🗞️ GenAI Catch-up Report 2026-10-13

6 topics from 2026-10-10 to 10-13 (selected from 32 candidates). Details: research note.

1️⃣ 🎯 OpenAI's Decisions API in public beta at $0.10 per million input tokens

POST /v1/decisions on GPT-6 Luna returns probabilities, choices and rubric scores for text and images, with no output or cache charges. Independent tests: level with Jev on exact matching, behind on nuanced judgements, at 1.6x to 3.1x its cost.

Impact highPotential highAttention highReleaseModelsAPI and cloud

2️⃣ ⚖️ Microsoft-Decision-1: a Qwen-based decision model at Jev's price

Post-trained from Qwen3.5-9B, it returns a probability for each fixed answer at $0.042 per million input tokens, output free, in preview on Foundry and OpenRouter. Microsoft's charts put it first on accuracy and speed; one independent test had it slower than Jev.

Impact midPotential highAttention highReleaseModelsAPI and cloud

3️⃣ 🗃️ SQLite takes in vec1: ANN vector search in the 3.54.0 source tree

One C file adds IVF indexes with PQ or OPQ codes and in-index metadata filters, built separately as a loadable extension. SQLite's own benchmark on 769K text embeddings gives recall@10 of 0.62 without reranking in SQL and 0.93 to 0.98 with it.

Impact midPotential highAttention highReleaseAgent building

GenAI Catch-up Report · 2026-10-13 · 🤖 Generated by Claude Opus 5.5

🗞️ GenAI Catch-up Report 2026-10-13

4️⃣ 🛡️ AWS's Strands Box: an agent sandbox with history-aware Dogwood policies

A developer preview for Apple-silicon Macs pairs Seatbelt containment with Cedar-based rules that can depend on what the agent did earlier, judged at a gateway, a shell, Python and an MCP broker. On EC2 Mac, a host program the policy let run still read a forbidden file.

Impact midPotential highAttention highReleaseSecurity and safetyCoding agents

5️⃣ ✈️ An aviation-based essay: keep the ability to take over from coding agents

Drawing on FAA and EASA guidance, an engineer argues for keeping the ability to take over from AI rather than to code without it, reading high-risk diffs by hand instead of meeting quotas. Its citations hold up; most featured Hatena comments dispute the analogy.

Impact midPotential midAttention highAnalysisCoding agents

6️⃣ 🧩 Dynamic workflows in Claude Managed Agents: up to 1,000 agents per run

The session's agent writes a program that Anthropic's servers run in phases, 64 threads at a time, billed as tokens plus $0.08 per session-hour. Anthropic claims 66 of 70 planted bugs found against 14 to 27 for one agent, and warns of high token use.

Impact midPotential highAttention midReleaseAgent buildingAPI and cloud

GenAI Catch-up Report · 2026-10-13 · 🤖 Generated by Claude Opus 5.5

1️⃣ 🎯 OpenAI's Decisions API in public beta at $0.10 per million input tokens

Opened to all developers on 10-06, a week after the DevDay preview, with docs, SDKs in five languages and a Playground; GA is expected within weeks.

Key points

  • POST /v1/decisions runs on gpt-6-luna only: one shared input of text and images, and a questions array of mixed types.
  • predicate returns a probability, choice a value with probabilities and a confidence, score a weighted level index.
  • $0.10 per 1M input tokens, no output or cache charges; Luna via the Responses API bills $0.50 output, $0.01 cached input.
  • Classmethod's 900 requests: Japanese text scored 93/100 on predicate, 84/100 on a three-way choice; medians 209–218 ms.
  • Against Jev: tied on exact matches (106/107), behind on semantic (90–91 vs 96–97, Classmethod), r = 0.960 on scores (7shi).

Perspectives and debates

  • OpenAI claims up to 10x the speed of Luna via Responses, without a benchmark; 7shi's batch ran 4.8x longer via Responses.
  • Simon Willison expects /v1/decisions to be a de facto standard; The New Stack sees TypeSafe's schema winning instead.
  • Constraints: no calibration data, choice skewed and order-sensitive in a forum test, cost 1.6x to 3.1x Jev's as measured.

Technical details

  • Request: model, input (a string, or user messages with input_text and input_image) and questions; answers echo name.
  • choice takes 2 to 255 unique string or boolean values; any answer may be a refusal; dependent decisions need a second call.
  • Images: the guide allows inline base64 only, while the API reference takes up to 128, public HTTPS URLs included.
  • SDKs call client.decisions.create(...): Python 3.26.0, JavaScript 7.30.0, Go 3.73.0, Ruby 0.101.0 and Java 4.78.0.
  • ZDR and HIPAA for eligible customers, US and Europe data residency; long-context and regional multipliers still apply.
  • Unstated: Decisions rate limits, caps on questions or score levels, the GA date, other models, Azure or Bedrock.

Sources: OpenAI · Classmethod · 7shi (Zenn) · Previous: 10-01 #5

GenAI Catch-up Report · 2026-10-13 · 🤖 Generated by Claude Opus 5.5

2️⃣ ⚖️ Microsoft-Decision-1: a Qwen-based decision model at Jev's price

Announced on 10-09 in public preview on Foundry and OpenRouter, three days after OpenAI opened its Decisions API; it drew 199 points on Hacker News.

Key points

  • Qwen3.5-9B post-trained to score fixed options in one pass, for yes/no, choice, rating and rubric questions; text only.
  • $0.042 per 1M input tokens with output free, Jev's price, on Foundry (US and EU data zones listed) and OpenRouter.
  • Microsoft's 36 benchmarks, about 150,000 questions: top accuracy (83.5%, The Register), 2.5x the runner-up's P50 speed.
  • Microsoft: eight perturbation types flipped 1.3% of cases on average; reordering or paraphrasing options flipped none.
  • Measured: p50 369 ms vs Jev's 283 ms at 0.72x its cost (Topfi, HN), 0.29 s on OpenRouter, 395 ms in a Zenn OCR gate.

Perspectives and debates

  • The New Stack: Qwen over its partner's model, the prize being agent traffic kept on Foundry; MAI or OpenAI rebases undated.
  • Speed claims moved: 4.5x over Quyet-1.0-Large at launch, now 2.5x over H2O-Lightning-4B; calibration ranks 2nd or 3rd.
  • Microsoft Learn: wording and option order shift scores, best used for ranking and thresholds; JevOut attacks untested.

Technical details

  • Endpoint <foundry-resource-endpoint>/providers/microsoft/v1/systemone; model is the deployment name, not the model name.
  • System One shape: a state and a questions map of noul, choice or score with criteria, unlike OpenAI's /v1/decisions.
  • Version 1 deploys as GlobalStandard or, in some regions, DataZoneStandard, with Microsoft Entra ID or an api-key header.
  • Catalog: Preview, 32,768-token context, text in, JSON out; no images, audio, video or open-ended generation.
  • OpenRouter serves microsoft/microsoft-decision-1 through its Decisions API, so chat-completions SDKs do not work with it.
  • Questions in one request run in parallel; a noul plus a score used 86 input and 2 free output tokens, about $0.0000036.

Sources: Microsoft · Microsoft Learn · The New Stack · Hacker News

GenAI Catch-up Report · 2026-10-13 · 🤖 Generated by Claude Opus 5.5

3️⃣ 🗃️ SQLite takes in vec1: ANN vector search in the 3.54.0 source tree

Merged into trunk on 10-01 without an announcement and tagged at 0.8 in 3.54.0; mattn's Zenn write-up of 10-11 drew 236 Hatena Bookmark users.

Key points

  • Dan Kennedy imported vec1 into ext/vec1/ on 09-30 and merged it on 10-01; the version-3.54.0 tag of 10-09 carries 0.8.
  • It stays out of the amalgamation and every package, built only by make vec1.so; the 3.54.0 release notes are silent on it.
  • ANN via a virtual table: IVFADC with OPQ, PQ or binary codes, L2 and cosine, float32 only, AVX2 or NEON, no dependencies.
  • SQLite's v0.7 test, 769,382 1024-d text vectors: recall@10 0.62 without reranking, 0.930–0.961 with it, at 2,251–1,290 QPS.
  • Prebuilt routes: APSW's PyPI wheels via apsw.sqlite_extra.load(db, "vec1"), and the cgo-free ncruces/go-sqlite3 at 0.7.

Perspectives and debates

  • Dan Kennedy (July) puts it in the ballpark of most ANN engines for trainable data; the 10-01 roadmap sets no 1.0 date.
  • mattn values plain SQL with joins and WHERE but would wait before production; Hatena readers see small local RAG.
  • Robustness: 37 August check-ins fixed crashes and index corruption, fixes went on in October, and testing tops the roadmap.

vec1_train on about 100 × nbucket vectors, then rebuild Query vector Scan the nprobe share of IVF buckets, 0.05 by default About 100 candidates by compressed OPQ distance Rerank in SQL by exact distance, then LIMIT 10

Technical details

  • USING vec1(vector, meta1, ...) takes up to 255 untyped metadata columns; query t(:v, '{K:10}').
  • Metadata comparisons filter in the search; a bare WHERE in_stock filters after, leaving fewer than K rows.
  • sqlite-vec: no ANN index in stable v0.1.9 (alphas only), but int8 and binary vectors and WASM today.

Sources: vec1 manual · vec1 benchmarks · SQLite check-in · mattn (Zenn)

GenAI Catch-up Report · 2026-10-13 · 🤖 Generated by Claude Opus 5.5

4️⃣ 🛡️ AWS's Strands Box: an agent sandbox with history-aware Dogwood policies

Released on 10-07 as an Apache 2.0 developer preview for Apple-silicon Macs; a Zenn test on EC2 Mac mapped what its policy governs.

Key points

  • v0.1.0 ships only an Apple-silicon macOS 15+ build; Linux builds are held back, and Windows is under consideration.
  • Seatbelt bounds the agent; Dogwood rules judge requests at a gateway, Strands Shell, Monty for Python and an MCP broker.
  • Rules add temporal clauses to Cedar: three Slack posts per 10 minutes, or git push only after a passing npm test.
  • Zenn: an awk admitted by shell:spawn, with project reads in its tool table, printed .env despite the no_env forbid.
  • Zenn: the documented write budget never counted shell >> or tee -a writes; Strands Shell added about 9 ms per command.

Perspectives and debates

  • Marc Brooker: the kernel sandbox funnels traffic to a protocol-aware policy layer, yet "bugs could mean bypassed policy".
  • InfoWorld's analysts: consistent across frameworks, but new parts to trust, and bad policies block work or leave gaps.
  • No resource limits, tenant isolation, tamper-proof log or built-in IMDS block; Brooker adds a per-session microVM.

Agent process in Seatbelt File tools on box.toml grants:no decision, no log Commands, HTTP, MCP throughBox's gateway, shell, Python, broker Dogwood judges each requestwith the event history Host program run by shell:spawn:its own file access, no policy Files, hosts and MCP servers

Technical details

  • Ten actions, e.g. fs:read, shell:spawn, mcp:call; an outbound call needs net:connect and http:request.
  • The gateway swaps the agent's placeholder for the real secret only on permitted requests, SigV4 included.
  • Tutorials: Claude Code 2.1.292 and Codex CLI 0.160.1 run with their own checks off, editing through the shell.

Sources: AWS · Marc Brooker · strands-agents/box · Zenn (kashiwabaray)

GenAI Catch-up Report · 2026-10-13 · 🤖 Generated by Claude Opus 5.5

5️⃣ ✈️ An aviation-based essay: keep the ability to take over from coding agents

Published on Zenn on 10-05 by an engineer who stalled before approving a CDK diff an AI review had passed; it drew 159 Hatena Bookmark users.

Key points

  • Thesis: aviation keeps pilots able to take over from automation, not to fly without it; engineers need the same with AI.
  • Its case rests on SAFO 13002 and 17007, rule 78 FR 67800, AC 120-123, EASA SIB 2013-05R1 and EBT, which say what it claims.
  • Daily: no reading quota, echoing SAFO 17007 on flying quotas, but hand-reading of IAM, data-deleting IaC and network diffs.
  • Drills after EBT: rebuild an anonymised production incident small and drill it as a team, ungraded, with AI misses added.
  • It concedes most software failures roll back, but AI errors raise no alarm, so drills centre on faults in approved code.

Perspectives and debates

  • Hatena: most featured comments reject the analogy, as broken apps don't go manual; one would use it to get training time.
  • Aviation practises by hand at low workload and automates at high, the reverse of hand-reading only the riskiest diffs.
  • No test of takeover: Anthropic's RCT measured learning (50% with AI, 67% without); Bainbridge: unknown faults defy drills.

Technical details

  • SAFO 13002 (2013): constant autoflight doesn't maintain manual skills; fly by hand at low workload, automate at high.
  • SAFO 17007 (2017): manual flight underpins other flying skills; no blanket rules or quotas without scientific evidence.
  • AC 120-123 (2022): each pilot must be able to fly by hand, and pilots not flying ready to intervene (PF and PM).
  • NASA study (2014): motor skills held, cognitive ones slipped; 75% of pilots followed a faulty altimeter past 300 feet off.
  • CDK drill: RETAIN to DESTROY and deletionProtection: false on production RDS delete it on stack deletion or replacement.
  • Other drills: IAM "Action": "*", an ALB 502 from logs and target settings, wrong AI fixes; all AWS, no agent named.

Sources: shimitaro (Zenn) · FAA SAFO 17007 · Anthropic · Hatena Bookmark

GenAI Catch-up Report · 2026-10-13 · 🤖 Generated by Claude Opus 5.5

6️⃣ 🧩 Dynamic workflows in Claude Managed Agents: up to 1,000 agents per run

In public beta since 10-09: the session's agent writes a program that Anthropic's servers run in the background, in phases, as a workflow run.

Key points

  • Enabled by the multiagent_20261001 type, where workflows default to on, under the managed-agents-2026-04-01 beta header.
  • Per run: 64 threads at once (not guaranteed), 1,000 agents in all, a 24-hour lifetime; 10 open runs per session by default.
  • No API call starts a run: the agent decides from the message, and permission policies cover the run's tools, not its start.
  • No price of its own: tokens at each model's rates plus $0.08 per session-hour, and a run keeps the session running.
  • Anthropic on X: workflows found 66 of 70 planted bugs in a 116k-line codebase in each of three runs, one agent 14 to 27.

Perspectives and debates

  • Anthropic suggests a scoped task first, as runs can use many tokens; The Decoder calls the cost-effectiveness debatable.
  • MIXED, vanja.io: no approval step, a budget that can overshoot by 64 requests per run, stopping only by asking the agent.
  • XenoSpectrum: 1,000 is a lifetime cap, not a parallel count; the bug test leaves budgets and false positives open.

Client sends user.messageand sets the session budget Session agent writes and startsthe workflow program Server runs the program in phasesand collects the threads' results Agent threads, 64 at once,sharing sandbox and credentials workflow_run.* events to the client;at the end the agent reads the result

Technical details

  • A completed run may hide failed threads; errors include timeout_error and thread_limit_error.
  • No webhooks or run list, and the code stays hidden; user.interrupt stops the agent's turn, not a run.
  • Claude Code's GA workflows differ: a readable local script, 16 agents at once by default, a prompt on first use.

Sources: Claude Platform Docs · @ClaudeDevs · MIXED · XenoSpectrum

GenAI Catch-up Report · 2026-10-13 · 🤖 Generated by Claude Opus 5.5