ai-powered-markdown-translatorArticle translated from fr to en with gpt-5.4-mini.
June 20, 2026 marks two complementary moves: Runway is radically simplifying the integration of AI video generation into products via its API Recipes, while OpenAI publishes LifeSciBench, an expert benchmark of 750 PhD-level tasks in the life sciences, and introduces GPT-Rosalind, its new specialized model that outperforms GPT-5.5 across the entire benchmark. On the GitHub side, a weekend promotion offers additional credits to Copilot Max subscribers.
Runway API Recipes — production video generation in one call
June 17 — Runway is launching Recipes on its API: preconfigured endpoints that encapsulate Runway’s internal expertise in prompting and video workflows. A single API call is now enough to embed a complete, production-grade media generation feature into a third-party platform.
The value of this approach is that it removes the burden of pipeline design and maintenance: the developer consumes a polished final result without having to assemble or optimize the intermediate steps. The Recipes package Runway team’s best practices directly.
“New on the Runway API: Recipes. Drop production-ready generative media features into your platform, with one API call. Recipes are Runway-built endpoints with our prompting and workflow expertise packaged in. Polished results, without building or maintaining the workflow.” — @runwayml
LifeSciBench and GPT-Rosalind — PhD-level evaluation in the life sciences
June 17 — OpenAI publishes LifeSciBench, an expert benchmark designed to assess whether an AI model can act as a true collaborator in scientific research — not merely answer biology questions. The benchmark involves 173 PhD contributors and 453 expert reviewers with more than twelve years of experience on average.
| Metric | Value |
|---|---|
| Total tasks | 750 |
| Artifacts (figures, PDFs, sequences) | 1,062 |
| Multi-step tasks (≥2 steps) | 79 % |
| Inter-rater agreement | > 96 % |
At the same time, OpenAI introduces GPT-Rosalind, a new model specialized in the life sciences, whose results on LifeSciBench surpass GPT-5.5 across all measured dimensions.
| Model | Overall pass rate | Scientific communication | Clinical translation |
|---|---|---|---|
| GPT-5.5 | 25.7 % | 56.3 % | 36.8 % |
| GPT-Rosalind | 36.1 % | 71.1 % | 57.7 % |
GPT-Rosalind’s identified weaknesses: tasks involving artifacts (28.1% vs. 45.1% for text-only tasks), exact numerical outputs (14.8%), and the generation of biological constructs.
Briefs
-
Copilot Max — $200 weekend credits — GitHub is offering Copilot Max subscribers an extra USD 200 in credits this weekend for the GitHub Copilot app, with announced offers to come for the Pro and Pro+ plans. 🔗 Source
-
Runway × UNHCR — FIFA World Cup film — upperfast used Runway to create a film for UNHCR about professional footballer refugees (including Alphonso Davies and Antonio Rüdiger). The film reached 1.2 million organic views in 24 hours and was presented at the UN General Assembly before entering permanent residence at UN headquarters in New York. 🔗 Runway customer story
-
OpenAI joins the Rust Foundation (Platinum) — OpenAI becomes a Platinum member of the Rust Foundation, the nonprofit organization that manages the Rust programming language, and makes a donation to the open source project. 🔗 Rust Foundation announcement
What this means
Integrating AI video generation into production is reaching a new level of accessibility. With Recipes, Runway offers a model analogous to preconfigured components in frontend frameworks: a developer consumes a finalized endpoint rather than assembling a pipeline from scratch. This abstraction lowers the entry cost for software publishers that want to embed AI video capabilities without in-house expertise in prompting or workflow orchestration. It is a signal that the API layer for generative media is being structured around higher-level primitives, comparable to what voice or translation APIs did a few years ago.
Evaluating models on expert scientific tasks reveals precise, measurable gaps. LifeSciBench goes beyond factual memorization benchmarks by testing real research workflows — experimental design, clinical translation, multi-step reasoning over artifacts. The fact that GPT-Rosalind exceeds 28% only on artifact-based tasks, while reaching 45% on text alone, points to a concrete limit: scientific multimodal understanding remains the main bottleneck. This type of expert benchmark, co-built with 173 PhD researchers, sets a new standard for measuring the real utility of models in professional settings.