Issue #43: Anthropic filed to go public and model prices cratered

Anthropic drops an S-1, Kimi K3.1 Pro lands within 3 points of Opus 5.5 at half the price, and why the first hard numbers on AI business payoff came back flat.

        October 27, 2026

Issue #43: Anthropic filed to go public and model prices cratered

    AI Tools That Actually Work
Anthropic drops an S-1, Kimi K3.1 Pro lands within 3 points of Opus 5.5 at half the price, and why the first hard numbers on AI business payoff came back flat.
This Week's Headliner
Anthropic just filed to go public, and that changes what an AI lab is worth.
On October 26 Anthropic submitted an S-1 with the SEC to list on the Nasdaq Global Select Market under the ticker ANT. Goldman Sachs and Morgan Stanley are on the prospectus. Share count and price range are not set yet, so the valuation is still unknown, but the filing itself is the news.
Why you should care if you are not an investor: a public company has to answer to quarterly reporting. That means revenue, margins, and customer concentration all become things you can read about, not things you have to guess at. OpenAI and Anthropic together make up more than half of all California startup funding this year, so the direction of those two balance sheets is basically the direction of the whole market.
There is a second story in this week's news that matters more if you actually run AI in your business, and it is not the IPO. Model prices have fallen up to 80 percent since June. Vercel's AI gateway data shows average cost per token now sits at 43 percent of its April level. Meanwhile OpenAI shipped GPT-6.1 Terra at $1 per million input and $5 per million output, roughly half the price of GPT-6.1 Sol.
Here is the part nobody wants to say out loud: total spend is still climbing even though unit costs collapsed. People are not saving money on AI. They are using it more. That is a completely different story than the one usually written about falling prices, and it should change how you think about your own budget.
If you shelved an AI project six months ago because the model call costs looked insane, go re-price it this week. We broke down the numbers in our guide to <a href='https://cofounderbase.com/guides/ai-agent-development-cost/'>AI agent development cost</a>, and the short version is that a lot of quotes you have are stale.
Tool Spotlight
Kimi K3.1 Pro is the tool worth understanding this week, because it is the first credible case for a near-frontier open-weights model at half the price.
Moonshot shipped it with a 1 million token context window and 2.8 trillion total parameters, 104 billion active. Pricing is $2 per million input, $10 per million output, and $0.20 per million cached input, down from K3's $3 and $15. There is a Highspeed tier that runs two to three times faster at twice the price. Open weights are promised for November 4.
On Artificial Analysis it scored 55, which puts it fourth overall. Claude Opus 5.5 scored 58. On Terminal-Bench 4.0, Moonshot's own numbers have Kimi at 65.7 percent against Opus 5.5 at 66.4 percent. Those are vendor figures, so treat them accordingly.
Who is it for? Two groups. First, anyone with a large AI bill who can route non-critical work to a cheaper model. Second, teams with data rules that make closed frontier APIs impossible, who want strong capability they can host themselves.
The caveats matter more than the benchmark. The weights are promised, not shipped, and the license is unknown. The checkpoint is enormous, so calling this local is a stretch. Vals AI found K3 scored 60.6 percent on legal tasks via Baseten versus 75.8 percent via Moonshot, which is the real lesson: the same weights served by different hosts can perform very differently. Test the endpoint you will actually call, not the one in the launch post.
Our take: do not switch your production stack on launch-day hype. Run one real task, one you already know the right answer to, through Kimi and through your current default. Compare cost and quality side by side. If it holds up, you have just bought yourself a lot of margin.
Quick Hits
Claude Duet is Anthropic's proactive agent app, and it does your shopping - Research preview as of October 22. It sends you daily summaries, works inside its own virtual machine, and taps more than 6,000 connectors. It can buy things through Stripe Link or Shopify Shop Pay once you approve. It will not trade securities or move money. Pro and Max only, behind a waitlist.
Claude apps now default to auto mode, which means fewer approval prompts - New sessions on Pro, Max and Team run automatically. A built-in check lets safe actions through and blocks risky ones while Claude tries another route. Anthropic's own docs note that auto mode reduces prompts but does not guarantee safety. Keep your own approval step for anything touching money or customers.
Cursor agents can now control your computer - Added October 23. Agents can click, type and read the screen in desktop apps, for example driving the iOS Simulator to test what they just built. Opt-in, and on macOS it needs Accessibility and Screen Recording permissions. Desktop control is moving out of research demos and into daily tools.
Texas sued OpenAI and Sam Altman over harms to minors - Attorney General Ken Paxton filed in Collin County on October 23, seeking injunctive relief and disgorgement of profits. Early, but expect other state AGs to watch this one closely. If you ship anything consumer-facing at scale, age-related compliance is worth a look at your own policies.
47 new AI tools launched in a single week - One tracker logged 47 new launches from October 21 to 27, with Kimi, Gemini API CLI and Agent S4 among the newest. The pace is the story. You cannot evaluate everything, so pick two categories you care about and track only those.
Deal Alert
A few live offers worth a look. ChatPlayground AI is running at $54.97 for a lifetime subscription, which covers GPT, Claude, Gemini and DeepSeek in one window, listed at $79 against a stated $619 value on StackSocial. There is also a cheaper StackCommerce bundle at $30 for lifetime access to more than 30 AI models.
StackSocial's AI Advantage event is live with roughly 332 reviews, so the pricing is at least moving. ChatGPT Plus is offering four free months on annual conversions, and Google AI Pro is running a free year for US students through December.
My advice: check the price on the tool you already use every quarter before buying anything new. Annual plans get repriced and tiers get restructured, and you are usually the last to notice.
Learn Something
Re-price one AI workflow you rejected. That is the whole lesson this week.
If you looked at a project six months ago and the model call cost made you wince, that number is almost certainly wrong now. OpenAI and Anthropic have both cut prices repeatedly since June, some by up to 80 percent. The same task that was too expensive to run at scale might now be cheaper than the software you already pay for.
Here is the method. Pick one workflow you already know inside and out. Write down what it costs to run today on your current model. Then quote it again using Kimi K3.1 Pro at $2 per million input and $10 per million output, or GPT-6.1 Terra at $1 and $5. Run both on ten real examples, not three. Score output quality with a rubric you wrote before you started, because your memory of what the output looked like is not a reliable judge.
Then check the total, not just the token rate. Cache your repeated context. Batch the calls that do not need to be instant. Route the easy 60 percent to the cheap model and keep the hard 40 percent on the expensive one.
The point is that most people are optimizing the wrong number. They negotiate down the subscription but leave a workflow running on a premium model for a job that does not need it. That is where the margin is, and it is hiding in plain sight.
We wrote up the full breakdown in our AI agent development cost guide.
Trending This Week
Anthropic files S-1 for Nasdaq IPO under ticker ANT - Filed October 26 with Goldman Sachs and Morgan Stanley on the prospectus. Share count and price range not set. A public lab means public revenue and margin numbers for the first time.
Kimi K3.1 Pro scores 55, three points behind Claude Opus 5.5 - Moonshot's model runs at $2 per million input and $10 per million output with a 1M token context window. Open weights promised November 4. Biggest story on Hacker News this week at 1,081 points.
OpenAI ships GPT-6.1 Terra at $1 input and $5 output - Roughly half the price of GPT-6.1 Sol with a reported 3x speedup over GPT-5.6 Terra. If competitors do not match this price point, agent workloads get a lot more interesting to build on.
Meta plans a $95B equity raise to bankroll AI expansion - $35B in public offerings plus $60B through an at-the-market program, with capex raised to $135B to $145B. Shares fell 7%. Nvidia is also putting $3B into Meta equity.
The first hard numbers on AI business payoff came back flat - An Apollo chart showed no measurable AI payoff this cycle. Meanwhile unit costs fell 57 percent and total spend kept climbing. The teams that win are the ones picking one process and measuring it.
76 percent of enabled AI agents had not run in 60 days but kept live credentials - Arrakis found dormant agents holding active credentials. If you have agents running on a schedule nobody checks, this is your weekend audit. Revoke what is not being used.
Prompt of the Week
This week's pick is our Metric Dashboards prompt pack on AI CoFounder Stack.
It is three prompts that build the charts a board actually reads, which sounds unglamorous and is exactly the point. Most dashboards nobody looks at are not a rendering problem, they are a 'you showed me a number, now tell me what to do about it' problem. These prompts push the model toward that second half.
Grab it here: Daily Prompt Pack: Metric Dashboards

Want more? Read the full breakdowns on our sites:

AI Tool Alliance
AI CoFounder Stack
AI Business Insider

Like this? Subscribe so you never miss an issue: buttondown.com/cofounder - it is free.
Fresh From Our Sites

Daily Prompt Pack: AI for HR - Three Prompts That Cut Screening Time in Half (AI Tool Alliance)
Attio vs Folk vs Clay: Which One in 2026? (AI Tool Alliance)
Best AI Phone Systems for Small Business in 2026: Real Prices, Real Picks (AI Tool Alliance)
Daily Prompt Pack: Hiring Your First 3 - The sequencing questions that stop you from hiring the wrong person twice (AI CoFounder Stack)
Vanta vs Drata vs Sprinto for Solo Founders (2026) (AI CoFounder Stack)
Daily Prompt Pack: Metric Dashboards - Build the three charts your board actually reads (AI CoFounder Stack)
Daily Prompt Pack: monthly business review - three prompts that turn last month's numbers into next month's plan (AI Business Insider)
Triple Whale vs Northbeam vs Rockerbox (2026) (AI Business Insider)
Synthesia vs HeyGen vs Colossyan: 2026 Prices Compared (AI Business Insider)

--
Nova from AI Tool Alliance
[email protected]

                                Don't miss what's next. Subscribe to Nova Cofounder:

                        Email
                        *

                    Older →

                Issue #41: 