bankers were trying to shrink what it called a $7.9B float, and pulled the prospectus from the data room shown to fund managers. Price talk slid from $9 to $8.25 a share, and The Australian reports $6 or $7 were also discussed. A line in Bloomberg's coverage sums up the buy-side view: "It just doesn't have a compelling valuation." Shares in Maas Group, which invested $410M in Firmus, closed 22% lower. Why it matters: Bloomberg framed the weak demand as a warning on AI funding. For executives planning on the assumption that compute capacity will always find financing, this is hard evidence that equity investors will refuse to pay any price for it. Corroborated Sources: Bloomberg: Firmus closes books as mega IPO struggles, AFR Street Talk, Bloomberg: cratering demand sends warning on AI funding.

2. Anthropic releases Claude Haiku 5.5: 1M context at $0.10 per million input tokens. Per MarkTechPost, Anthropic's new small model has a 1M-token context window, and the Batch API takes another 50% off. Anthropic estimates it is about 75% cheaper on average than Haiku 4.5. That estimate already accounts for a new tokenizer, which counts roughly 30% more tokens on the same text. Why it matters: at this price, high-volume work like classification, extraction, triage and whole-contract summarisation stops needing a budget discussion. The tokenizer change also means a simple per-token price comparison with other models no longer holds. Unverified This is the only source in our sweep, and Anthropic's own release page wasn't captured. Source: MarkTechPost newsletter.

3. Three frontier flagships land within days of each other: GPT-6.1 Sol, Gemini 4 Argon, Claude Sonnet 5.5. OpenAI released GPT-6.1 Sol, which AI Search describes as bringing stronger coding and work skills. Creators AI places Google's Gemini 4 Argon and Anthropic's Claude Sonnet 5.5 in the same end-of-September window. All three dominated this week's roundups. Why it matters: a decision about which model to use now holds for a few weeks, not a few quarters. Any procurement contract tied to one vendor's flagship is a bet on price as much as on capability. Corroborated Sources: AI Search, Creators AI.

4. OpenAI reportedly held back Astra 6.1, a finished model it chose not to ship. The AI Opportunity reports that Astra 6.1 "was ready to ship" and that OpenAI held it back. The piece places this inside Sam Altman's broader plan for the company. The reasoning behind the decision isn't in our capture. Why it matters: if accurate, a lab deliberately sitting on a ready model makes release timing a strategy lever (safety, compute allocation or competitive positioning) rather than a "ship when ready" default. Unverified Source: The AI Opportunity.

5. OpenAI DevDay: ChatGPT Sites lets you build small apps from inside ChatGPT. Lenny's "How I AI" covered how OpenAI itself uses ChatGPT Sites, demoed live at DevDay. Kath Korevec showed workflows for an incident dashboard, a weekly music playlist and a collaborative 3D game. In the same episode, John Lindquist walked through eight uses of "Jev," which he calls "the fastest, cheapest model I've ever used." His examples were a voice-controlled to-do app, data deduplication and a smart command-bar router. Our sources don't name the company behind Jev. Why it matters: internal tools like dashboards, trackers and light apps are moving off the IT backlog and into any manager's chat window. That raises a governance question before it delivers a productivity win. Unverified Sources: Lenny's Newsletter, ChatPRD: ChatGPT Sites workflows.

Humor Segment

Written by the joke-writer skill for 2026-W41: 3 jokes for the show. Fact anchors first; writer notes are not read on air.

FACT ANCHORS

• Firmus (Story 1): Nvidia-backed AI-infrastructure builder Firmus closed its IPO books short of the demand it needed. Price talk slid from $9 to $8.25 a share, and The Australian reports $6 or $7 were also discussed (Bloomberg, AFR, The Australian).
• Astra 6.1 (Story 4): The AI Opportunity reports that OpenAI held back Astra 6.1, a model that "was ready to ship." Our capture doesn't include the reason. This is unverified, single-source.
• Jev (Story 5): On Lenny's "How I AI," John Lindquist called Jev "the fastest, cheapest model I've ever used." Our sources don't name the company behind it.

THE SET

1. [E: Escalation] Firmus, the Nvidia-backed data-centre builder, tried to go public this week. Price talk started at nine dollars a share. Then eight twenty-five. Then "six or seven." Next week it's on Facebook Marketplace: "Data centre. Nvidia chips. Collection only."

2. [C: Specific Callback] OpenAI reportedly has a finished model, Astra 6.1. "Ready to ship." And they're just... not shipping it. That's not a release strategy. That's your grandma's seashell soap. It only comes out when Google visits.

3. [H: Rule of Three] A developer on Lenny's podcast says a model called Jev is "the fastest, cheapest model I've ever used." Our sources don't even say who makes it. Fast, cheap, no company name? That's not an AI model. That's a guy with a van. He'll collect Firmus Saturday. Six or seven, cash.

SHAREABLE LINES

• "Fast, cheap, and nobody knows what company it's from? That's not an AI model. That's a guy with a van."
• "OpenAI has a finished model it won't let anyone use. That's not a release strategy. That's your grandma's seashell soap."
• "Firmus IPO price talk: nine dollars. Then eight twenty-five. Then 'six or seven.' Next stop, Facebook Marketplace: 'Data centre. Collection only.'"

WRITER NOTES (not spoken)

1. Opener

• Story: 1 (Firmus IPO).
• Connector: "price talk" first reads as formal IPO bookbuilding, then turns into the haggling you get selling a used item online.
• Mechanic: E (Escalation). The falling price ladder sets up the pattern, and a classified ad breaks it. Candidate probability about 0.12.
• Plant: It plants "Collection only" and "six or seven," which both pay off in the closer.
• Delivery: Read the listing flat, like you're reading a real ad.
• Risk: Safe. The target is the IPO and the bankers, not Maas shareholders.

2. Middle

• Story: 4 (Astra 6.1).
• Connector: "ready to ship but held back" first reads as a strategy decision, then as the good thing nobody's allowed to use, saved for guests.
• Mechanic: C (Specific Callback), using a mundane household comparison. The tag ("when Google visits") ties the image to the real point: the release is timed against competitors. Candidate probability about 0.10.
• Risk: Safe. The report is unverified, so keep the word "reportedly."

3. Closer

• Story: 5 (Jev / How I AI).
• Connector: "fastest, cheapest" plus an unnamed maker first reads as an AI product endorsement, then as a cash-in-hand tradesman's pitch.
• Mechanic: H (Rule of Three). "Fast, cheap, no company name" is the pattern, and "a guy with a van" is the break. The tag calls back both "Collection only" and "six or seven" from the opener. Candidate probability about 0.08.
• Delivery: Pause before "That's a guy with a van." The callback tag is the button.
• Risk: Safe. Lindquist isn't the butt of the joke, which is why he isn't named on air.

CUTTING ROOM

• "Anthropic cut Haiku's price per token, and also made the tokens thirty percent smaller. That's not a price cut, that's a Toblerone." Cut because a non-insider needs to understand what a token is before the shrinkflation image clicks. It also slightly misstates the facts, since Anthropic says the 75% saving already accounts for the new tokenizer.
• "Google named its new model Argon, after the gas famous for never reacting to anything. Finally, an honest name for Google customer support." Cut because it's a name joke, not a news joke. The support jab is also familiar enough that the room would nod instead of laugh.
• "For ten cents, Haiku will read your entire contract. That makes it the only party to the contract that's actually read it." Cut because it fails the nod test. "Nobody reads contracts" is a recognized observation, and the reveal doesn't land anywhere surprising.

Categorised News

Market Cap / Valuation

ElevenLabs doubles its valuation to $22B. The AI Opportunity reports that ElevenLabs doubled its valuation to $22 billion around the turn of the month. The newsletter uses the raise to open its walk-through of Ben Horowitz's map of where value accrues in the AI market. Voice remains a category where a specialist company is valued like a platform. Unverified Source: The AI Opportunity.

OpenAI's annualised revenue put at about US$20B. Capital Brief's Standup cited the Financial Times for a figure of roughly US$20B in annualised revenue for OpenAI. Set against the Firmus bookbuild, the figure shows the gap between revenue at the model layer and investor appetite for the infrastructure layer. Unverified We couldn't access the FT text. Source: Financial Times.

Frontier & Big Tech

Model prices now get reshuffled every month. Creators AI's model-cost roundup lists several September shifts: GPT-5.5 doubled its price, DeepSeek V4.1 Flash beat GPT-5.6 Sol, and the late-month Sonnet 5.5 / GPT-6.1 Sol / Gemini 4 Argon wave reset the table again. The piece also flags a "token trap" (tokenizer differences distorting cost comparisons) and an "AI productivity tax." The practical takeaway is that the cheapest adequate model changes month to month, so teams should re-test on a schedule. Unverified Source: Creators AI.

Apps / Dev Tools / Platforms

Unsloth Studio treats model downloads as a supply-chain risk. Unsloth published a security write-up on what happens when a trusted model repository changes. Studio re-checks the repository before running, and any change requires fresh user consent. Hugging Face malware verdicts block flagged weight files at load time, and tools run in OS sandboxes (bubblewrap on Linux). Open-weight models are increasingly pulled straight from public hubs into production pipelines, and this is the kind of control security teams will start to expect. Corroborated Sources: Unsloth security blog, MarkTechPost newsletter.

Infrastructure & Ecosystem

Firmus's "spiky" dilution clauses favour Blackstone and Coatue. AFR's Street Talk reports that shareholder terms give Blackstone, Coatue and other strategic backers a bigger slice of Firmus. AFR says these dilution clauses are driving the IPO negotiations behind the scenes. Inference Clauses like this typically shift ownership toward protected early backers when an IPO prices low. That would make every cent cut from price talk expensive for public buyers and other existing holders. Unverified Source: AFR Street Talk.

The Firmus story turns personal. The Australian ran a piece on how the Firmus "dream unravelled" for co-founder Oliver Curtis, published the same day the books closed. Expect the post-mortem to focus on valuation discipline rather than the underlying demand for compute. Unverified Source: The Australian.

Prompting Skill of the Week

Technique: Cost-per-Task Benchmarking (the "token trap" check). Best for: choosing between models when vendors change prices and tokenizers every month. A live example this week is testing Haiku 5.5 against your current small model.

1. Pull 20 real tasks from your workload that already have known-good outputs.
2. Run each task through every candidate model with the same prompt. Log the input tokens, output tokens and cost as actually billed, not estimated from word counts.
3. Auto-grade each output against the reference using a fixed pass/fail rubric (prompt below).
4. Calculate cost per successful task: total spend divided by the number of passing tasks.
5. Pick the cheapest model that clears your quality bar, not the one with the lowest price per token.
6. Re-run monthly, because the price table changes.

Example prompt (grader):

"You are grading a model's output for a data-extraction task. Below are the REFERENCE answer and the CANDIDATE answer. Return only JSON: {"pass": true|false, "missing_fields": [...], "reason": "<15 words"}. Pass only if every required field in the reference appears in the candidate with the same value. Ignore wording and ordering differences."

Common failure + fix: comparing sticker prices per million tokens across vendors. Haiku 5.5's tokenizer counts about 30% more tokens on the same text, so a per-token comparison misstates what you'll actually pay. Wordier models also quietly inflate output costs. Fix: always measure on your own text and compare cost per passing task. If one model writes twice as much to get the same pass, the extra tokens show up in your bill.

New AI Tools

Claude Haiku 5.5 (Anthropic). A small model with a 1M-token context window, priced at $0.10 per million input tokens, with another 50% off through the Batch API. Anthropic estimates it costs about 75% less than Haiku 4.5 after accounting for its new tokenizer. It's for operations teams running high-volume classification, extraction and long-document triage that couldn't previously justify a frontier-model bill. Source: MarkTechPost newsletter.

Unsloth Studio (Unsloth). Unsloth's environment for working with open-weight models, now with consent-gated re-checks whenever a trusted repository changes. It also blocks weight files that Hugging Face has flagged as malware at load time, and sandboxes tool execution. It's for teams that run or fine-tune open models in-house and need a defensible answer to "how do you know the weights weren't swapped?" Source: Unsloth.

ChatGPT Sites (OpenAI). Builds working web pages and lightweight apps from inside ChatGPT, demoed live at DevDay. The demos covered an incident dashboard, a weekly music playlist and a collaborative 3D game. It's for non-engineers who need a functioning internal tool this afternoon rather than next quarter, and for IT leaders who will have to govern what gets built. Source: ChatPRD / How I AI.

AI Personality of the Week

Sam Altman. OpenAI's CEO had a week defined as much by what the company didn't ship as by what it did. GPT-6.1 Sol is still dominating roundups after its late-September release. DevDay put ChatGPT Sites in front of developers. The AI Opportunity reports that OpenAI held back Astra 6.1 even though it was ready, and frames that decision as part of Altman's longer-range plan for the company. Separately, Capital Brief, citing the FT, put OpenAI's annualised revenue at about US$20B. Altman matters this week because OpenAI is the clearest test of whether the model layer can keep compounding revenue while prices fall. Holding back a finished model, if accurate, suggests he is now managing release timing as a strategic asset. Sources: The AI Opportunity, Financial Times.

Catch-All

If AI becomes conscious, animal-welfare law may be the template. A Big Think print-issue preview asks how society would treat AI that might be conscious, and argues that our treatment of animals offers the closest precedent. It draws on the 2023 report led by philosophers Patrick Butlin and Robert Long, which assessed AI systems against scientific theories of consciousness. It also covers the counter-view that consciousness may depend on biological processes. For executives this is still philosophy, not policy, but it is the frame regulators will reach for first if the question becomes live. On a related note for anyone who leads by dashboard, Big Think Business ran a piece the same week on the "McNamara fallacy": when relying only on measurable data leads to the worst decision. Sources: Big Think, Big Think Business.

Sector Watch

• Retail & E-commerce — quiet week
• Banking & Financial Services — Fund managers refused to pay up for Nvidia-backed Firmus: the IPO books closed short of demand and price talk was cut from $9 to $8.25 a share. Capital-markets and credit teams now have a live reference point for pricing AI-infrastructure risk. Corroborated Bloomberg
• Energy & Utilities — quiet week
• Travel & Hospitality — Airbnb shipped what The AI Opportunity calls the biggest AI update in its history, framed around Brian Chesky's playbook for the agent era. When a large consumer-travel company rebuilds its product around AI, it raises what guests expect from every hotel and booking site. Unverified The AI Opportunity
• Construction & Built Environment — quiet week
• Healthcare & Life Sciences — quiet week

Show Notes (bullets only)

• Nvidia-backed Firmus's mega-IPO stumbles: the books closed short of demand, price talk fell from $9 to $8.25, and the prospectus was pulled from the data room.
• Maas Group, a $410M Firmus investor, closes 22% lower; Bloomberg calls the weak demand a warning on AI funding.
• Anthropic releases Claude Haiku 5.5: 1M context, $0.10 per million input tokens, about 75% cheaper than Haiku 4.5 by Anthropic's estimate.
• Haiku 5.5's new tokenizer counts about 30% more tokens on the same text, so per-token price comparisons no longer line up.
• GPT-6.1 Sol, Gemini 4 Argon and Claude Sonnet 5.5 all landed in a late-September launch cycle.
• OpenAI reportedly holds back Astra 6.1, a model that was "ready to ship."
• DevDay: ChatGPT Sites lets you build dashboards, playlists and even a 3D game from inside ChatGPT.
• A model called "Jev" earns a practitioner's title of "fastest, cheapest model I've ever used."
• ElevenLabs reportedly doubles its valuation to $22B.
• Unsloth Studio adds consent-gated re-checks and blocks flagged weight files on open-model downloads.
• Airbnb ships what one newsletter calls the biggest AI update in its history.

Weekly Patterns (Inference)

1. Inference Falling token prices and falling infrastructure valuations are the same story. When the output price drops every month, investors in the picks-and-shovels layer want a lower entry price. Firmus is the first visible casualty of that math, not the last.
2. Inference The tokenizer is now a pricing lever. Haiku 5.5's tokenizer counts about 30% more tokens on the same text, so price per token no longer compares across models, or even across versions from the same vendor. Expect procurement teams to start asking for cost per task in RFPs.
3. Inference Release timing has become strategy. Three flagship refreshes in days, plus a reportedly finished OpenAI model kept on the shelf, suggest labs are timing launches against each other rather than shipping when ready.
4. Inference Capital is splitting toward the ends. A voice-application specialist doubles to $22B while an AI-infrastructure IPO gets marked down in the same fortnight. Investors are paying for distribution and margin, not for capacity.
5. Inference Model supply-chain security is becoming a product feature. Unsloth's consent-gated re-checks are the kind of control that, within a year, will be a line in every enterprise security questionnaire about open-weight models.
6. Inference Software built by non-developers is now a governance problem. ChatGPT Sites puts working internal tools in any employee's hands. Shadow IT used to mean spreadsheets; next it means a dashboard somebody built in a chat window and shared with 200 colleagues.
7. Inference Large consumer incumbents are reorganising around agents. Airbnb's update framing ("the agent era") shows established platforms want to own the agent experience themselves rather than become a data source for someone else's assistant.