Ringier, the Swiss media group, and OpenAI have said they are now working together. The announcement, published under the headline "Ringier and OpenAI launch collaboration", is brief: the OpenAI Ringier collaboration exists, and both sides have launched it. What the available material does not set out is the scope of the work, which OpenAI products are involved, which of Ringier’s markets are covered, how long the arrangement runs, or whether money changes hands. Everything beyond the existence of the partnership is inference rather than reported detail, and this article marks the difference throughout.
Key takeaways
- Ringier and OpenAI have confirmed a collaboration. The announcement as published does not specify scope, financial terms, duration or the OpenAI products involved.
- For developers, the significance is the pattern rather than the deal: access between news organisations and model providers is increasingly settled by contract, not left to crawler policy alone.
- Cost context from Convly’s own database: GPT-5.6 Sol lists at $5.00 in and $30.00 out per million tokens with a 1.05M-token context window, so archive-scale text work is priced per token.
- Video is priced by the second: Sora 2 from $0.05 and Sora 2 Pro from $0.15, against Veo 3.1 at $0.05, Wan 2.5 at $0.05 and Kling 2.5 Turbo Pro at $0.07.
- Open-weight routes such as DeepSeek V4-Pro at $0.435 in and $0.87 out, or Llama 3.3 70B at $0.10 and $0.32, remain far cheaper for high-volume, low-stakes pipeline stages.
- No contract value has been published, so treat any specific figure circulating about this partnership as unverified.
What the OpenAI Ringier collaboration confirms, and what it leaves open
The confirmed facts are narrow. Ringier and OpenAI have launched a collaboration, announced by Ringier. No contract value, product name, market list, timeline or executive quote appears in the material available, so none appears here either.
Stating that narrowness matters, because arrangements of this shape usually acquire detail later. The categories that tend to follow are content licensing, editorial and production tooling, distribution inside an assistant, or some combination of the three. Which of those applies to Ringier is not established. What a collaboration does imply, at minimum, is that a large European publisher will be working with a frontier model provider’s technology somewhere inside its media operations. The interesting part for readers of this publication is not the ceremony of the announcement but the operational arithmetic that any such arrangement runs into.
Why a publisher deal with OpenAI matters to AI model users
Taken alone, one partnership between a model provider and a media group is a small item. Taken as part of a pattern, it is the mechanism by which the news industry’s relationship with AI systems is being settled: contract by contract. Three consequences follow for people who build with models, offered here as context rather than as reported fact.
First, provenance. Providers that sign publishers gain material they can describe as licensed, which changes the risk profile of any product built on top of those models. Second, attribution. When a publisher is inside the arrangement, its journalism is more likely to be surfaced with a link than paraphrased anonymously, and that matters to any developer whose application depends on citations holding up under scrutiny. Third, tooling. Media groups run large, repetitive language workloads: transcription, translation, tagging, summarisation, headline variants, archive search. Those are precisely the workloads that expose whether frontier pricing survives contact with volume.
What frontier-model text pricing looks like at newsroom scale
Whatever shape the work takes, someone pays per token. The figures below are list API rates from Convly’s AI models database. They describe the market a publisher buys into, not the terms of this specific partnership, which remain unpublished.
| Model | Provider | Context window | Input / output per 1M tokens |
|---|---|---|---|
| GPT-5.6 Sol | OpenAI | 1.05M | $5.00 / $30.00 |
| GPT-5.5 | OpenAI | 1.05M | $5.00 / $30.00 |
| Gemini 3.1 Pro | 1.05M | $2.00 / $12.00 | |
| Claude Sonnet 5 | Anthropic | 1M | $2.00 / $10.00 |
| DeepSeek V4-Pro | DeepSeek | 1M | $0.435 / $0.87 |
| Llama 3.3 70B | Meta | 128K | $0.10 / $0.32 |
The spread is the point. A single million-token pass over archive text costs $5.00 in input on GPT-5.6 Sol and about $0.44 on DeepSeek V4-Pro, before any output is generated. Multiply that by a catalogue of several million articles and the choice of model stops being an editorial preference and becomes a budget line. Our AI API cost calculator and the AI price-performance index exist for exactly that arithmetic. OpenAI publishes its own current rates on its API pricing page, which is the document to check before a budget is committed, since list prices move and secondary write-ups go stale.
Per-second video pricing now sits in the publishing stack
Text is only half of a modern publishing pipeline. Generated and edited video is billed by the second, and the numbers are small enough per clip to be misleading at scale. On Convly’s figures, Sora 2 starts at $0.05 per second and Sora 2 Pro at $0.15 per second, which puts a 30-second cut at roughly $1.50 and $4.50 respectively. Rival rates sit in the same band: Veo 3.1 from $0.05 per second, Wan 2.5 from $0.05 and Kling 2.5 Turbo Pro from $0.07.
For a media group producing hundreds of short clips a day across several markets, the difference between $0.05 and $0.15 per second compounds into a meaningful annual figure. Nothing in the announcement indicates that video tooling forms part of this collaboration; the pricing is included because it is the cost environment any publisher-provider arrangement operates in during 2026.
Self-hosting versus API access for a media group
Large publishers with legal duties around data residency often ask whether some workloads should run on their own hardware rather than a vendor endpoint. The open-weight numbers make that a real question rather than a theoretical one. DeepSeek V4-Pro needs roughly 800 GB of VRAM at 4-bit, which is a serious cluster. Mid-sized options are far more approachable: Llama 3.3 70B at about 40 GB, Qwen3 32B at about 20 GB, and Gemma 3 27B at about 16 GB, all with 128K context windows.
The practical answer is usually a split. Sensitive or extremely high-volume steps such as bulk tagging and first-pass translation run locally or on a cheap open-weight endpoint, while the smaller share of work that demands frontier reasoning goes to a hosted model. Sizing that split is what our free VRAM calculator and self-hosting vs API calculator are for, and the wider economics are set out in our open vs closed AI cost study.
Open questions after the announcement
Four things would change how this partnership should be read, and none of them is currently on the record. Whether Ringier content becomes retrievable inside OpenAI’s consumer products, and with what attribution. Whether the arrangement covers editorial tooling for journalists, commercial systems, or both. Which of Ringier’s markets are in scope. And whether any payment flows in either direction. OpenAI posts partnership and product announcements in its own newsroom, so that is where an expanded description of the terms would surface if one is published. Until then, the honest summary is the one at the top of this article.
Frequently asked questions
What did Ringier and OpenAI actually announce? That they have launched a collaboration. The announcement confirms the partnership exists and does not, in the material available, describe its scope, products, markets, duration or financial terms.
Does the OpenAI Ringier collaboration mean Ringier journalism will appear in ChatGPT? That is not stated. Content distribution inside an assistant is one common form these arrangements take, but nothing published so far confirms it here, so treat claims to that effect as unverified.
Has a contract value been disclosed? No. No figure, payment direction or term length appears in the announcement. Any specific number attached to this deal elsewhere should be treated as unsourced until a party publishes it.
What does it cost to run this kind of work on OpenAI’s models? On Convly’s published figures, GPT-5.6 Sol lists at $5.00 in and $30.00 out per million tokens with a 1.05M-token context window. A publisher processing millions of archived items should model that at volume rather than per request, which is what our AI API cost calculator does.
Is there a cheaper route for high-volume pipeline stages? Yes, for work that does not need frontier reasoning. DeepSeek V4-Pro lists at $0.435 in and $0.87 out per million tokens, and Llama 3.3 70B at $0.10 and $0.32, roughly an order of magnitude below flagship rates.
The bottom line
The OpenAI Ringier collaboration is, as announced, a single confirmed fact wrapped in unpublished detail. It is worth noting anyway, because each of these agreements moves the news industry’s dealings with model providers further into contract territory and further away from being decided by crawlers and courts alone. Developers building news-adjacent products should watch the attribution and licensing terms that eventually emerge from arrangements like this one, since those terms shape what content can legitimately be surfaced and cited. And whichever way the specifics land, the arithmetic underneath does not change: per-token and per-second rates, checked against the vendor’s own pricing page, decide what any of this costs to run.
Sources: news.google.com. Reported August 31, 2026.
