先週のAIニュース #335 - Opus 4.6、Codex 5.3、Gemini 3 Deep Think、GLM 5、Seedance 2.0

Last Week in AI #335 - Opus 4.6, Codex 5.3, Gemini 3 Deep Think, GLM 5, Seedance 2.0

A crazy packed edition of Last Week in AI! Plus some small updates.

Last Week in AIFeb 16, 20264135ShareEditor’s note: I apologize for the inconsistent release date of the newsletter and podcasts in recent months. I’ll aim to start releasing on Saturday/Sunday consistently from now on! This edition of the newsletter covers a bit more than a week as a result.

I am also going to be adding an ‘Editor’s Take’ for Top News to add a bit commentary and extra context beyond the news summary.

Anthropic releases Opus 4.6 with new ‘agent teams’

Anthropic launches Claude Opus 4.6 as AI moves toward a ‘vibe working’ era

Summary: Anthropic released Claude Opus 4.6, a major upgrade focused on larger, faster, and more collaborative work. The headline feature is “agent teams,” a research-preview capability that lets multiple coordinated agents split a complex task into parallel subtasks, communicate directly, and finish faster than a single sequential agent. Opus 4.6 also expands the context window to 1 million tokens (on par with Sonnet 4/4.5), enabling work over large codebases and long documents. Another key addition is a native PowerPoint side panel, letting users draft and edit slides directly inside PowerPoint with Claude instead of generating decks externally and importing them.

Anthropic says Opus has evolved from a coding-first model to one suited for a broader set of knowledge workers, citing adoption beyond software engineers to product managers and financial analysts. Advancements in Claude Code and the productivity suite Claude Cowork are feeding a shift from “vibe coding” to “vibe working,” reflecting end-to-end task execution and higher-quality professional outputs.

Editor’s Take: Opus 4.6 comes only a few months after 4.5, and yet showcases impressive gains across a variety of benchmarks beyond what the 0.1 version bump might suggest. It feels like the frontier labs may have gotten to the point of continously post-training their models via RL, and I wouldn’t be surprised if we see more impressive gains in just a few months.

With GPT-5.3-Codex, OpenAI pitches Codex for more than just writing code

OpenAI says new coding model helped build itself

OpenAI’s new Spark model codes 15x faster than GPT-5.3-Codex - but there’s a catch

A new version of OpenAI’s Codex is powered by a new dedicated chip

OpenAI launches new macOS app for agentic coding

Summary: OpenAI unveiled GPT-5.3-Codex, a frontier coding model available via CLI, IDE extension, web, and a new macOS Codex app (API coming later). The model outperforms GPT-5.2-Codex and GPT-5.2 on SWE-Bench Pro and Terminal-Bench 2.0, and runs 25% faster with lower compute usage. While headlines claim it “built itself,” OpenAI clarifies the model was “instrumental in creating itself,” with early versions used to debug training, manage deployments, analyze test results, clean datasets, and monitor training runs—not fully self-building. The model is pitched for the full software lifecycle beyond code generation: debugging, deploying, monitoring, writing PRDs, editing copy, user research, tests, and metrics. OpenAI also designated GPT-5.3-Codex as its first “high-capability” model for cybersecurity per its Preparedness Framework.

A companion release, GPT-5.3-Codex-Spark, targets real-time “conversational” coding with a smaller, latency-first research preview that generates code up to 15x faster than GPT-5.3-Codex but trails it on SWE-Bench Pro and Terminal-Bench 2.0, and is not expected to meet the “high-capability” cybersecurity threshold. Spark achieves big latency gains—roughly 80% lower round-trip overhead, 50% faster time-to-first-token, and 30% lower per-token overhead—via session and streaming optimizations and persistent WebSocket connections, and it defaults to lightweight, targeted edits without auto-running tests. Initially limited to $200/month Pro users, Spark is the first milestone in OpenAI’s multi-year, >$10B partnership with Cerebras, running inference on the WSE-3 waferscale chip (4 trillion transistors).

Complementing the models, OpenAI’s new macOS Codex app orchestrates multiple agentic coding assistants in parallel, supports scheduled automations and selectable agent “personalities,” and has seen 500,000+ downloads, with OpenAI aiming toward dual modes: long-horizon reasoning/execution alongside real-time collaboration, and background sub-agents for longer tasks.

Editor’s Take: OpenAI seems to have succesfully caught up with Anthropic in terms of having a model that is optimized for improving workplace productivity, after having previously focused more on improving ChatGPT as a general purpose chatbot for anybody and everybody. Pretty big deal!

Google Unveils Gemini 3 Deep Think for Science & Engineering

Is This AGI? Google’s Gemini 3 Deep Think Shatters Humanity’s Last Exam And Hits 84.6% On ARC-AGI-2 Performance Today

Summary: Google unveiled a major upgrade to Gemini 3 Deep Think, a specialized “extended reasoning” mode tuned for science, research, and engineering where accuracy and multi-step logic matter more than latency. The system emphasizes test-time compute—letting the model think longer—and internal verification to prune incorrect reasoning paths, a direct response to competitors like OpenAI’s o1 and Anthropic’s Claude. Google reports a verified 84.6% on ARC-AGI-2 from the ARC Prize Foundation, a benchmark designed for generalization to novel tasks rather than memorization, and 48.4% on Humanity’s Last Exam without tools across logic-dense academic questions. Deep Think also claims gold medal–level performance on the written sections of the 2025 International Physics and Chemistry Olympiads and gold-level on the 2025 IMO, plus 50.5% on the CMT-Benchmark for advanced theoretical physics.

For engineering and coding, Gemini 3 Deep Think now posts a 3455 Elo on Codeforces—“Legendary Grandmaster” territory—demonstrating capability in complex data structures, algorithmic optimization, and system design, with Google noting a 35% accuracy lift in software engineering tasks over prior versions.

Editor’s Take: rounding off the impressive model releases, this one may just be the most impressive of all. Even having gotten used to major advancements, the ARC-AGI-2 result surprised me and many other commentators. Similar to Opus 4.6, this makes me wonder if scaling up RL / post-training is behind such rapid and signifiacnt improvements.

DeepSeek boosts AI model with 10-fold token addition as Zhipu AI unveils GLM-5

Zhipu leads rally in Chinese AI stocks, surging 30% as a wave of new releases hits market

MiniMax M2.5 promises “intelligence too cheap to meter” as Chinese labs squeeze Western AI pricing

China’s Zhipu AI launches new major model GLM-5 in challenge to its rivals

Z AI launched GLM-5, new open-source model on chat and APIs

Source

Summary: China’s AI labs rolled out major upgrades focused on long context, coding, and agentic workflows. DeepSeek expanded its flagship model’s context window from 128,000 to over 1,000,000 tokens, enabling much larger “memory” for complex reasoning, multi-file code work, and sustained tasks. Zhipu AI (Z.ai) launched GLM-5, an open-weight flagship positioned for “agentic engineering” rather than “vibe coding,” emphasizing end-to-end software delivery, deep debugging, and multi-step tool use. GLM-5 scales to a Mixture-of-Experts architecture with roughly 744B total parameters and 40B active per token, increases pretraining to 28.5T tokens, and integrates DeepSeek Sparse Attention to balance long-context performance with serving efficiency.

Benchmarks and pricing underscore intensifying competition. Zhipu reports GLM-5 leads among open models on coding and agentic tasks, with self-reported scores like 77.8 on SWE-bench Verified and strong showings on BrowseComp and MCP-Atlas; it ranks #1 among open models in Text Arena and aims to approach Claude Opus 4.5, though it still lags Anthropic’s Claude on coding by Zhipu’s own admission. MiniMax released M2.5 under MIT as open weights, trained with reinforcement learning across hundreds of thousands of complex environments to optimize autonomous planning and long-horizon efficiency.

Editor’s Take: Starting with DeepSeek V3/R1, it has felt like the open source models coming out of China are getting increasingly close to matching the performance of the flagship models of western frontier labs. These releases continue that trend, which makes me wonder if open source options will soon start eating away at the profits of closed source API providers.

Bytedance shows impressive progress in AI video with Seedance 2.0

ByteDance’s new model sparks stock rally as China’s AI video battle escalates

ByteDance’s new Seedance 2.0 supposedly ‘surpasses Sora 2′

ByteDance’s Seedance 2.0 Might be the Best AI Video Generator Yet

ByteDance’s new model sparks stock rally as China’s AI video battle escalates

Hollywood isn’t happy about the new Seedance 2.0 video generator

Summary: ByteDance pre-released Seedance 2.0 to select users, showcasing a multimodal video generator that accepts up to 12 inputs at once (max: nine images, three videos, three audio files) alongside text, and outputs 4–15s clips with auto-generated sound effects/music. The headline upgrade is precise reference capability: users can lift camera motion, editing grammar, effects, and action beats from uploaded reference videos, swap/insert characters, extend existing shots, and chain multi-scene storylines while maintaining style and character consistency. Early testers report sharper 2K exports, roughly 30% faster generation vs Seedance 1.5, smoother camera moves, improved visual consistency, and convincing motion accuracy and lip-sync, with results that are watermark-free.

The launch intensified China’s AI video race days after Kuaishou’s Kling 3.0 and sparked a stock rally: COL Group hit its 20% limit, Shanghai Film and Perfect World rose ~10%, and Huace Media ~7%, while the CSI 300 ticked up 1.63%. Swiss consultancy CTOL and other reviewers claim Seedance 2.0 surpasses OpenAI’s Sora 2 and Google’s Veo 3.1 in practical tests, citing lifelike characters, fine-grained edit control, and “cinematic” natural-language workflows—though cherry-picked demos likely show best-case output.

Meanwhile, Hollywood and rights groups blasted the tool’s lack of guardrails after viral clips mimicked celebrities (e.g., Tom Cruise vs. Brad Pitt) and Disney/Paramount IP; Disney and Paramount sent cease-and-desist letters alleging indistinguishable reproductions of franchises like Spider-Man, Darth Vader, and Grogu.

Editor’s Take: this feels like a Nano-Banana moment for video generation — (I created that em-dash, not AI) a huge and sudden leap far beyond what existing text-to-video models can do. The ability to edit/compose inputs so well vastly increases the usefulness of this model, and I wouldn’t be surprised if we start seeing a lot more primarily AI-generated videos within a year.