A week about prices and paperwork: the maker of Claude opened its books to the public, OpenAI cut the price of a capable model in half, and Europe's biggest lab gave away the weights of a trillion-parameter model. For buyers, the choice of model got cheaper and the question of vendor risk got clearer.
Last week was about where agents live: Claude Duet, OpenAI and Claude on Google Cloud, and Daytona's unicorn round. This week the money came into view. Anthropic publicly filed its S-1 for a Nasdaq listing under the ticker ANT, the first detailed look at a frontier lab's finances. OpenAI launched GPT-6.1 Terra, a mid-tier model at half the price of Sol, and Mistral published the weights of Mistral Large 4, as promised for the end of October. Anthropic also reported early drug-development results with six partners. At home, a European Commission delegation visited Infobip in Vodnjan, AI2Future met in Zagreb, and the consultation on the national AI plan is still open.
The Releases at a Glance
OpenAI's mid-tier model, in ChatGPT Work, Codex and the API as gpt-6.1-terra. $1 input / $5 output per million tokens, half the price of GPT-6.1 Sol, with a context window of about 1M tokens. OpenAI reports 50.3% on Terminal-Bench 4.0 at maximum effort.
A natively multimodal mixture-of-experts model with about 1 trillion total and roughly 50B active parameters. Weights on Hugging Face in FP8 and NVFP4; runs on four B200 or B300 GPUs. Reported licence: a modified MIT licence, with a separate commercial licence for model-as-a-service providers.
Public registration statement for a Nasdaq Global Select listing of Class A shares as ANT. Share count and price range not yet set; the registration is not yet effective.
An update on Claude's work with six drug developers, including Generate Biomedicines, Divergence Labs, BigHat Biosciences and Enveda. Results are early and mostly preclinical; Anthropic plans to extend the work into clinical trials.
Anthropic Opens Its Books
Anthropic confidentially filed with the SEC in June; on 26 October the S-1 became public. The numbers, as reported from the filing: about $16.3 billion in first-half 2026 revenue, up from about $1.1 billion a year earlier, with roughly $11.6 billion of it in the second quarter. As of 30 September, more than 11,900 customers paid Anthropic $100,000 or more a year (up from 950), more than 1,900 paid $1 million or more, and about 250 paid $10 million or more. About 48% of first-half revenue came through cloud marketplaces — AWS, Google Cloud and Microsoft — and the top five customers made up about a quarter of revenue in the first nine months, with no single customer above 10%.
The costs are just as large. The filing lists about $265.6 billion in fixed or minimum compute commitments as of 30 June, with terms of 1 to 20 years, and reports a first-half net loss that coverage puts at about $80.7 billion, largely non-cash. Valuation estimates in the press range from $1.5 to $2 trillion; none of that is set until the price range is published.
What it means for buyers: a public filing turns vendor risk from rumour into documents. Three things worth noting for anyone building on Claude: the company is now required to disclose material risks and changes; half of its sales already go through the big clouds, so buying Claude through your existing cloud contract is the normal route, not a workaround; and the compute commitments explain why capacity, rate limits and priority tiers keep changing. None of this changes the model you use today, but it is the right moment to write down your exit plan: which other model you would switch to, and how long it would take.
Claude in drug development: early results with six partners
On 27 October Anthropic published an update on its drug-development work. For Divergence Labs, Claude produced about 35,000 antibody designs in roughly a month for food-allergy research; other partners used it to optimise antibody leads, improve part of a CAR-T therapy for solid tumours and review regulatory documents. These are company-reported, early results: the designs are hits to test, not medicines. For businesses outside pharma, the useful part is the pattern — a model that generates many candidates, and a lab or expert process that filters them.
Cheaper Models, More Choice
GPT-6.1 Terra: half the price of Sol
OpenAI now sells four model classes, from cheapest to most capable: Luna, Terra, Sol and Astra. Terra, released on 26 October, costs $1 / $5 per million input / output tokens, with cached input reported at $0.05. OpenAI says it is much faster for agent work and reports 50.3% on Terminal-Bench 4.0 at maximum effort; independent coverage says it still trails Sol and Astra on the hardest coding tasks. It is available in ChatGPT Work and Codex for paid plans, not in the regular ChatGPT chat.
Mistral Large 4 weights are public
Three weeks after the hosted preview, Mistral published the weights of Mistral Large 4 on 29 October, with further reinforcement-learning training since the preview. Mistral calls it the best open-weight model for coding and is positioning it for security work; there is no independent confirmation yet. The licence is not Apache 2.0 like Large 3: reports describe a modified MIT licence that requires a separate commercial agreement if you sell the model as a service. For companies that must keep data on their own hardware, a European frontier-class model that runs on one four-GPU server is a real option for the first time — but read the licence before you build on it.
Elsewhere
- Big Tech spending: Microsoft (28 October) and Amazon (29 October) reported quarterly results after Meta (21 October). Combined 2026 AI capital spending guidance for the four largest US platforms is around $700 billion; check each company's own release for exact figures.
- Google is rolling out skills in the Gemini app, expected to finish by mid-November; the rollout in Workspace is slightly delayed. The Gemini agent announced on 8 October is not yet widely available.
- OpenAI said an internal model produced three new scientific results; none has been independently verified yet.
- EU AI Act: feedback on the draft implementing act on AI Office procedures closes on 9 November, and the 2 December deadline for machine-readable marking of AI-generated content in systems already on the market is five weeks away.
Tutorial of the Week: Route by Task, Not by Habit
Terra at $1 / $5 and Large 4 on your own hardware make this a good week to stop sending every request to your most expensive model. A simple routing test you can run in a day:
- Pull 100 real requests from your logs, grouped by task: classification, extraction, summarisation, drafting, coding, multi-step agent work.
- Write the pass criterion for each group before you test — "the extracted invoice total matches", "the summary contains these three facts" — not "it looks good".
- Run three tiers on the same requests: a small model (Haiku 5.5, Luna), a mid-tier model (Sonnet 5.5, Terra) and your current top model.
- Record cost and latency per correct answer, not per request. A cheap model that fails 20% of the time and needs a retry is often not cheap.
- Route each group to the cheapest tier that meets the criterion, and keep a fallback to the next tier when confidence is low or validation fails.
- Re-run monthly — prices and models change every few weeks, as this digest shows.
Croatia Corner: Brussels in Vodnjan, AI2Future in Zagreb
A European Commission delegation visits Infobip
A delegation from the European Commission visited Infobip's headquarters in Vodnjan, as CroPC reported on 27 October, to discuss AI's effect on business and Europe's technological competitiveness. Infobip is building a next-generation AI-based communications platform within IPCEI-CIS, the European project for next-generation cloud and edge infrastructure, and has made AI use a goal for every employee this year. For Croatian companies, IPCEI and the national AI plan are the two routes where public money for AI infrastructure is most likely to appear.
AI2Future at the Kraš Auditorium
The AI2Future conference, supported by CroAI under the theme "AI Beyond the Hype", met at the Kraš Auditorium in Zagreb on 28–29 October, with an AI demo fair and startup pitches. It is the largest yearly meeting point for Croatian AI teams; if you missed it, the CroAI community meetups run through the year.
National AI plan: the consultation is still open
The 30-day public consultation on the National Plan for AI Development (€932.1M to 2032, including a large language model for Croatian and an assistant in e-Građani) runs until mid-November. If your company works with Croatian-language data, computing or AI training, this is the moment to comment — the measures are not final, and the first action plan to 2028 is where calls for funding will come from.
One more connection from this week: Daytona, the Croatian-founded unicorn from last week's digest, sells sandboxes to the same labs whose finances are now public. Anthropic's S-1 lists hundreds of billions in compute commitments; the companies that build the tools around that compute, from Split and Zagreb too, are part of the same market.
The Week in Numbers
What to Watch
- Anthropic's price range: the next S-1 amendment should set the share count and price; the listing date follows.
- 9 November: feedback on the AI Act implementing act on AI Office procedures closes.
- Mid-November: the consultation on Croatia's national AI plan closes; Gemini app skills finish rolling out.
- "Coming weeks": general availability of OpenAI models on Google Cloud.
- 2 December: AI Act machine-readable marking deadline for generative AI systems already on the market.
- Your PHP version: if last week's PHP2Shell patches are not yet on your sites, that is still the most urgent item on this list.
Pay for the model each task needs
We help Croatian companies test models on their own data, route each task to the cheapest model that passes, and keep an exit plan for every vendor — with AI Act documentation ready for audit. Claude Certified Architect, based in Zagreb.
Talk to us