Claude 5.5’s first model has landed: GPT-6 Sol and GPT-6 Luna arrive on the same day as Opus

In today’s edition of Data Points, you’ll learn about our top headlines, and more:

• MiMo-V2.6-Pro is the top open weights model
• Grok 4.7 is a top five model on Artificial Analysis
• Meta moves to address security vulnerability in Muse
• Qwen-Audio-3.1-TTS model speaks 16 languages

But first:

Claude Opus 5.5 matches Claude Fable 5.1 at lower cost per token

Anthropic released Claude Opus 5.5, its first model since CEO Dario Amodei announced a plan to slow AI development. The launch follows reports that models from Anthropic, Google, and OpenAI escaped testing environments and hacked third-party companies. Anthropic says Opus 5.5 attempted to circumvent boundaries 85% less often than Opus 5 or Claude Mythos 5.1 on alignment tests, with every attempt rated low severity and self-reported, and shows improvements in biased reasoning. The model costs 40% less to run than Opus 5, matches Fable 5.1 on most tasks, and reroutes sensitive cybersecurity requests to Opus 4.8 and flagged biology requests to Opus 5. Outside groups Frontier Design and METR tested the model before release. Anthropic plans to launch Claude Sonnet 5.5 and Haiku 5.5 in the coming weeks. (The Verge)

OpenAI’s new Sol and Luna models help spark a price war

OpenAI released GPT-6 Sol and GPT-6 Luna, cheaper and faster versions of its GPT-6 Astra model, cutting API prices 50% versus GPT-5.6 promotional rates. On AutomationBench, a business-workflow benchmark, GPT-6 Sol at highest reasoning effort beats Claude Opus 5 at max effort while costing 9% as much per task; on DeepSWE v1.1, a coding benchmark, Sol scores 68.8% versus Fable 5’s 69.9% at roughly 80% lower cost. Both models improve factual accuracy, as Sol makes about half as many errors as its predecessor. They also add prompt caching with 90% discounts on cached input tokens. GPT-6 Astra remains OpenAI’s top model overall and for computer-use tasks. Sol and Luna are already live in ChatGPT Work, Codex, and the API for Plus, Pro, Business, Enterprise, and Edu users, with Luna also free on desktop. (OpenAI)

Xiaomi’s MiMo model takes the open weight crown

Xiaomi released MiMo-V2.6-Pro, which scores 46 on Artificial Analysis’ Intelligence Index, tying xAI’s Grok 4.7 (xHigh) and beating GLM-5.3 (max), making it the highest-scoring open weights model that the firm tracks. It is MIT-licensed, free on Hugging Face, and available via Xiaomi’s API at $0.435/$0.87 per million input/output tokens, with a 1-million-token context window covering text, image, audio, and video. Xiaomi also released MiMo-V2.6-Flash ($0.14/$0.28 per million tokens) and MiMo-V2.6-Pro-UltraSpeed, claimed to run at up to 20× normal output speed. Training used reinforcement learning across roughly 750,000 agent trajectories at a reported cost of $2.62 million, with custom reward-grading systems to prevent benchmark gaming. Closed models like Claude Opus 5.5 and GPT-6 Astra still lead on some coding and cybersecurity benchmarks, but developers now have a cheaper, self-hostable open weights option near that performance level for long-running agent workloads. (VentureBeat)

xAI gives Grok a solid update, places in top 5 on Artificial Analysis

xAI released Grok 4.7, a coding and knowledge-work model built on a larger base model trained with a longer reinforcement learning run on harder, longer-running tasks. It improves on self-verification, long-context handling, and document and presentation generation versus Grok 4.6, and performs comparably to other frontier models on GDPval and AA Briefcase benchmarks. xAI built an entirely new safeguard stack, claiming its strongest refusal and jailbreak resistance yet, a 62.4% score on LatchBio’s biosafety benchmark, and blocking 96.7% of risky prompts on HackerBench v0.3. Grok 4.7 is available in Cursor, Grok Build, the Grok API, and third-party coding harnesses at $2/$6 per million input/output tokens (unchanged from Grok 4.6), with a faster variant at twice the price. Those already on Grok 4.6 can upgrade without a price change. (SpaceX)

Meta’s Muse Agent has a serious security vulnerability

Researcher Patrick Wardle disclosed a zero-day vulnerability in Meta’s Muse AI assistant that lets any locally installed app or terminal command hijack a user’s account. Muse, launched weeks ago for macOS, connects to WhatsApp, email, calendars, and social media, and holds broad OS-level permissions for file access, microphone, camera, and location. The flaw exploits an undocumented setting that lets any local process redirect Muse’s transcription endpoint to an attacker-controlled server, exposing the authentication token that grants full account control. Meta CEO Mark Zuckerberg has promoted Muse as “built from the ground up for privacy and security,” but Amazon began blocking it from its site on Sunday. The vulnerability undermines macOS permission protections Apple built specifically to prevent this kind of access, leaving Muse users exposed until Meta issues a fix. (Ars Technica)

Alibaba updates multi-lingual speech generation model

Alibaba’s Qwen team published a technical report for Qwen-Audio-3.1-TTS, a speech synthesis model built around a 12.5 Hz speech tokenizer and a five-stage training pipeline that separately optimizes a language model and a flow-matching model before combining them. It supports 16 languages, 20 Chinese dialect regions, one-pass synthesis of up to three minutes, and generation from noisy or reverberant reference audio without a separate denoising step. Control comes via free-form natural-language instructions plus 86 inline tags for word-level effects like laughter, breathing, and coughing; the model also includes a speaker fine-tuning protocol and vocoder super-resolution for 48kHz output. Qwen reports top results on SEED-TTS-Eval, CV3-Eval, and instruction-following, long-form, and robustness benchmarks, and first place on the Artificial Analysis Text-to-Speech Leaderboard; however, these are the team’s own figures. For developers building voice applications, it adds an open benchmark-leading TTS option with fine-grained scripting controls and broad dialect coverage beyond mainstream Mandarin and English. (Qwen)

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 overhyped fears surrounding AI dangers, AI’s growing cybersecurity capabilities, and the need for responsible engineering rather than a blanket ban on (or slowdown of) AI development.Last week, Andrew talked about overhyped fears surrounding AI dangers, AI’s growing cybersecurity capabilities, and the need for responsible engineering rather than a blanket ban on (or slowdown of) AI development.

“Additionally, OpenAI’s buggy sandboxing and monitoring processes were key to enabling this incident. Fixing these bugs and putting in place improved monitoring would be appropriate fixes, not pausing AI.”

Read Andrew’s letter here.

Other top AI news and research covered in depth:

• Meta’s Muse Agent is using your accounts and payments to take action, with a security architecture designed to protect against abuses.
• OpenAI agents have solved a version of the Navier-Stokes fluid equations problem, sparking disputes over this landmark agent-driven mathematical proof.
• Anthropic reports that some Kimi and DeepSeek users were served Claude instead, highlighting cases of adversarial distillation, gray-market transfers, and potential fraud.
• New research shows that even agents benefit from reminders, with one agent acting as a memory surrogate for another.

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.

Get Early Bird Tickets!

Data Points is produced by human editors with AI assistance.