# AI lab coverage audit — October 19, 2026

## Scope and method

This is a dated editorial audit, not a model leaderboard or an exhaustive catalog of every AI organization and announcement. It adds 17 research assessments across 15 lab series: 12 existing series reused and three new series for Moonshot/Kimi, MiniMax, and Z.ai/GLM. OpenAI and Meta each receive two articles because their uncovered developments raise separate research questions.

The audit compared the existing blog bodies, the series registry, and the previous [October 6 source ledger](/research/ai-labs-october-2026/coverage-and-sources.md), including articles revised October 16. Novelty was determined from the actual coverage, not merely the publication date. Searches also covered the reference types, foundations, and implementation docs before describing a dedicated coverage gap. The live site's headers and sitemap were retrieved on October 19; a successful HTTP response does not prove that every repository article is deployed.

Sources are official announcements, documentation, research papers, repositories, and direct evaluator reports. Article citations support the adjacent claims. Announcement dates, later page updates, and the October 19 access date are separate. Living references are identified as such. Preview, beta, promised weights, and available products are not interchangeable.

AI assisted the research, writing, and review. No independent human review is claimed. No paid model calls, cloud deployments, GPU benchmarks, voice trials, financial simulations, or scientific-result reproductions were performed. Experiments, adoption gates, and economic examples in the articles are original proposals or explicitly illustrative arithmetic. External results remain attributed to their publishers or evaluators.

## Coverage decisions

| Lab | Earlier coverage inspected | New assessment and research contribution | Announcement dates | Series action |
|---|---|---|---|---|
| OpenAI | DevDay economics, Astra, developer platform, multimodal stack, science | [Live Decisions and latency budgets](/blog/openai-october-live-decisions-latency-budget): transport versus generation versus confirmed completion. [Science verification capacity](/blog/openai-october-science-verification-capacity): mathematical manuscripts, research access, and the verification bottleneck | October 6, 8, 15, 16 | Reuse `openai-research` |
| Anthropic | Opus/Sonnet 5.5, Claude platform, safety and product stack | [Haiku delegation](/blog/anthropic-haiku-5-5-delegation-partitioning): context partitioning, effort routing, pricing breakpoint, and total parent/child cost | October 7 | Reuse `anthropic-research` |
| Google DeepMind | Restricted Argon/private memory and October 19 Harness² article | [Argon ADK optimization](/blog/google-argon-adk-optimization-acceptance): public rollout and the acceptance funnel from six proposals to two shipped optimizations | October 13 | Reuse `google-deepmind-research` |
| Meta | Muse stack, security architecture, small-business approvals | [Wonder memory/focus](/blog/meta-muse-wonder-memory-focus-evaluation): separate memory from persistence. [Spark Open serving economics](/blog/meta-muse-spark-open-serving-economics): self-hosting, weight memory, precision, and complete serving costs | October 12/13 and 15 | Reuse `meta-research` |
| DeepSeek | Harness audit and September V4.1 Flash/cache/migration assessment | [Search, Fetch, and Flash-1017](/blog/deepseek-search-fetch-flash-1017-evidence-economics): evidence acquisition and model changes require separate baselines | October 17 | Reuse `deepseek-research` |
| Alibaba Qwen | Planner-Agent co-evolution | [E-Commerce Bench](/blog/qwen-ecommerce-bench-profit-solvency-evaluation): long-horizon solvency, fraud exposure, memory experiments, and residual variation in the environment | August 31 preprint; September 3 index entry | Reuse `qwen-research`; explicitly a missed-coverage catch-up |
| xAI | Grok 4.7 and Team Bots | [Grok Voice 3.0](/blog/xai-grok-voice-3-resolution-latency-economics): speech playback, tool state, benchmark denominators, and cost per resolved call | October 14 | Reuse `xai-research` |
| Mistral | Vibe knowledge and artifacts | [Large 4 and Timestral](/blog/mistral-timestral-large-4-forecasting-routing): general reasoning versus numerical forecasting, issue-time covariates, and hybrid routing | October 6 and 15 | Reuse `mistral-research` |
| Microsoft AI | Streaming MAI audio and October 12 trust-architecture analysis | [MAI-Base-2 and Code Flash](/blog/microsoft-mai-base-2-code-flash-efficiency-evidence): fitted training efficiency versus accepted-code cost and local-deployment economics | October 7 and 15 | Reuse `microsoft-ai-research` |
| Amazon | Bedrock Managed Agents preview | [Nova 2.5 Sonic and Strands Bidi](/blog/amazon-nova-2-5-sonic-bidi-conversation-continuity): conversation consistency through reconnection, delayed tools, and real audio | October 5 | Reuse `amazon-research`; missed subject predating the previous article |
| NVIDIA | OpenShell/Sentry containment | [Nemotron 3.5 Super VL/Bolt](/blog/nvidia-nemotron-3-5-super-vl-specialization-economics): specialization, serving mechanisms, quantization, and training-artifact limits | October 15 | Reuse `nvidia-research` |
| Cohere | Embed 5 and multilingual retrieval | [Coding-agent teams](/blog/cohere-coding-agent-teams-planning-economics): information gathering, lead attention, partial reward, and integration economics | October 14 | Reuse `cohere-research` |
| Moonshot/Kimi | Kimi appears in FrontierHarness and inference articles | [Cache and search economics](/blog/moonshot-kimi-cache-search-evidence-economics): separate TTL/write economics from retrieval completeness and source validity | September 16 and 28 | Create `moonshot-research`; explicitly a missed-coverage catch-up |
| MiniMax | No dedicated lab series or M3.1 Flash assessment in the inspected registry/content | [M3.1 Flash](/blog/minimax-m3-1-flash-agent-economics-assessment): cheap-token claims, mandatory reasoning, adapter compatibility, and acceptance experiments | October 19 | Create `minimax-research` |
| Z.ai/GLM | GLM appears as a frozen executor in Harness² | [GLM-Legal](/blog/zai-glm-legal-domain-specialization-evidence): co-training, external evaluator evidence, criterion coverage versus all-pass completion | October 12; AutoClaw cross-post October 13 | Create `zai-research` |

## Primary source index

All sources were checked October 19, 2026. Detailed dates, claim scope, and limits appear in each article's source ledger.

- OpenAI: [API changelog](https://developers.openai.com/api/docs/changelog), [native voice Decisions](https://developers.openai.com/api/docs/guides/decisions-voice), [Ultrafast](https://developers.openai.com/api/docs/guides/ultrafast-mode). Science sources are linked individually in the science article; different releases and research-access programs are not treated as one model announcement.
- Anthropic: [Haiku 5.5 launch](https://www.anthropic.com/claude-haiku-5-5), [overview](https://platform.claude.com/docs/en/models/haiku-5-5/overview), [migration](https://platform.claude.com/docs/en/models/haiku-5-5/migration-guide). The linked system card exceeded the retrieval tool's size limit and its detailed results are not assessed.
- Google: [public Argon rollout](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/), [ADK optimization account](https://developers.googleblog.com/optimizing-a-200000-line-python-codebase-overnight-with-gemini-4-argon/). First-party case study, not an independent benchmark run.
- Meta: [Wonder launch](https://about.fb.com/news/2026/10/introducing-muse-wonder-models/), [research post](https://research.meta.ai/blog/muse-wonder), [methodology card](https://research.meta.ai/static/muse-wonder-model-card). Spark Open's release and model documentation are linked in its separate article.
- DeepSeek: [October 17 changelog](https://api-docs.deepseek.com/updates/), [Search and Fetch guide](https://api-docs.deepseek.com/guides/search/). Living billing and capacity terms are not a tested account invoice.
- Qwen: [versioned paper](https://arxiv.org/abs/2608.30730v1), [full text](https://arxiv.org/html/2608.30730v1), [research index](https://qwen.ai/research), [official repository](https://github.com/QwenLM/E-CommerceBench), [project results](https://ecbench.github.io/). Later project results are distinguished from the paper's original roster.
- xAI: [Voice 3 release](https://x.ai/news/grok-voice-3), [speech guide](https://docs.x.ai/developers/model-capabilities/audio/speech-to-speech), [Artificial Analysis](https://artificialanalysis.ai/speech-to-speech). Evaluator metrics retain their definitions; a final tool call is not proof of a durable customer outcome.
- Mistral: [Large 4](https://mistral.ai/news/mistral-large-4/), [Timestral](https://mistral.ai/news/timestral/), [inference repository](https://github.com/mistralai/timestral). Preview API access, promised weights, and differing code/weight licenses remain separate.
- Microsoft: [MAI-Base-2 pretraining](https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/), [Code Flash](https://microsoft.ai/news/mai-code-1-1-flash-br-better-faster-at-a-quarter-of-the-cost/). Post-training and later local-product integration were not treated as completed releases.
- Amazon: [Sonic launch](https://aws.amazon.com/about-aws/whats-new/2026/10/amazon-nova-2.5-Sonic/), [Bidi GA](https://strandsagents.com/blog/bidi-agents-now-ga/), [service card](https://docs.aws.amazon.com/ai/responsible-ai/nova-2-5-sonic/overview.html), [living model catalog](https://aws.amazon.com/nova/models/).
- NVIDIA: [family launch](https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/), [technical report](https://developer.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning/), [model reference](https://docs.api.nvidia.com/nim/reference/nvidia-nemotron-3_5-super-vl-120b-a12b), [training recipe](https://github.com/NVIDIA-NeMo/Nemotron/blob/main/docs/nemotron/super35/README.md).
- Cohere: [October 14 team-scaling study](https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents), [ProgramBench background](https://arxiv.org/abs/2605.03546v1). The study's use of third-party models is not presented as a Cohere checkpoint launch.
- Moonshot: [September 16 search](https://forum.moonshot.ai/t/new-web-search-api-is-now-avaliable/606), [September 28 cache](https://forum.moonshot.ai/t/context-caching-upgrade-5-min-1-hour-ttl-options/613), [current caching rules](https://platform.kimi.ai/docs/guide/context-caching). Arithmetic uses mutually exclusive write/read/miss accounting.
- MiniMax: [October 19 launch](https://www.minimax.io/blog/minimax-m31f), [model catalog](https://platform.minimax.io/docs/guides/models-intro), [compatible API](https://platform.minimax.io/docs/api-reference/text-anthropic-api). Launch HTML was retrieved directly after the web reader failed to open it.
- Z.ai: [GLM-Legal](https://z.ai/blog/glm-legal), [AutoClaw cross-post index](https://autoclaw.z.ai/blog/), [Vals model record](https://www.vals.ai/models/zai_glm-legal), [LAB methodology](https://www.vals.ai/benchmarks/hlab). The vendor page's indexed text was readable while a direct web open returned empty; Vals was opened directly. No general-availability or weight-license claim is inferred from an evaluator's catalog label.

## Unresolved source differences

- AWS's October 5 Nova 2.5 Sonic announcement specifies 256K context; the current Nova models page specifies up to 1M. The article preserves both and requires endpoint-specific verification before capacity planning. This audit does not invent an intervening expansion date.
- NVIDIA's broad release language and its narrower repository status differ in scope. The documented exact SFT/RL blends and some training components remain pending. Running released weights, domain adaptation, and reproducing the original training are different claims.
- Living evaluator tables can disagree with older narrative paragraphs about model population and rank. GLM-Legal's article uses explicitly identified metric values instead of asserting a permanent position.

## Additional labs checked and selection limits

ByteDance Seed and Baidu were also checked. The inspected [Seed updates](https://seed.bytedance.com/en/) list an August 5 audio-visual release and October 14 recruiting; the [foundation-model index](https://seed.bytedance.com/en/blog_list/foundation) lists Seed2.1 on June 23. ByteDance's September HarnessDev research already appears in our [self-improving harness article](/blog/self-improving-harnesses-harnessdev-hoh). The inspected [ERNIE model blog](https://ernie.baidu.com/blog/) lists its latest model-release article on May 9. No additional recent, substantive, uncovered announcement was verified for either in this pass. That is a bounded search result, not a claim that neither organization published anything elsewhere.

Fundraising, hiring, promotional token events, rumors, future announcements, and already-covered subjects were not used to manufacture a fresh article. “Big AI labs” has no universal fixed membership; this audit makes its 15-series publication scope and two additional checks explicit. It does not claim complete coverage of every science, policy, model, or product update from those organizations.

## Repository verification

- All 660 tests in 40 test files passed; existing publishing and category-assignment tests were unchanged.
- ESLint, TypeScript, and the production build passed. The build generated 686 pages, and Pagefind indexed 502 public pages.
- All 17 new article HTML files and three new series HTML files were checked for generation, article titles, series targets, and audit links.
- Browser checks inspected the three new series and the MiniMax article, including its rendered provenance and tables. Other new articles were compiled and checked as generated HTML, rather than individually browsed.
- These checks establish repository integration and rendering. They do not validate vendor model claims, reproduce research results, or establish a live production deployment.