The Bleeding Edge Weekly — W41: Haiku 5.5 drops to a dime, and an Nvidia-backed IPO can't find buyers

Nvidia-backed Firmus closed its IPO books short of the demand it needed. Its share-price talk was cut from $9 to $8.25, and its prospectus was pulled from the data room shown to fund managers. In the same week, Anthropic reportedly priced Claude Haiku 5.5 at ten cents per million input tokens. Bloomberg called the Firmus result a warning on AI funding. Taken together, the two stories show the model layer getting cheaper while investors put a lower price on AI infrastructure.

The week in models

Claude Haiku 5.5: 1M context for $0.10 (unverified). According to the MarkTechPost newsletter, Anthropic's new small model has a 1M-token context window and costs $0.10 per million input tokens. 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. MarkTechPost is our only source, and we didn't capture Anthropic's own release page, so treat the numbers as reported rather than confirmed.

Three flagship models released within days. OpenAI released GPT-6.1 Sol, which the AI Search newsletter describes as having 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. The same Creators AI roundup notes that GPT-5.5 doubled its price during September and that DeepSeek V4.1 Flash beat GPT-5.6 Sol. The practical result: the cheapest model that is good enough for a given job now changes every month, not every year.

ChatGPT Sites puts app-building inside ChatGPT (unverified). OpenAI demoed ChatGPT Sites at DevDay. It builds working web pages and lightweight apps without leaving ChatGPT. On Lenny's How I AI, Kath Korevec showed three examples: an incident dashboard, a weekly music playlist and a collaborative 3D game. Small internal tools that used to wait in the IT queue can now be built by any manager in a chat window.

Unsloth treats model downloads as a supply-chain risk. Unsloth published a security write-up explaining how Unsloth Studio protects against tampered models. Studio re-checks a trusted model repository before running it, and asks the user for consent again if anything has changed. Weight files that Hugging Face has flagged as malware are blocked when they load. Tools run in OS sandboxes, using bubblewrap on Linux. More companies now pull open-weight models straight from public hubs into production, and security teams will start to expect controls like these.

Devices & robotics

Our sources had no device or robotics story worth your time this week, so we've kept this section to one line rather than pad it.

What it means for leaders

For CEOs: The Firmus bookbuild is the evidence to bring to your next board meeting. Investors declined to buy AI capacity at the price the bankers asked, and Maas Group, a $410M Firmus backer, closed 22% lower. Over the same weeks, the models themselves kept getting cheaper. The question to be ready for: if the compute partners our AI plan depends on had to raise money at Firmus's price, would our plan still hold?

For CIOs/CTOs: If the Haiku 5.5 tokenizer change is confirmed, comparing models on price per token stops working. Measure cost per successful task on your own text instead. Three flagship refreshes in one window argue for contracts that run a quarter, not a year. If your teams run open-weight models, check that you can prove the weights weren't swapped, the way Unsloth Studio's re-checks do. Our read: test Haiku 5.5 now against your current small model on 20 real tasks, and stay flexible on flagship commitments.

For AI transformation leads: ChatGPT Sites turns shadow IT into a chat-window problem. A dashboard someone builds this afternoon can be shared with 200 colleagues by Friday, before anyone has asked what data it reads. Your governance framework should cover these tools before people start relying on them. The experiment to run this month: give one team two weeks to build internal tools in ChatGPT Sites, and keep a simple register of who built each one, what data it touches and who can see it.

The pattern

Two of this week's biggest stories involve a price whose unit changed. With Haiku 5.5, if the reports are right, a "token" is no longer the same amount of text from one model version to the next. That makes the sticker price harder to compare, and it pushes buyers toward measuring cost per completed task. Firmus shows the same thinking at the infrastructure end. Fund managers looked at a data-centre business selling into a market where prices fall every month, and declined to pay the asking price for its capacity. Speculation: both sides of the AI market may be moving from pricing inputs to pricing results. Buyers will pay for finished work, not raw tokens, and investors will pay for proven returns, not floor space and GPUs. Cheap tools like ChatGPT Sites could produce enough new workloads to fill the data centres. Whether that volume justifies capacity priced at $9 a share is the open bet.

---

This post is also published on our Substack newsletter at edge-ai.forum. Subscribe for the weekly roundup direct to your inbox — fresh AI news, executive context, and devices + robotics every Friday morning.