Mistral shows European AI has more to offer: Reflection releases low-cost LLM with open weights
OpenAI gives mathematicians extra work. Analysis shows Claude subscriptions’ bang for buck. Google’s open weight embedding model gets an upgrade. Optional text watermarks come to OpenAI products.
In today’s edition of Data Points, you’ll learn about our top headlines, and more:
- OpenAI gives mathematicians extra work
- Analysis shows Claude subscriptions’ bang for buck
- Google’s open weight embedding model gets an upgrade
- Optional text watermarks come to OpenAI products
But first:
“Le Chonk” offers users worldwide a powerful open option
Mistral released Mistral Large 4 in Research Public Preview, with open weights for the 1 trillion parameter (49 billion active) model planned for late October. It scores 38 on the Artificial Analysis Intelligence Index, comparable to GPT-6 Luna (max, 38) and DeepSeek V4.1 Flash (max, 39), making it the most intelligent model from outside the US and China according to Artificial Analysis. It also scores 50 on the firm’s Cyber Index, level with GLM-5.3-Flash, and supports a 512,000-token context window with text and image input. Standard pricing is $1.36/4.18 per million input/output tokens, putting its Cost per Task at $1.13 — over four times that of similarly-capable open weights models like GLM-5.3-Flash ($0.25) and DeepSeek V4.1 Flash ($0.27), although a 50 percent launch discount cuts this to $0.57 for the first two weeks. The benchmark results suggest Mistral has closed the intelligence gap with top US and Chinese models, but developers will pay a notable premium until pricing or efficiency improves. (Artificial Analysis)
Reflection’s much-anticipated Beam boasts cost efficiencies
Reflection AI, a startup founded in 2024 by former DeepMind researchers, released its first open weight model, called Beam, positioning it as a US-built alternative to open weight models built in China. CEO Misha Laskin says Beam excels at coding and agentic tasks and needs three to four times less compute to reason through a problem than comparable open models, making it faster and cheaper to run. Reflection says Beam’s performance is comparable to Z.ai’s GLM-5.2 and approaches Alibaba’s Qwen 3.8-Max, though both still trail top closed models from OpenAI and Anthropic. The New York-based company, funded by Nvidia and other top investors, is already training a more capable successor model. For developers, Beam adds another open weight option from a Western lab alongside Mistral and Thinking Machines Lab, aimed at users who want to avoid models built in China but still run and control their own weights. (Semafor)
Unknown OpenAI model solved a giant math problem set
OpenAI released a set of new mathematical results produced by an internal, unreleased frontier model, publishing them in a GitHub repository with protocols for citations and paper revisions. The repository includes formalizations of many proofs in Lean, a language that lets computers verify mathematical proofs, plus 10 summaries of the model’s reasoning and statistics on how many problems it attempted. OpenAI estimates the average result took roughly three hours of ChatGPT Pro-equivalent compute to produce. The company developed its disclosure approach with input from the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and says it plans to fund workshops and conferences to help the math community understand AI-generated results. For developers, this signals OpenAI is testing its ability to release unverified AI research output before making the underlying model itself available. For mathematicians, this poses verification headaches comparable to OpenAI’s earlier Navier-Stokes proof, only at greater scale. (OpenAI)
Claude subscription plans are more cost-efficient than OpenAI’s
SemiAnalysis built a dashboard to measure the real API-equivalent value of AI subscription plans and found Anthropic’s Claude plans deliver roughly five times more value per dollar than OpenAI’s for agentic workloads. Subscriptions normally show only a usage meter, not per-token pricing, so SemiAnalysis ran controlled prompts and tracked how far each plan’s meter moved per token type. The firm says OpenAI recently cut the API-equivalent value of its $200 Pro plan in half, introduced a $500 tier offering only 21 percent more usage of its top model, and quietly lowered cached-token pricing for GPT-6.1 Sol in the same update. Anthropic’s tiers, by contrast, offer consistent per-dollar value across all plan levels, and its Opus 5.5 plan still beats OpenAI’s equivalent even adjusted for OpenAI’s cheaper per-token pricing. SemiAnalysis also writes that providers can silently change subscription limits without announcement. This makes continued tracking necessary for anyone modeling AI lab economics. (SemiAnalysis)
Open weight embedding model offers developers an alternative
Google released EmbeddingGemma 2, an open embedding model that unifies text, code, images, video, and audio in a single shared embedding space. Built on the Gemma 4 architecture and licensed under Apache 2.0, it has 740 million parameters total but can run with as little as 270 million for text-only workloads, plus optional 170 million-parameter vision and 300 million-parameter audio encoders. It supports an 8K token context window (four times the original EmbeddingGemma), can truncate output vectors from 768 down to 128 dimensions via Matryoshka Representation Learning for up to 6x storage savings. The model needs roughly 191MB of RAM for text or 567MB for full multimodal use, making it capable of running on a Pixel 11 Pro or comparable smartphone. Google reports a 9.92-point gain on MTEB Code over the original EmbeddingGemma and says the model leads other sub-1B multimodal embedders on benchmarks including MTEB Code and MAEB, though it hasn’t published comparisons against larger proprietary models beyond those claims. The model is available now on Hugging Face and Kaggle with support from tools like Ollama, llama.cpp, vLLM, and Qdrant, giving developers a free, on-device option for building offline semantic search and retrieval-augmented generation across multiple data types. EmbeddingGemma2 is behind Google’s and Cohere’s paid models in accuracy, but open embedding models provide developers continual access without having to pay again to embed new data. (Google)
ChatGPT will get text watermarks, but only optional outside EU
OpenAI introduced a text watermarking system called textGrain to comply with the EU AI Act’s requirement that generated text be machine-identifiable. Already, API customers worldwide can opt in to watermarking for select models, with it off by default; ChatGPT and Codex in the EU will get invisible watermarks on eligible outputs in the coming weeks, but this is not becoming a global default. The watermark embeds a statistical signal in word choice, and OpenAI says it matched or beat Google’s SynthID for text in testing, though detection drops sharply with edits: replacing 10 percent of words with synonyms cut detection from 92 percent to 66 percent, and 25 percent synonym replacement dropped it to 17 percent. Access to the detector is limited at launch to approved researchers and expert organizations, not the public, because of risks of false positives and negatives. For developers building in the EU, this means watermarking is available but optional outside ChatGPT/Codex, and the technology is explicit that a detected watermark proves nothing about authorship, ownership, or accuracy. It’s unclear whether this approach will meet the AI Act’s requirements, but opt-in use helps assuage worries that even invisible statistical watermarking may affect output quality. (OpenAI)
Want to know more about what matters in AI right now?
Read the latest issue of The Batch for in-depth analysis of news and research.
Last week, Andrew talked about the promising cyber capabilities of the open weight model GLM-5.3, its potential for cyberdefense, and the importance of engineering solutions to improve AI safety.
“Similarly, even though mass media describes the OpenAI-Hugging Face hack as dangerous agents ‘going rogue,’ this incident is leading teams everywhere to engineer better sandboxing and monitoring – important changes that are making AI safer. (On OpenWorker, our agent harness that supports security workflows, Rohit Prsad, Devika Verma, and I are building on Nvidia's new safety framework to improve sandboxing.)”
Read Andrew’s letter here.
Other top AI news and research covered in depth:
Other top AI news and research covered in depth:
- Xiaomi emerges as an unexpected leader in open weights with its MiMo-V2.6-Pro RL and MiMo-V2.6-Flash models.
- Google’s Gemini 3.8 goes live, integrating voice, video, and reasoning in one with extended thinking capabilities.
- DeepSeek’s V4.1 architecture leapfrogs its Pro version again by significantly reducing cache reads.
- AREX introduces agents that can check their own work piece by piece, condensing and reviewing their own research.
A special event for our community

AI Dev brings together developers who build with AI every day. You'll hear from engineers at the companies shipping agents, models, and infrastructure, then meet them in person on the demo floor. Join us in New York City on November 30 and December 1.
Data Points is produced by human editors with AI assistance.