# AI lab research refresh — November 14, 2026

This pass looked for new and materially expanded work since the [October 19 coverage audit](/research/ai-labs-october-19-2026/coverage-and-sources.md). It includes 23 assessments across 20 lab series. The October 23 [AuthorizationBench assessment](/blog/authorizationbench-success-does-not-grant-authority) was included when checking what the site had already covered. Announcement dates below are dates shown by sources, not dates on which a tenant was tested.

## Selection

We searched official research and product indexes, release notes, papers, model cards, and repositories. We checked existing blog files for release names, source URLs, and themes. Some posts extend earlier coverage when something material changed: Mistral's weights and Studio preview after the initial Large 4 preview, Grok Bot 1.0 after Team Bots beta, and OpenAI Decisions general availability after its beta. We kept new releases and prior comparison points separate.

This is a selection of technically consequential items with enough evidence for a research assessment. It is not an inventory of every company announcement, every region, or every new model listing. The five new series are Tencent Hunyuan, ByteDance Seed, Apple ML Research, Ai2 and IBM Research; all fifteen existing lab series receive new coverage.

| Lab | Source date, 2026 | Assessment | Primary starting points | Evidence boundary |
|---|---|---|---|---|
| [OpenAI](/blog/series/openai-research) | Nov 9 | [Read the assessment](/blog/openai-control-room-revocation-window) | [Control Room](https://openai.com/index/control-room/) and [misalignment monitoring](https://openai.com/index/how-we-detect-frontier-model-misalignment/) | The 35-second objective is scoped to internal inference revocation after receipt of a stop instruction. |
| [OpenAI](/blog/series/openai-research) | Nov 13 | [Read the assessment](/blog/openai-chatgpt-wallet-approval-order-reconciliation) | [Wallet launch](https://openai.com/index/introducing-chatgpt-wallet/) and [purchase guide](https://help.openai.com/en/articles/20001654-making-purchases-with-chatgpt-wallet) | Initial dots rollout; approval is distinct from a merchant order and fulfillment. |
| [Anthropic](/blog/series/anthropic-research) | Nov 9 | [Read the assessment](/blog/anthropic-fable-5-2-prototype-verification) | [Fable and Mythos 5.2](https://www.anthropic.com/claude-fable-and-mythos-5-2) and [electronics prototypes](https://www.anthropic.com/research/claude-electronics-design) | The documented builds are functioning prototypes, with qualification outside the evidence. |
| [Google DeepMind](/blog/series/google-deepmind-research) | Nov 10 | [Read the assessment](/blog/google-alphaproof-code-proof-boundary) | [AlphaProof Code](https://deepmind.google/blog/introducing-alphaproof-code/) and [verification artifacts](https://google-deepmind.github.io/code-verification/) | Proofs cover selected properties of a compiled harness under stated assumptions. |
| [Google DeepMind](/blog/series/google-deepmind-research) | Nov 12 | [Read the assessment](/blog/google-gemini-hydrogen-orchestrator-handoff-economics) | [Hydrogen](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/) and [Antigravity Orchestrator](https://antigravity.google/blog/gemini-4-hydrogen-in-google-antigravity) | Standalone and routing results are separate. The routing benchmark version and enterprise entitlements differ across first-party pages. |
| [Google DeepMind](/blog/series/google-deepmind-research) | Nov 11 | [Read the assessment](/blog/google-gemini-decisions-confidence-routing) | [Gemini Decisions](https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/) and [OpenAI Decisions GA](https://openai.com/index/decisions-api/) | Compare confidence semantics and batching before using the APIs for selective routing. |
| [Meta](/blog/series/meta-research) | Nov 9 | [Read the assessment](/blog/meta-muse-ember-workflow-cost) | [Muse Ember](https://research.meta.ai/blog/muse-ember) and [evaluation methodology](https://research.meta.ai/static/muse-ember-evaluation-report) | Completion metrics vary by benchmark; repricing above 200K input matters at the trajectory level. |
| [DeepSeek](/blog/series/deepseek-research) | Nov 13 | [Read the assessment](/blog/deepseek-managed-application-apis-control-boundaries) | [API updates](https://api-docs.deepseek.com/updates/) and [chatbot guide](https://api-docs.deepseek.com/guides/build_a_chatbot/) | Managed application APIs wrap tool use and change which parameters and evidence the caller controls. |
| [Alibaba Qwen](/blog/series/qwen-research) | Nov 11 | [Read the assessment](/blog/qwen4-max-preview-shared-output-budget) | [Qwen4-Max-Preview](https://help.aliyun.com/en/model-studio/qwen4-max-preview) | Token Plan preview, combined reasoning/answer output budget, and documented retirement. |
| [xAI / SpaceXSI](/blog/series/xai-research) | Nov 2 | [Read the assessment](/blog/xai-grok-bot-1-0-scheduled-analytics-evidence) | [Grok Bot 1.0](https://x.ai/news/grok-bot-1-0) and [Data Bot](https://x.ai/news/data-bot) | Recurring analysis needs a metric owner, a permitted audience and a defensible labor counterfactual. |
| [Mistral](/blog/series/mistral-research) | Oct 29 / Nov 5 | [Read the assessment](/blog/mistral-studio-durable-agent-recovery-boundaries) | [Large 4 weights](https://mistral.ai/news/mistral-large-4-oss/) and [Studio](https://mistral.ai/news/build-with-studio/) | Worker placement, remote session storage, and external effect confirmation are distinct. |
| [Microsoft AI](/blog/series/microsoft-ai-research) | Nov 5 | [Read the assessment](/blog/microsoft-pdd-kit-proof-boundary-review) | [PDD Kit](https://www.microsoft.com/en-us/research/blog/formally-verifying-agent-written-full-stack-web-apps-with-pdd-kit/) | A demonstrated Rust login proof and its assumptions are assessable without implying product-wide correctness. |
| [Amazon / AWS](/blog/series/amazon-research) | Nov 12–13 | [Read the assessment](/blog/amazon-agentcore-v2-snapshot-session-economics) | [Runtime V2 default](https://aws.amazon.com/about-aws/whats-new/2026/11/new-agentcore-runtime-default/) and [snapshot patterns](https://aws.amazon.com/blogs/machine-learning/architecture-patterns-for-fast-session-startup-on-amazon-bedrock-agentcore-runtime/) | Dated rollout and developer-guide defaults differ; benchmark warmed state and session accounting explicitly. |
| [NVIDIA](/blog/series/nvidia-research) | Oct 21 / Nov 4 | [Read the assessment](/blog/nvidia-nemo-compass-eval-author-trace-holdout) | [Agent Toolkit](https://blogs.nvidia.com/blog/enterprise-platforms-agent-toolkit/) and [trace-to-test walkthrough](https://developer.nvidia.com/blog/use-nvidia-nemo-to-turn-agent-failures-into-tests-and-harness-fixes/) | Trace-derived diagnosis and tests need independent adjudication and held-out incidents. |
| [Cohere](/blog/series/cohere-research) | Oct 21 / Nov 4 | [Read the assessment](/blog/cohere-rerank-5-context-focus-evidence-retention) | [Rerank 5](https://cohere.com/blog/rerank-5) and [OpenRouter availability](https://cohere.com/blog/embed-rerank-5-openrouter) | Compare evidence survival as well as ranking; API truncation and endpoint feature parity require checks. |
| [Moonshot / Kimi](/blog/series/moonshot-research) | Oct 24 | [Read the assessment](/blog/kimi-k3-1-pro-harness-and-open-weight-economics) | [K3.1 Pro](https://www.kimi.com/en/blog/kimi-k3-1-pro) and [weight repository](https://huggingface.co/moonshotai/Kimi-K3.1-Pro) | Separate matched-harness model comparisons from hosted and self-hosted deployment economics. |
| [MiniMax](/blog/series/minimax-research) | Nov 11 | [Read the assessment](/blog/minimax-h3-image-accepted-asset-economics) | [H3-Image](https://www.minimax.io/blog/minimax-h3-image) and [model card](https://huggingface.co/MiniMaxAI/MiniMax-H3-Image) | Compare accepted asset cost after requirements, failed candidates, and repairs. |
| [Z.ai / GLM](/blog/series/zai-research) | Oct 20–Nov 13 | [Read the assessment](/blog/zai-zcode-worktrees-remote-session-authority) | [ZCode changelog](https://zcode.z.ai/en/changelog) and [remote development](https://zcode.z.ai/en/docs/remote-development) | Worktrees isolate checkouts; long-lived sessions still need separate process, identity, and continuation checks. |
| [Tencent Hunyuan](/blog/series/tencent-research) | Nov 4 | [Read the assessment](/blog/tencent-hy4-multimodal-evidence-handoffs) | [Hy4 launch](https://www.tencent.com/tencent-launches-hy4-joining-the-ranks-of-the-worlds-leading-open-source-ai-models/) and [preview artifacts](https://github.com/Tencent-Hunyuan/Hy4-preview) | Final hosted Hy4 and August preview weights are different artifacts; visual reading needs record verification. |
| [ByteDance Seed](/blog/series/bytedance-seed-research) | Nov 9 | [Read the assessment](/blog/bytedance-seedream-5-5-edit-invariants) | [Seedream 5.5](https://seed.bytedance.com/en/blog/more-realistic-images-more-stable-editing-across-100-rounds-introducing-seedream-5-5) and [ModelArk docs](https://docs.byteplus.com/en/docs/modelark/seedream-5-5) | Repeated edit, extracted layers and exported raster have different preservation criteria. |
| [Apple ML Research](/blog/series/apple-ml-research) | Nov 5 | [Read the assessment](/blog/apple-amortized-retrieval-index-drift) | [Retrieval highlight](https://machinelearning.apple.com/research/faster-retrieval) and [July ICML paper](https://proceedings.mlr.press/v306/olausson26a.html) | This is a November research highlight of earlier work; learned routing couples the index to a query distribution. |
| [Ai2](/blog/series/ai2-research) | Nov 2–3 | [Read the assessment](/blog/ai2-statebench-current-state-reconstruction) | [StateBench preprint](https://arxiv.org/html/2611.03094v1) and [research release](https://allenai.org/blog/statebench) | Controlled update dependencies expose failures that plain long-context retrieval can miss. |
| [IBM Research](/blog/series/ibm-research) | Nov 12 | [Read the assessment](/blog/ibm-granite-llmon-context-structure) | [LLMON release](https://huggingface.co/blog/ibm-granite/llmon-an-llm-native-markup-language-now-in-granite) and [8B model card](https://huggingface.co/ibm-granite/granite-4.1-8b-llmon-preview) | Preview checkpoints make instruction-to-input bindings explicit; general benchmark parity varies by task. |

## Differences and limits that affect interpretation

- **Availability:** The Wallet Help Center describes the initial dots rollout, despite a broader product description. Tencent announced final weights and a system card as forthcoming; the public Hugging Face query on November 14 listed only preview variants. Moonshot's release entry and model repository dates are distinct observations. The Qwen article flags its scheduled endpoint retirement.
- **Version and accounting:** The two Google launch descriptions name different Terminal-Bench versions for the same Orchestrator figures. The AWS dated announcement and developer guide describe different defaults for omitted versions. The corresponding articles preserve both accounts and propose checks for the actual deployed configuration. Product quota is not API cost.
- **Publication types:** Apple's November highlight discusses July ICML research. Ai2's preprint, release post, and evolving data repository are cited separately. Formal results from Microsoft and Google are bounded to the stated properties, artifacts and assumptions.
- **Research limits:** These are AI-assisted source reviews with original engineering analysis. Vendor, researcher, or partner measurements remain attributed. We did not reproduce model benchmarks, run user accounts or vendor deployments, conduct payments, train models, build proof artifacts, or measure the proposed adoption tests.

## Other developments in the window

The scope does not treat a single article as complete coverage of a lab. OpenAI's [product index](https://openai.com/news/product-releases/) also lists GPT-6.1 Terra, Images 2.6 and GPT-6 Cyber; this batch focuses on inference revocation, purchase reconciliation, and the Decisions GA comparison. Anthropic's [newsroom](https://www.anthropic.com/news) includes other science and company announcements; the Fable assessment focuses on the documented electronics builds. Google's [Carbon availability announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/) provides context to two posts but has no separate deployment assessment. Mistral's [research index](https://mistral.ai/news/) also lists Voxtral 3. Ai2's [research index](https://allenai.org/research) lists XYBench and ToyBlocks. IBM's [Granite Switch patch](https://huggingface.co/blog/ibm-granite/granite-switch-4-1-1) and [speech releases](https://huggingface.co/blog/ibm-granite/granite-speech-5) are distinct from the selected LLMON work. These are documented follow-up candidates; the linked release facts are not reproduced results.
