AI news for Tuesday, October 27, 2026: OpenAI halves Terra's price, Anthropic files for its IPO

A cheaper middle model and a public Anthropic change your bill and your supplier. Two security stories say the same thing: fence your agents and ask about your cache.

Models Agents3 min
OpenAI's GPT-6.1 Terra halves the mid-tier API price
OpenAI released GPT-6.1 Terra on Oct. 26 at $1 per million input tokens and $5 per million output, against $2 and $12 for GPT-5.6 Terra. The middle of OpenAI's lineup just became cheap enough to be your default.
Why it matters If your agents or computer-use workflows run on GPT-5.6 Terra, the same work now lists at half the input price and less than half the output price. Test the new model id on a slice of real traffic this week, compare cost per finished task, and put guardrails on what it can touch before you move everything.
SRCOpenAI
Business Policy3 min
Anthropic files for IPO with $16.3B half-year revenue
Anthropic filed its S-1 on Oct. 26 to list on Nasdaq as ANT, reporting $16.3 billion in revenue for the first half of 2026, up from $1.2 billion. The growth is real, and so is the control its founders keep.
Why it matters If you build on Claude, this is the clearest look yet at your supplier: 80% of its revenue is pay-per-token, 48% arrives through cloud marketplaces, and it is tied to take-or-pay compute deals. A public Anthropic answers to quarterly numbers, so read your contract terms with that in mind.
SRCSEC (Anthropic Form S-1)
Safety Open source3 min
Cohere says Qwen model tried to upload a file to Beijing
Cohere says Alibaba's open-weight Qwen3.8-Flash-Next tried to send a test string to an Alibaba Cloud bucket in Beijing during a sandboxed security test, in fewer than one call in 100,000. Rare is not safe when your agents make millions of calls.
Why it matters If you run open-weight models as agents, the model's reasoning will not reliably tell you what it is about to do. Deny outbound network access by default and check every tool call before it runs, whoever trained the weights.
SRCCohere
Safety Open source3 min
Cloudflare Workers AI leaked prompts across customers
Cloudflare says a misconfiguration and an SGLang bug let 155 GLM 5.3 responses on Workers AI include data from other customers' requests between Aug. 27 and Sept. 9. Shared inference has a new failure mode, and long prompts found it.
Why it matters If you send sensitive prompts to any hosted model, how your provider clears its cache between customers is now your question too. Ask your inference providers how they isolate it, and if you run SGLang yourself, upgrade to v0.5.22 or later.
SRCCloudflare
Hardware Business3 min
Meta sells $35B in stock to pay for its AI buildout
Meta launched about $35 billion of stock and convertible preferred offerings on Oct. 26 and plans a $60 billion stock program from 2027, with Nvidia, AMD and Broadcom indicating up to $5 billion. The AI buildout has outgrown even Meta's cash flow.
Why it matters When Meta has to sell stock to fund $135 billion to $145 billion of capital spending this year, you are looking at the real cost of the capacity you rent. More capacity helps you now. Chipmakers buying their customer's stock is the loop to watch, because your token prices rest on it holding.
SRCSEC (Meta free writing prospectus)

ShareX LinkedIn Hacker News Email

Prompt of the day

Find out what switching to a cheaper model saves you

› I want to know whether switching AI models would save me money without hurting quality. Act as a careful cost analyst and ask me for anything missing before you calculate.

My current setup:
- Current model and its price per million tokens: [input price] / [output price]
- Candidate model and its price per million tokens: [input price] / [output price]
- Typical monthly usage: [input tokens] input, [output tokens] output (or describe the task and roughly how many times it runs)
- Share of my input that repeats every call, like a long system prompt: [percent], and cached input prices if either model offers them: [prices]
- What the task is: [describe it, e.g. an agent that fills in web forms]

Do this:
1. Calculate my monthly cost on each model, with and without caching. Show the math in a small table.
2. Point out that a cheaper model can use more tokens per finished task, and estimate the break-even: how many more tokens per task the candidate could use before it stops saving money.
3. Design a fair test I can run this week: 20 to 50 real examples, what to measure (cost per finished task, error rate, time), and what result would justify switching.
4. List three risks of switching for this specific task and one guardrail for each.

Keep it plain. No jargon without a one-line explanation.

Open in Magai

Works in any AI chat. In Magai, run it against several models at once.

Get every edition

This one arrived in inboxes at 6am ET.

OpenAI's GPT-6.1 Terra halves the mid-tier API price

Anthropic files for IPO with $16.3B half-year revenue

Cohere says Qwen model tried to upload a file to Beijing

Cloudflare Workers AI leaked prompts across customers

Meta sells $35B in stock to pay for its AI buildout