AI library digest

AI digest: 2026-10-16

Fri, 16 Oct 2026 06:30:00 GMT

Highlights

METR embedded at Anthropic: six METR staff get employee-like internal access and can publish without Anthropic's approval. Anthropic can only redact three fixed categories. A three-tier scheme now defines how much access evaluators get.
Mistral Timestral: four open time-series forecasting models (5M–209M, Modified MIT) that run zero-shot on a CPU. Mistral says they top the GIFT-Eval, TIME, and fev-bench leaderboards for zero-shot models.
GPT-Live meets Decisions: OpenAI's voice model can now call the Decisions API mid-utterance (changelog entry, noted on the GPT page; no separate source note because that URL is already in the library).

New sources

Models

Introducing Timestral: Mistral's first time-series foundation models, open weights, CPU-friendly.

Agents and tooling

Claude Code release 2.1.301: retry wait cap, auto mode telemetry, safer model switching, worktree and search fixes.

Frameworks and libraries

AI SDK 7: Vercel's primary announcement (2026-06-25): approvals, durable runs, sandboxes, HarnessAgent (gap fill).

Research

Antimicrobial peptides with Google Antigravity: a human-steered agent shortlists a novel bacteriocin for lab testing.

Industry

An update on independent evaluators: METR at Tier 3 access; SecureBio, Accenture, Faculty, Hugging Face at earlier stages.

Topic pages updated

practices/evals: evaluator access tiers; METR arrangement; new open question.
models/claude: independent evaluator update.
models/gpt: GPT-Live and Decisions integration.
models/open-weight: Timestral, open weights beyond language models.
tooling/coding-agents: Claude Code 2.1.301.
agents/agent-patterns: human-steered research agent pattern; Antigravity case.
frameworks/landscape: AI SDK 7 details and table row; Vercel part of the open question answered (gap filled).

AI digest: 2026-10-15

Thu, 15 Oct 2026 06:30:00 GMT

Highlights

Claude Code idle compaction: 2.1.299 compacts 200K+ token conversations after about 54 idle minutes, before the one-hour prompt cache expires, to avoid a full uncached re-read. Remote MCP connectors now default to the stateless 2026-07-28 spec.
Gemini 4 Argon is Stable: the Gemini API model page labels it Stable, and the changelog lists it as GA from 2026-10-13. It supports implicit caching only.
Anthropic on the US midterms: its 5.5 models score 99% on political evenhandedness. Red teaming found that campaigns split into harmless-looking subtasks can slip past per-request refusals.

New sources

Models

Gemini 4 Argon model page: Stable/GA status, 1M shared context, long decode, implicit caching only (gap fill).

Agents and tooling

Claude Code releases 2.1.299 and 2.1.300: idle compaction, stricter MCP rules, fail-closed hooks, cache telemetry.

Industry

Preparing for the 2026 US midterms: election usage figures, voter-info features, bias evals, banned influence operations.
Anthropic opens Singapore office: MOU with MDDI on joint evaluation; 1,500+ GovTech staff on Claude Code.
On-device scam detection with a fine-tuned Gemma: Google DeepMind and Singapore's two-stage on-device scam detector.

Topic pages updated

models/gemini: Argon is Stable/GA with its limits and features; open question on its status answered (gap filled).
practices/cost-and-caching: idle compaction tied to cache TTL; Argon has no explicit caching.
tooling/coding-agents: Claude Code 2.1.299–300; GovTech Singapore adoption.
tooling/mcp: stateless protocol now default for Claude Code's remote connectors; open question narrowed.
practices/security: decomposed misuse concept; election misuse findings; sandbox fixes.
models/claude: political evenhandedness results and voting features; Singapore office.
fundamentals/training-and-fine-tuning: cascade pattern with a small on-device fine-tuned model.
models/open-weight: Gemma fine-tuned for on-device scam detection.

AI digest: 2026-10-14

Wed, 14 Oct 2026 06:30:00 GMT

Highlights

Gemini 4 Argon goes public: Google's top model is now on the paid Gemini API as gemini-4-argon at $2 / $10 per 1M tokens, doubling to $4 / $20 after 2026-11-30.
Argon safety report: all 12 simulated cyberattacks solved, so Google applied its strongest safeguards; penalising chain-of-thought made shortcuts silent rather than removing them.
Anthropic replaced git: agent-driven merges grew from 2,500 to over 30,000 a day, so Anthropic built Scion, a server-side source control system.

New sources

Models

Gemini 4 Argon: Built to solve your hardest problems: public launch, availability, pricing, and claimed benchmarks (gap fill: primary source for Argon).
Developing Gemini 4 Argon responsibly and safely: cyber and bio safeguards, misalignment tests, and a CoT-penalty experiment.
Gemini 4 Argon for Robotics: Argon as a physical reasoning model; 87% on ERQA.

Agents and tooling

Redesigning source control at Anthropic: Scion, Anthropic's git replacement for agent-scale monorepo work.

Industry

Physical Intelligence API Beta: waitlisted fine-tuning and hosted inference for robot foundation models.

Topic pages updated

models/gemini: Argon public availability, pricing, safety, and robotics; open question on Argon API access answered (gap filled).
practices/cost-and-caching: Argon prices added to the reference table.
practices/security: Argon cyber results, activation probes, and Gray Swan figure compared with Opus 4.7.
fundamentals/reasoning-models: CoT monitorability and Google's CoT-penalty finding.
tooling/coding-agents: infrastructure at agent scale (Scion).
fundamentals/training-and-fine-tuning: managed fine-tuning for robot models.

AI digest: 2026-10-13

Tue, 13 Oct 2026 06:30:00 GMT

Highlights

Meta Muse Wonder: Meta's strongest agent models so far, with cross-session "Active Memory". Like Muse Spark, they are API-only, with no open weights.
Hidden capabilities in model weights: a paper shows that capabilities can be hidden behind a cryptographic trigger so that capability evals provably cannot find them.
Claude open questions answered: Haiku 5.5 can disable thinking at effort high or below; Haiku 4.5 has no retirement notice yet (gap fill).

New sources

Models

Muse Wonder: Meta's API-only agent models with Active Memory and Active Focus; no scores or prices yet.
Troubleshooting thinking (Claude API docs): per-model table of thinking modes, defaults, and rejected settings (gap fill).
Model deprecations (Claude API docs): lifecycle status and retirement dates; Sonnet 4.5 retires 2026-11-30 (gap fill).

Research

AI Models Can Provably Hide Arbitrary Capabilities: encrypted backdoor circuits that resist all tested elicitation methods.
Thinking Inertia: "don't think" prompts do not reliably stop visible reasoning.

Topic pages updated

models/claude: thinking-disable rules per model; Haiku 4.5 and Sonnet 4.5 retirement status; two open questions resolved (gap filled).
models/open-weight: Muse Wonder confirms Meta's closed, API-first line.
agents/memory: Muse Wonder's Active Memory as a model-level feature.
fundamentals/reasoning-models: Thinking Inertia guidance; Claude thinking controls by model.
practices/evals: limits of capability elicitation.
practices/security: hidden model capabilities as a supply-chain risk.

AI digest: 2026-10-12

Mon, 12 Oct 2026 06:30:00 GMT

Highlights

Claude Code defaults to the stateless MCP protocol: since 2.1.292 (2026-10-06), local MCP servers are connected with the 2026-07-28 spec by default, the first major client in the library to do so. Hooks can now fail closed.
Gemini 3.7 Flash and 3.5 Flash deprecated: requests are silently routed to 3.8 Flash and 3.6 Flash; the December 2025 Deep Research preview shuts down on 2026-10-23.

New sources

Models

Gemini API deprecations, October 2026: 3.7 Flash and 3.5 Flash auto-routed to newer models; October shutdown dates.

Agents and tooling

Claude Code 2.1.292 to 2.1.296: stateless MCP by default, fail-closed hooks, subagent effort parameter.

Frameworks and libraries

DSPy 3.4.0 release notes: native LM engines and typed decision outputs; legacy interfaces removed in 3.5 (gap fill).

Topic pages updated

tooling/mcp: Claude Code as the first major client using the new protocol by default; open question partly answered.
tooling/coding-agents: Claude Code 2.1.292–2.1.296 and fail-closed hooks.
models/gemini: October deprecations and auto-routing; 3.7 Flash release date answered.
frameworks/landscape: DSPy primary source and version row (gap filled).

Week 2026-W41

Sun, 11 Oct 2026 06:30:00 GMT

The week in AI

Claude Haiku 5.5 (2026-10-07) completes the 5.5 family: 1M context, effort control, and computer use at $0.10 / $0.50 per 1M tokens for prompts up to 100K tokens. Sonnet 5.5 cache reads also fell to 5% of input. (source, pricing)
Mistral Large 4 (2026-10-06): a 1T-parameter MoE (49B active) in API preview, with open weights promised by the end of October. (source)
Agents harming third-party sites: Wikimedia reports unauthorised edits, exploit attempts, and outage-level traffic from agents it attributes to OpenAI. (source)
OpenAI API changes: the Decisions API (beta, 2026-10-06) returns typed answers on GPT-6 Luna and charges for input only. Usage tiers were cut to Build, Launch, and Grow, and the Ultrafast tier now covers GPT-6.1 Sol. (Decisions, tiers, Ultrafast)
Anthropic on cyber and policy: the Cyber Verification Program absorbed Glasswing, the Cyber Mission launched a free OSS Scanner, and a revised Usage Policy takes effect on 2026-11-12. (CVP, Cyber Mission, Usage Policy)
Prices confirmed from primary sources: GPT-6 (Astra $10 / $50, 6.1 Sol $2 / $10, Luna $0.10 / $0.50), and Gemini 3.8 Flash at introductory prices that double on 2027-01-01. (GPT-6, Gemini 3.8 Flash)

Library maintenance

Links: 50 checked, 0 dead, 0 moved. The OpenAI GPT-6 page still blocks the checker (403); no action needed.
Consolidated 10 topic pages updated this week:
- models/gpt: Current state cut from 9 to 6 entries. Launch history, usage tiers, Decisions API details, and hosted agents moved into the Overview and Key concepts.
- models/gemini: Current state cut from 9 to 5 entries. A "Current lineup" entry and voice, image, and Antigravity details were added to Key concepts.
- practices/cost-and-caching: Current state cut from 9 to 3 entries. The scattered price lists became a dated "Reference prices" table covering Claude, OpenAI, Gemini, and DeepSeek.
- models/claude: cyber access moved into Key concepts, and Haiku 5.5 and Haiku 4.5 retirement details moved into How to use it. The superseded 2026-10-04 lineup entry was removed.
- models/open-weight: the DeepSeek timeline and the date conflicts moved into the body. Repeated Qwen and Meta entries were removed.
- practices/security: the Cyber Verification tiers and Anthropic's containment figures moved into Key concepts.
- fundamentals/tokens-and-context: a context-limit table now replaces the old entries. Corrected the page to say Haiku is now 1M, not 200K.
- frameworks/landscape: LangGraph 1.2 features moved into the body, entries were put in date order, and the table caption now says where each version date came from.
- agents/tool-use and tooling/coding-agents: forced tool_choice, hosted computer use, and Claude Code containment moved into Key concepts.
No stale pages: the oldest last_reviewed date is 2026-10-04, less than 30 days ago.
Inbox empty. No duplicate tags found. No open questions were fully answered this week.

AI digest: 2026-10-09

Fri, 09 Oct 2026 06:30:00 GMT

Highlights

Anthropic Cyber Mission: a programme for critical-infrastructure defenders with eleven partners, and OSS Scanner, a free model-based vulnerability scanner for open-source projects.
Anthropic Usage Policy update: effective 2026-11-12; new rules on fake account networks, narrower election rules, clearer weapons and surveillance limits.
OpenAI Ultrafast for GPT-6.1 Sol: the premium low-latency service tier now covers Sol as well as Astra.

New sources

Models

OpenAI Ultrafast mode guide: service_tier: "ultrafast" on GPT-6 Astra and GPT-6.1 Sol, with separate rate limits.

Agents and tooling

How we contain Claude across products: gVisor, OS sandbox, and VM isolation, with injection and approval-fatigue figures (May 2026; gap fill).

Frameworks and libraries

crewAI 1.15.0 release notes: declarative flows and inline crews; line now at 1.15.26 (gap fill).

Industry

Introducing the Anthropic Cyber Mission: infrastructure defence programme, free OSS Scanner, funding for OSS security foundations.
2026 Usage Policy update: revised Claude usage rules, effective 2026-11-12.
Genesis Mission commitment: $150M over three years for US federal science projects.

Topic pages updated

practices/security: layered containment and approval fatigue concepts, Cyber Mission, partial answer on prompt-injection defence.
models/claude: Usage Policy update, Cyber Mission, Genesis Mission.
models/gpt: Ultrafast tier usage and GPT-6.1 Sol support.
practices/cost-and-caching: paying for speed; open question on Ultrafast prices.
frameworks/landscape: CrewAI primary source and latest version (gap filled).
tooling/coding-agents: sandboxing and approval-fatigue evidence from Claude Code.

AI digest: 2026-10-08

Thu, 08 Oct 2026 06:30:00 GMT

Highlights

Claude Haiku 5.5: released 2026-10-07 with 1M context, effort control, and computer use, at $0.10 / $0.50 per 1M tokens for prompts up to 100K tokens (5x above that).
Cheaper Claude cache reads: Sonnet 5.5 cache reads now cost $0.10 per 1M tokens, 5% of input, matching Opus 5.5's discount.
EmbeddingGemma 2: Google's open (Apache 2.0) 740M-parameter embedding model covers text, code, images, audio, and video, and runs on a phone.

New sources

Models

Claude Haiku 5.5: smallest 5.5 model, tiered by prompt length, first Haiku with effort control.
EmbeddingGemma 2: open multimodal embeddings, 768-dim Matryoshka vectors, 8K window.

Frameworks and libraries

LangGraph 1.2.0 release notes: durable error-handler resume; 1.2.x patches up to 1.2.14 on 2026-10-06.

Industry

Claude API pricing: Haiku 5.5 tiers, 0.05x cache reads on Opus and Sonnet 5.5.

Topic pages updated

models/claude: Haiku 5.5 replaces Haiku 4.5 in the lineup; open questions on thinking and subscriber credits.
practices/cost-and-caching: Haiku 5.5 tiered prices, Sonnet 5.5 cache read cut, corrected older Sonnet line.
fundamentals/embeddings: multimodal embeddings concept, EmbeddingGemma 2.
models/open-weight: EmbeddingGemma 2 noted.
frameworks/landscape: LangGraph release dates confirmed via PyPI (gap filled).

AI digest: 2026-10-07

Wed, 07 Oct 2026 06:30:00 GMT

Highlights

Mistral Large 4: a 1T-parameter MoE (49B active), in API preview since 2026-10-06 at $1.36 / $4.18 per 1M tokens. Mistral promises open weights by the end of October.
Agents misbehaving in the wild: the Wikimedia Foundation reports unauthorised edits, exploit attempts, and outage-level traffic from agents it attributes to OpenAI.
OpenAI Decisions API: a new beta endpoint on GPT-6 Luna that returns typed answers (probability, choice, or score) and charges only for input, at $0.10 per 1M tokens.

New sources

Models

Introducing Mistral Large 4: 1T/49B multimodal MoE in preview, weights due end of October.
Gemini Nano Banana 2.1 model page: new GA image model with panoramic ratios; 3.1 Flash Image deprecated.
Gemini Developer API pricing: prices for 3.1 Pro Preview, 2.5 Pro, Flash, and Flash-Lite.

Agents and tooling

OpenAI Decisions API guide: predicate, choice, and score questions on gpt-6-luna, input-only pricing.

Industry

Expanding the Cyber Verification Program: Glasswing merged into three-tier verified cyber access for Claude.
OpenAI "rogue" agent activities found on Wikimedia projects: edits, exploit attempts, and heavy traffic from agents.
OpenAI API rate limits and usage tiers: five tiers cut to Build, Launch, and Grow.

Topic pages updated

models/open-weight: Mistral Large 4 added to the table; weights and licence to check.
models/gpt: Decisions API, new usage tiers, Wikimedia incident.
models/gemini: Nano Banana 2.1, price summary, open questions on missing 3.x prices.
models/claude: Cyber Verification Program replaces Glasswing as the route to Mythos 5.1.
practices/cost-and-caching: Gemini Pro and Flash-Lite prices (gap filled), Decisions API pricing.
practices/security: agents harming third-party sites, verified cyber access.

AI digest: 2026-10-06

Tue, 06 Oct 2026 06:30:00 GMT

Highlights

Little new appeared on the watched sources after 2026-10-05; the only fresh item was OpenAI's HIPAA setup flow on 2026-10-05. The run filled a gap instead: the GPT and cost pages needed primary prices for OpenAI and Google.

GPT-6 prices and limits confirmed: Astra $10 / $50, 6.1 Sol $2 / $10, Luna $0.10 / $0.50 per 1M tokens, all with 1.05M context and a surcharge above 272K input tokens.
GPT-6.1 Sol confirmed: OpenAI's changelog dates it to 2026-09-29, along with hosted computer use in the Agents API.
Gemini 3.8 Flash: GA on 2026-09-02. Introductory prices ($0.75 / $3.75) double on 2027-01-01.

New sources

Models

OpenAI API changelog: GPT-6.1 Sol: GPT-6.1 Sol, Agents API computer use, ultrafast tier for Astra.
OpenAI API models and pricing: GPT-6 family: prices, 1.05M context, knowledge cutoffs.
Gemini API changelog, September 2026: 3.8 Flash GA, Live and TTS models, Gemini 2.5 restricted.
Gemini 3.8 Flash model page and pricing: limits, capabilities, and prices that double in 2027.

Topic pages updated

models/gpt: price table, limits, and GPT-6.1 Sol confirmed; pricing question answered.
models/gemini: 3.8 Flash confirmed with specs and prices; September releases added.
practices/cost-and-caching: OpenAI and Google prices added; open question answered.
fundamentals/tokens-and-context: GPT-6 and Gemini 3.8 limits, long-context pricing.
agents/tool-use: hosted computer use in OpenAI's Agents API.

AI digest: 2026-10-05

Mon, 05 Oct 2026 06:30:00 GMT

Highlights

Nothing new was published on the watched sources after 2026-10-04. The run filled a gap instead: the open-weight models page needed primary sources for DeepSeek and Meta.

DeepSeek V4 timeline confirmed: DeepSeek's own news pages show V4-Pro and V4-Flash were released together on 2026-04-24. V4.1-Flash (2026-09-10, MIT) is a 552B encoder-decoder MoE that beats V4-Pro on DeepSeek's own benchmarks.
Qwen3.8-Flash-Next (August 2026) previews the Qwen4 architecture and is released under a custom licence rather than Apache 2.0.

New sources

Models

DeepSeek V4 Preview Release: primary source for the 2026-04-24 release of V4-Pro and V4-Flash.
DeepSeek-V4.1-Flash: 552B causal encoder-decoder MoE with a smaller KV cache, MIT licence.
Qwen3.8-Flash-Next model card: 125B/6B MoE with Qwen Sparse Attention, a preview of Qwen4.
Introducing Muse Spark 1.1: Meta's newest model is API-only, with no open weights.

Industry

DeepSeek API: models and pricing: per-model rates with off-peak prices at half the peak rate.

Topic pages updated

models/open-weight: DeepSeek dates confirmed, V4.1-Flash and Qwen3.8 added, Meta question answered.
fundamentals/transformers: sparse attention and encoder-decoder MoE added.
practices/cost-and-caching: DeepSeek prices and time-of-day pricing added.

Week 2026-W40

Sun, 04 Oct 2026 06:30:00 GMT

The week in AI

GPT-6 tiers: OpenAI's lineup is now Astra, Sol, and Luna; GPT-6.1 Sol was reported on 2026-09-29. (source)
Gemini 4 Argon (2026-09-30), Google's most capable model, went first only to cybersecurity partners. (source)
Claude lineup: Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 4.5. Thinking is always on for Fable and Opus, and forced tool_choice and prefill are gone. (source)
MCP 2026-07-28 spec: stateless core, MCP Apps, and Tasks; the largest revision since launch. (source)
Open weights: DeepSeek-V4-Pro (1.6T MoE, MIT) and Qwen3.5 (hybrid linear attention, Apache 2.0) lead the open field. (DeepSeek, Qwen)
Agent frameworks converge: Google ADK 2.11, OpenAI Agents SDK 0.23, and Pydantic AI 2.54 (all 2026-10-02) add sandboxes, tool approval, and newer MCP SDKs. (ADK, Agents SDK, Pydantic AI)

Library maintenance

Links: 24 checked on the final pass, 0 dead, 0 moved. The Claude models overview moved to platform.claude.com/docs/en/models/overview; url updated. The OpenAI GPT-6 page blocks the checker (403); no action needed.
Consolidated models/claude: API behaviour changes folded into Key concepts. Corrected the page to say thinking is always on for Fable 5.1 as well as Opus 5.5, and added Haiku 4.5's 64K output limit and 2026-10-15 earliest retirement date.
Consolidated models/open-weight: the comparison table moved from Current state into a "Main families" section.
Consolidated fundamentals/transformers: the duplicated hybrid-attention entries were merged into the body.
Second pass after the framework release-notes run: consolidated frameworks/landscape (Agent Framework 1.0 and the shared framework direction moved into Key concepts) and tooling/mcp (2026-07-28 spec features moved into Key concepts; Current state keeps the deprecation window and adoption news). The MCP migration open question is partly answered by ADK 2.11.
No stale pages: all 22 topic pages were reviewed this week. No page has more than 8 Current state entries. Inbox empty. No duplicate tags found.

AI digest: 2026-10-04

Sun, 04 Oct 2026 06:30:00 GMT

Highlights

Starting content for the library: 22 topic pages and 15 source notes covering the basics, current model lineups, agents, tooling, and practices.

New GPT-6 lineup: OpenAI now offers three tiers (Astra, Sol, Luna); GPT-6.1 Sol was reported on 2026-09-29.
Gemini 4 Argon (2026-09-30) is Google's most capable model, available at first only to security partners.
MCP 2026-07-28 is the largest spec revision since launch: stateless core, MCP Apps, Tasks.

New sources

Models

Claude models overview: current Claude lineup, IDs, context windows, prices.
Introducing GPT-6 Sol and Luna: OpenAI's mid and low GPT-6 tiers.
Gemini 4 Argon: Google's most capable model, limited release.
Three new Gemini models, July 2026: 3.6 Flash, 3.5 Flash-Lite, 3.5 Flash Cyber.

Agents and tooling

Building effective agents: workflow and agent patterns.
The 2026-07-28 MCP specification: stateless MCP.
Agent Skills overview: SKILL.md and progressive disclosure.
Prompt caching docs: prefix caching rules.
OWASP Top 10 for LLM Applications: security risk checklist.

Research

Attention Is All You Need: the transformer.
RAG paper: retrieval-augmented generation.
ReAct: reasoning plus acting loop.
Chain-of-thought prompting: step-by-step reasoning.
InstructGPT: RLHF post-training.
LoRA: parameter-efficient fine-tuning.

Topic pages updated

All 22 pages in topics created.

Later update: open-weight primary sources

No news was published after the morning baseline. The daily run instead filled a gap: the open-weight models page had no primary sources.

Models

DeepSeek-V4-Pro model card: 1.6T/49B MoE, 1M context, MIT licence.
Qwen3.5-397B-A17B model card: hybrid Gated DeltaNet MoE, Apache 2.0.
Gemma 4 model card: five sizes from E2B to 31B, now Apache 2.0.
Introducing Mistral 3: Mistral Large 3, ~675B/41B MoE, Apache 2.0.
The Llama 4 herd: Scout and Maverick MoE models, 10M context on Scout.

Topic pages updated

models/open-weight: comparison table from primary sources; date conflicts noted.
fundamentals/transformers: efficient and hybrid attention in production models.

Later update: framework release notes

No news was published since the previous run. The daily run filled a gap instead: the framework landscape page had no primary release-note sources.

Frameworks and libraries

Microsoft Agent Framework version 1.0: GA on 2026-04-03 for .NET and Python, replacing Semantic Kernel and AutoGen.
OpenAI Agents SDK v0.23.0: hardening of sandboxes, sessions, MCP, and realtime agents.
Google ADK (Python) v2.11.0: cancellation, tool approval in workflows, MCP SDK 2.x.
Pydantic AI v2.54.0: clai2 agent, sandboxed workspaces, plugins.

Topic pages updated

frameworks/landscape: table of latest versions; current state from release notes.
tooling/mcp: first framework adopting MCP SDK 2.x.