AI Week: Oct 5–11 — GateKeeper

EN translation of the Russian original (issue 004, product language is Russian) — original. Names, dates, figures and links unchanged.

GateKeeperissue 004window Oct 5–11, 2026collected Oct 12, 2026

AI Week: Oct 5–11

Anthropic closed out the Claude 5.5 lineup with a budget Haiku, and Mistral shipped a trillion-parameter Large 4 — but Mistral's own materials disagree on price and architecture details. In parallel: a record AI-memory profit at Samsung coincided with an Anthropic billing outage and a hard model-sunset date at Google.

4+1

agents: collection + validation

raw items collected

in this issue

dropped or merged as duplicates

claims re-checked by a separate agent

8/3/1

confirmed as-is / revised / excluded

Three takeaways

Anthropic closed out the Claude 5.5 lineup with a budget Haiku — price dropping almost three-quarters

Haiku 5.5 completes the 5.5 lineup (alongside Opus and Sonnet): $0.10/$0.50 per million input/output tokens versus $1/$5 for Haiku 4.5, Haiku's first adjustable "effort" tier, and 72.4% on OSWorld 2.1 versus 15.7% for the previous version. All figures independently confirmed.

Mistral shipped a trillion-parameter Large 4 — but the company's own materials contradict each other

1 trillion parameters, trained on 3,800 Nvidia Grace Blackwell chips — an impressive claim to the European frontier. But the active-parameter count (49 or 52 billion?) and the real price (list versus the current 50% promo discount) disagree between Mistral's announcement and its own model card; the benchmarks are vendor claims with no independent check.

A "quiet" week that cost real money: Samsung's AI-memory record, an Anthropic outage, a hard Google deadline

Samsung posted Korea's first-ever quarterly operating profit above 100 trillion won — driven mainly by HBM memory for AI servers. In the same window, Anthropic mistakenly blocked some organizations over spend limits, and Google set a hard shutdown date for its Deep Research model. Three different flavors of "expensive infrastructure" in one window.

What breaks: deadlines

The only part of this issue that needs a calendar entry.

When	What	Who it affects
Oct 23, 2026	Gemini API retires deep-research-pro-preview-12-2025 — a hard date per Google's changelog; replacements are deep-research-preview-04-2026 / -max-04-2026.	Anyone with that specific preview model's ID hardcoded — calls start failing after the date.
already (Oct 8)	gemini-3.7-flash and gemini-3.5-flash deprecated, auto-routed to gemini-3.8-flash and gemini-3.6-flash respectively.	Anyone with those two model IDs hardcoded — behavior and latency can shift quietly with no code change.
roughly end of Oct 2026	Mistral promises open weights for Large 4 "by the end of the month" — no exact date given.	Anyone waiting for self-hosted weights instead of the API preview on Mistral Studio.
roughly later in Oct 2026	Reflection AI promises Beam's weights under Apache 2.0 "later in October" — no exact date given.	Anyone considering Beam as an open alternative to closed APIs.

Model and product releases

Oct 7Anthropic

Claude Haiku 5.5: three-quarters cheaper, and the first switchable effort tier

Haiku 5.5 completes the Claude 5.5 lineup. Price for prompts up to 100K tokens is $0.10/$0.50 per million input/output tokens versus $1/$5 for Haiku 4.5 (above 100K: $0.50/$2.50). This is the first Haiku with adjustable "effort" levels; on Anthropic's own OSWorld 2.1 benchmark it scores 72.4% versus 15.7% for Haiku 4.5.

why it mattersFor high-volume agentic and classification work, the economics shift sharply — workloads that used to need an expensive model purely because of call volume can now run on a cheap one with little loss on this benchmark.

confirmed: primary source + 3 independent outlets Anthropic

Oct 6Mistral AI

Mistral Large 4: a trillion parameters — and inconsistencies in the company's own materials

Large 4 is a multimodal MoE model with 1 trillion parameters (52 billion active per the model card, though the announcement itself said 49 billion), trained on a cluster of 3,800 Nvidia Grace Blackwell chips. Available in preview on Mistral Studio; the official price is $1.36/$4.18 per million input/output tokens, but independent reviews report a 50% promo discount currently in effect ($0.68/$2.09). Open weights are promised "by the end of the month" with no exact date.

why it mattersA claim to a trillion-parameter open-weight European frontier model — but at minimum the active-parameter count and the real price don't match between Mistral's own materials and independent reviews.

primary source exists but contradicts itself; benchmarks (Cybench 93%, AA-Briefcase 1393 Elo) are vendor claims Mistral AI Simon Willison

Oct 6OpenAI

OpenAI opened a public beta of the Decisions API

The new POST /v1/decisions endpoint runs on the GPT-6 Luna model and returns typed predicate/choice/score responses instead of free text; priced at $0.10 per million input tokens, with output and cache free. OpenAI claims roughly 10x speedup versus the Responses API — its own measurement, with no independent check.

why it mattersA new API primitive for classification and routing that competes directly with structured output and niche "decision" providers.

independently confirmed (aiweekly.co, letsdatascience.com); the 10x speedup is OpenAI's own measurement; openai.com returns 403 to automated requests developers.openai.com

Oct 7TII

TII released Falcon ASR — speech recognition built around Arabic

The Technology Innovation Institute (Abu Dhabi) published Falcon-ASR, a 1.6-billion-parameter model for Arabic (including the Emirati dialect) as well as English, French, Spanish and Portuguese. Claimed average WER is 20.92% across six Arabic test sets (versus 23.17% for the best previously published result) and 5.74% on English across seven sets, with word-level timestamps.

why it mattersA rare case of a regional (Arabic/dialect) ASR model beating published benchmarks — an area major US and Chinese labs usually don't cover.

primary source; WER figures are the authors' own claims Hugging Face

Oct 5AWS

AWS added four frontier models and OpenAI-based agents to Bedrock in one week

Amazon Bedrock gained four new models — OpenAI GPT-6.1 Sol, a speed-optimized GPT-6 Astra "UltraFast" tier (up to 300 tokens/sec per OpenAI's claim), Anthropic Claude Sonnet 5.5, and xAI Grok 4.7 — and moved "Bedrock Managed Agents powered by OpenAI" (built on a customized OpenAI Agents API) into public preview, launchable via AgentCore Runtime or self-hosting.

why it mattersAWS consolidated multi-vendor model access and OpenAI-based agent tooling onto one platform in a single week — a competitive move against Azure's and Google Cloud's model marketplaces.

primary source aws.amazon.com; speed/cost figures are OpenAI's claims as relayed by AWS AWS

Expensive and quiet

Oct 8Samsung

the week's biggest story by money

Samsung posted Korea's first-ever quarterly operating profit above 100 trillion won

Samsung Electronics' preliminary Q3 2026 guidance: operating profit of roughly 107.4 trillion won (+782.5% YoY) on sales of roughly 195 trillion won (+126.6% YoY) — the first Korean company ever to top 100 trillion won in quarterly operating profit. The growth is attributed mainly to demand and pricing for HBM memory (including for Nvidia's Vera Rubin platform); final figures land October 29.

why it mattersAI-driven memory price inflation is becoming a direct, measurable cost multiplier for anyone buying servers for training and inference — not just for the chipmakers themselves.

independently confirmed: Korea Times, Korea JoongAng Daily, Herald Business Samsung Korea Times

Oct 7Anthropic

Anthropic outage: some organizations wrongly locked out over spend limits

On October 7, some organizations were incorrectly flagged as having hit their spend limit, which blocked requests to the Claude API, Claude.ai and Claude Code; the incident was closed the same day (marked resolved at 21:23 UTC). Separately that day, there were elevated errors loading usage data on platform.claude.com. No full postmortem from Anthropic itself could be found.

why it mattersA billing glitch that mistakenly cuts off paying customers from production is exactly the kind of incident that never makes it into a newsletter but still costs downtime.

confirmed by an independent status tracker (apistatuscheck.com), not by a detailed report from Anthropic itself status.claude.com

Oct 8Google

Google deprecated three Gemini API models in one day, one with a hard shutdown date

On October 8, the Gemini API changelog got three entries: gemini-3.7-flash and gemini-3.5-flash deprecated with auto-routing to gemini-3.8-flash and gemini-3.6-flash respectively, and deep-research-pro-preview-12-2025 deprecated with a hard shutdown date of October 23, 2026 (replacements: deep-research-preview-04-2026 and deep-research-max-preview-04-2026).

why it mattersAnyone with a model ID hardcoded may see behavior and cost shift quietly from the auto-routing, and Deep Research stops working entirely after October 23.

primary source (Google's changelog); no independent outlet found confirming these three entries specifically ai.google.dev

Research and open weights

Two open model releases (Beam, the Nemotron recipe) plus a harvest of dated papers from Hugging Face Daily Papers.

Reflection AI unveiled Beam, a 501-billion-parameter open model

Reflection AI (former Google DeepMind researchers; backed by Nvidia, Sequoia, Lightspeed) announced Beam: a sparse MoE model with 501 billion parameters (23 billion active), trained on 23.8 trillion tokens, with a claimed 1M-token context (capped at a quarter of that in the live beta API). The company claims parity with Zhipu's GLM-5.2 at 3-4x lower inference compute — a claim the company itself notes hasn't been independently verified. Weights were unreleased as of October 11, promised "later in October" under Apache 2.0.

independently confirmed (letsdatascience, eesel.ai, forklog); benchmarks are vendor claims with no independent checkTechCrunch

Nvidia's Nemotron scored gold on IMO 2026 under official judging

Nvidia reports that a Nemotron-3-Ultra-based pipeline scored 30 of 42 points on IMO 2026 under official judging rules (gold threshold: 29), and separately 535.4 of 600 on IOI 2026 — above the gold threshold (361.12) and the best human score (498.27), though that run was unofficial and not reviewed by IOI's own judges. The fine-tuning recipe and data (22,000 coding problems, 414,890 SFT examples for IMO) were released openly.

the IMO result is independently confirmed (official judging + independent outlets); the IOI result is Nvidia's own unofficial claimHugging Face

vLLM 0.31.0

717 commits from 307 contributors: FlashMLA attention with NVFP4-compressed KV cache as the default for SM100 under DeepSeek-V4.1-Flash, a new weight-cache preload daemon ("vllm preload"), experimental fast restart via CRIU ("vllm snapshot"), Model Runner V2 speculative decoding, and stricter default safety settings for multimodal kwargs.

primary sourceGitHub

Cross-tokenizer on-policy distillation: from alignment coverage to supervision reliability

Strict one-to-one token matching between teacher and student already covers most of the student's tokens despite different tokenizers; reverse-KL distillation over the top-16 shared vocabulary performs as well as full-coverage distillation, while adding span-level MSE supervision for full coverage actually lowers accuracy. 210 upvotes on Hugging Face Daily Papers this week.

abstract checked directly; figures are the authors' claimsarXiv 2610.08448

CheckerBench: can agents synthesize static analyzers

An executable benchmark of 300 tasks drawn from 297 real CVEs across 167 repositories and 85 CWE categories. Across 21 model+harness configurations, average Pass@1 was 32.30%, with the best configuration reaching 45.33%.

abstract checked directly; figures are the authors' claimsarXiv 2610.07557

World Models' Last Exam in Physics

A benchmark of 40 controlled physics scenarios (mechanics, optics, fluids, heat/phase transitions, electromagnetism, surface tension) with no reference video: across eight video-generation models and 1,280 generated videos, the best score was 57.76 out of 100, with physical inconsistency persisting across all models.

abstract checked directly; authors' claimsarXiv 2610.08791

NeMo-DCR: bit-exact delta refit for trillion-scale agentic RL

Synchronizes RL policy weights by transmitting only changed parameters while guaranteeing bit-for-bit identity with a full transfer; at 3-5% of parameters changed across models from 30 billion to 1 trillion, retraining runs 12-40x faster, and relay-sync of a 1-trillion model at a 3% change shrinks from 87.5 minutes to roughly 150 seconds.

primary source; authors' affiliation (presumably Nvidia) isn't stated explicitly on the page, figures are authors' claimsarXiv 2610.08430

Making LLMs Say What They Think: the CoT-Interpretability Alignment metric

The CIA metric measures how well a model's stated chain of reasoning matches what interpretability tools actually detect inside the model; across three tasks and three models, CIA ranges from 44.8% to 75.9%, and training with a reward for "parametric honesty" meaningfully raises the score without losing accuracy. The paper's first version is from September 30 (outside the window); a substantive v2 revision from October 9 falls inside it.

primary source; figures are authors' claims; "this week" refers to the v2 revision, not the first publicationarXiv 2609.38972

Business and policy

Oct 8Manus

Manus raised over $500M — its first round since the Meta deal fell through

Butterfly Effect (Manus's parent company) announced a round of over $500M led by Boyu Capital and IDG Capital, with existing shareholders Tencent, HSG (formerly Sequoia China) and ZhenFund participating. The exact valuation wasn't disclosed (pre-round reports cited roughly $4B, unconfirmed by the company). This is the first round since April 2026, when Chinese authorities forced the reversal of Meta's roughly $2B acquisition of Manus.

why it mattersA clear signal that a major Chinese AI-agent company can raise significant capital (and keep Tencent/HSG as investors) after a forced reversal of a deal with a US tech giant.

independently confirmed: kucoin.com, aiweekly.co, webrazzi.com, vktr.com TechCrunch

Oct 8Arena

AI evaluation platform Arena raised $200M at a $3.1B valuation

Arena (formerly LMArena) raised a $200M Series B co-led by Lightspeed Venture Partners and Khosla Ventures (with Salesforce Ventures, Dell Technologies Capital, a16z, Felicis and others also participating), announced October 8; valuation roughly doubled to $3.1B. A $2.88B figure that appears in some sources, dated September 22, refers to a private secondary-market valuation via Forge Global, not the round itself, and isn't used in this issue. The company claims roughly $100M ARR as of June 2026 and launched an "Alignment Index" benchmark for unsafe/unauthorized model behavior.

why it mattersA tenfold jump in revenue and valuation in the model-evaluation/benchmarking category signals that safety-testing infrastructure is becoming its own well-funded layer of the AI stack.

primary source TechCrunch + independent outlets (winzheng.com, mezha.net, hedgeco.net) TechCrunch

Oct 6Crunchbase

AI-startup acquisitions are up 14% in 2026 — driven mostly by repeat buyers

Per Crunchbase News, from the start of 2026 through September 29, venture-backed AI startups were acquired 195 times — up 14% from all of 2025, while the number of distinct buyers grew only about 2% (the increase came from repeat buyers). The most active acquirers in 2026: OpenAI (10 deals this year, 20 over three years), Anthropic and Legora (5 each), Harvey (4), Sierra and Cursor (3 each), Cohere (2).

why it mattersA concrete, dated figure showing market consolidation: frontier labs and vertical leaders (legal, healthcare) are buying startups mainly for team and technology rather than competitors buying competitors.

primary source read directly, independently confirmed (aiweekly.co, metirai.com) Crunchbase News

What's not in this issue

Empty spots are a result too

Gemini 4, GPT-6.1 Astra, the next Meta models, Qwen4, DeepSeek V4.1 Pro, Grok 5: for each one, sources either disagree on the date or give a plain "coming soon" with no confirmed release inside the window. Dropped rather than guessed at.
EU AI Act and new US regulation: nothing new inside the window — status quo is the already-active Digital Omnibus delay and the December 2, 2026 deadlines set earlier.
No new court rulings, CEO/Chief AI Officer appointments, or new AI-adoption surveys with a confirmed date inside Oct 5–11.
Dropped for single-sourcing: a report about a delay incident with Costs data in OpenAI's Compliance API — confirmed by only one outlet (technobezz.com), and the "restored by October 11" date wasn't found anywhere else; excluded entirely rather than kept with a flag.
Dropped as outside the window (context, not an event): Broadcom's TPU-lease credit line for Anthropic (~$42B, Oct 1), reporting on Anthropic's mid-November IPO target, Mistral's €3B Series D (Sep 8), Instinct's $1B Series C (Sep 28), the unconfirmed FieldAI $700M round (Oct 2).

How this was collected

Four collector agents worked in parallel on non-overlapping clusters — vendor releases; pricing, deprecations and incidents; research and open weights; business and policy. Each was given the same rules: window October 5–11, 2026, a primary-source URL and a date required for every item, a confidence label, and no filling in gaps from the model's own memory — training-data knowledge doesn't count as evidence, only what search actually turns up.

A fifth agent re-checked 12 key claims (including everything in the Three Takeaways and the deadline table) against sources independent of the ones already cited: 8 confirmed as-is, 3 revised (Mistral Large 4 details, the Gemini deprecation dates, the Arena valuation), 1 dropped entirely as insufficiently confirmed (the OpenAI Compliance API incident).

This issue is a validation trial for GateKeeper, a service for automated monitoring of information on the web. Running the collection by hand with agents first shows which sources and which schema fields are actually needed, before any code gets written.