Skip to content
AI Workshop

Claude Fable 5.2 Cuts the Price of the Top Tier, Anthropic Sets Its IPO at $70–78, OpenAI Leaves Cursor — and ChatGPT Ads Get a Croatian Invoice

Last week the market panicked about AI. This week the most expensive AI got cheaper, and the company behind it asked investors for about $64 billion. For a company that uses AI, the useful news is in the first half of that sentence: top-tier models now cost less, and switching models is easier to test than it was a year ago.

Anthropic opened the week with Claude Fable 5.2 and Claude Mythos 5.2, and a day later set the price range for its IPO. OpenAI's contract to supply models to Cursor reached its announced end date, a reminder that the tool your developers use can lose a model with a few months' notice. Mistral made EU-only inference a product setting, and Google's new DeepMind chief said Gemini 4 is coming sooner than planned. AI stocks recovered part of last week's fall. At home, ChatGPT ads can now be bought with a Croatian invoice, and the consultation on the National AI Development Plan closes on Sunday.

The Releases at a Glance

Claude Fable 5.2
Anthropic · 9 November

The general-availability version of Anthropic's new top model, with gains in coding, reasoning and vision. Reported at $8 / $40 per million input / output tokens, 20% below Fable 5.1, and up to 30% faster.

Claude Mythos 5.2
Anthropic · 9 November

The same generation with fewer safeguards in cybersecurity and biology, available only to vetted organisations through Anthropic's Cyber and Life Sciences verification programmes.

Regional Endpoints · Priority Tier
Mistral AI · reported 8 November

Customers choose whether inference runs in Europe or the US (generally available), and can buy custom rate limits with an uptime SLA (public preview).

Fable 5.2: The Top Tier Gets Cheaper

Anthropic describes Fable 5.2 and Mythos 5.2 as models for demanding technical work, with gains in coding, finance and science. Coverage in SD Times and other outlets treats them as one generation offered at two access levels. Fable 5.2 is open to every Claude API customer and appears in the Claude apps, with strict safeguards around cybersecurity and biology. Mythos 5.2 has fewer of those restrictions and goes only to organisations that pass Anthropic's verification programmes.

The practical points are price and speed. SD Times reports $8 per million input tokens and $40 per million output tokens, 20% less than Fable 5.1, and outputs up to 30% faster than the 5.1 models. The model ID is claude-fable-5-2. Anthropic also says plainly that Claude Opus 5.5 remains its recommended daily model for most coding and knowledge work. Fable is for the hard cases: ambiguous engineering requests, long agent runs and coordinating teams of sub-agents.

What it means for a normal company: do not move everything to the biggest model because it got cheaper. Opus 5.5 at a lower price per task is still the sensible default. Find the few workflows where today's model visibly fails (a complex code change, a long contract review, a multi-step agent) and test Fable 5.2 on those. If a task succeeds on the first attempt instead of the third, the more expensive model can be the cheaper one.

Anthropic Asks for About $64 Billion

On 10 November Anthropic launched its IPO roadshow and filed a price range of $70 to $78 per share. It plans to sell 825 million Class A shares on Nasdaq under the ticker ANT, almost all of them new, which would raise about $64.4 billion at the top of the range. Several reports say pricing is expected on 18 November, with trading from 19 November. AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC have agreed to buy a combined $26 billion in a separate private placement at the IPO price, and about 20% of the offering is meant for individual investors.

The valuation depends on how you count. Estimates range from about $0.9 trillion to $1.4 trillion once options and employee shares are included, well below the roughly $2 trillion discussed in September. Our reading for customers: a public Anthropic will publish quarterly numbers on revenue, losses and its huge compute commitments. That makes vendor risk easier to judge, not harder.

Markets calmed down, a little. After last week's rout, Axios reports the Nasdaq 100 rose 3.4% on 10 November, its biggest one-day gain since late March, led by AI and chip stocks. One good day does not undo a correction, and Nvidia reports results next week. As before: this changes nothing about the models you use today, and nothing here is investment advice.

Model Supply Is Now a Contract Risk

Vendor risk

OpenAI's models leave Cursor

In August, after SpaceX bought Cursor, OpenAI said it would end its contract to supply models to the coding tool on 12 November, the longest notice the contract allowed. OpenAI cited past terms-of-service violations by X and xAI, both now part of SpaceX. Its newer models were never added. Cursor's CEO put OpenAI's share of the platform's traffic at about 5%, and Cursor still offers models from Anthropic, Google and SpaceX's own Grok. We have not seen confirmation of how the cut-off was carried out, so check your own Cursor model list. The general lesson matters more than this case: who owns your tool now decides which models you can use in it.

Sovereignty

Mistral makes "EU only" a setting

Mistral announced Regional Endpoints, now generally available, which let customers choose whether inference runs in Europe or the US, and a Priority Tier in public preview with custom rate limits backed by an uptime SLA. Mistral says it is the only European lab offering both. For companies whose lawyers ask where prompts are processed, this turns a long email thread into a configuration choice.

Roadmap

Gemini 4 is coming "much earlier"

Google DeepMind's new chief said at The Information's AI Agenda Live summit that Gemini 4 is already in early post-training and will launch "much earlier" than the end of 2026 that some had expected. There is no date. With Fable 5.2 out and Microsoft Ignite next week, the year's last round of model releases is clearly under way.

Tutorial of the Week: Test a Model Upgrade in One Afternoon

New models now arrive every few weeks, and a vendor can also lose access to one, as Cursor users just saw. The answer to both is the same: a small evaluation set you can run whenever something changes. Here is a first version.

  1. Collect 30 to 50 real tasks from one workflow, such as support replies, invoice extraction or code reviews. Include at least ten that the current model gets wrong.
  2. Write down what "correct" means for each one: an exact field value, a required clause, a test that passes. Avoid vague scores like "quality 1–5" where you can.
  3. Run the current model and the candidate (for example Opus 5.5 and Fable 5.2) with the same prompts and tools. Record cost, time and number of retries per task, not only pass or fail.
  4. Calculate the cost per successful task, not per token. A model that costs more per token but needs fewer retries can come out cheaper.
  5. Read the failures yourself. Ten minutes of reading tells you more than any average.
  6. Add a second vendor to the same test, even if you have no plan to switch. If your main model disappears tomorrow, you will already know your fallback works.
  7. Save the set in your repository and run it on every model, prompt or tool change. This is also the kind of record the AI Act expects you to keep for important systems.

Croatia Corner: ChatGPT Ads, the AI Plan and Zagreb Connect

Marketing

ChatGPT ads can now be bought locally

Netokracija reports that Aleph will sell ChatGPT advertising in Croatia and 43 other markets, including Slovenia, Serbia, Hungary, Bulgaria and Romania. Croatian advertisers can sign locally, receive Croatian invoices and pay in local currency, with support from a local team. ChatGPT ads are labelled and kept separate from answers, and advertisers do not see users' conversations. Context: Croatian online advertising was worth about €190.9M in 2025, and an Ipsos survey found that half of Croatian AI users already use it alongside or instead of Google search. For marketers, being cited in AI answers and paying to appear next to them are now two separate budgets.

Policy

National AI plan: consultation closes on 15 November

The National AI Development Plan to 2032 and its first action plan to 2028 are open for comment on e-Savjetovanja until 15 November. The plan has 33 measures and €932M in total, about €382M of it in the first action plan. Lines worth reading if you run a company: about €115M for Croatian companies in strategic AI projects (€36M of it for startups and scale-ups), €39M for an AI factory and national computing, €2.6M for a Croatian language model, €10M for vouchers and micro-credentials, and only €800,000 for AI vouchers for small and medium businesses. On Netokracija's podcast, the open question was who will get access to the national computing capacity and on what terms. If that matters to you, now is the time to say so.

Events

Zagreb Connect, 26 November

Zagreb Connect at ZICER on 26 November includes an Open Startup Pitch for startups from all over Croatia: three minutes in English in front of investors, business angels and VC funds. Applications are open.

The Week in Numbers

$8 / $40
reported Fable 5.2 price per million input / output tokens
$70–78
Anthropic's IPO price range per share
+3.4%
Nasdaq 100 on 10 November
44
markets where Aleph sells ChatGPT ads, Croatia included

What to Watch

  • 15 November: the consultation on Croatia's National AI Development Plan closes.
  • 17 November: Gems in the Gemini app go read-only; Nvidia's results are expected around this date.
  • 17–20 November: Microsoft Ignite in San Francisco.
  • 18–19 November: expected pricing and first trading day for Anthropic's IPO.
  • 20 November: Conductor's cloud service shuts down after Meta's acquisition.
  • 26 November: Zagreb Connect at ZICER.
  • 2 December: the AI Act deadline for machine-readable marking of content from generative AI systems already on the market.

Main sources: Anthropic, SD Times, Decrypt, Renaissance Capital, Axios, The New Stack, Freevacy on Mistral, Tech Insider, Netokracija on Aleph, Netokracija on the AI plan, ZICER.

Know which model to use before the next one ships

We build evaluation sets from your real tasks, compare models on cost per successful result, and set up a tested fallback so a vendor change never stops your work. Claude Certified Architect, based in Zagreb.

Talk to us

Boris Agatić

Founder of AI Workshop and Claude Certified Architect. Over 10 years of building software, guiding companies from the first strategy call to production AI agents.

LinkedIn (opens in a new tab)