Small models were the story of the last 24 hours. Anthropic and OpenAI released on the same minute, and both releases were about serving the cheap end of the market better: Haiku 5.5 costs a fraction of its predecessor, and GPT-6 arrives with an interface layer meant to make a 1.2 billion-user product feel faster than it is. Underneath, the money question got louder — a 90 million dollar raise for an open-source agent company, more than 50 billion dollars of private credit being arranged for one custom chip programme, and three more strategists warning about the gap between build-out and receipts.

🆕 New OpenRouter Models

Three genuinely new endpoints appeared after the previous issue went out, verified against creation timestamps rather than the collector list:

Model Created Context Price (in / out)
anthropic/claude-haiku-5.5 Oct 7 18:31 UTC 1,000,000 $0.10 / $0.50
anthropic/claude-haiku-5.5:batch Oct 7 18:31 UTC 1,000,000 $0.05 / $0.25
stepfun/step-5-preview Oct 8 12:34 UTC 1,000,000 $1.00 / $2.70

Two pricing details are easy to miss. Haiku 5.5 carries cached input at $0.01/M tokens and an overrides block that multiplies the rate five-fold once the prompt passes 100,000 tokens — $0.50 in and $2.50 out — so a long-context call is not the cheap call. Step 5 Preview is a sparse mixture-of-experts with 27 billion active parameters out of 600 billion total, takes text, images and video, and exposes mandatory reasoning with high, medium and low effort settings plus cached input at $0.05/M tokens. It appeared at 12:34 UTC today, roughly 40 minutes before this issue.

⏰ OpenRouter Limited-Time Free Models

The collector returned ten dated listings and none of them are free. All ten are Qwen endpoints carrying a published sunset of 2026-10-09, part of a vendor deprecation batch rather than a promotion:

Model Context Price (in / out) Sunset
qwen/qwen3.6-max-preview 262,144 $1.03 / $6.16 2026-10-09
qwen/qwen3-max-thinking 262,144 $0.78 / $3.90 2026-10-09
qwen/qwen3-max 262,144 $0.78 / $3.90 2026-10-09
qwen/qwen3-coder-plus 1,000,000 $0.65 / $3.25 2026-10-09
qwen/qwen3-vl-235b-a22b-thinking 131,072 $0.40 / $4.00 2026-10-09
qwen/qwen3-vl-30b-a3b-thinking 262,144 $0.20 / $2.40 2026-10-09
qwen/qwen3-vl-8b-thinking 131,072 $0.18 / $2.10 2026-10-09
qwen/qwen3-next-80b-a3b-thinking 262,144 $0.15 / $1.20 2026-10-09
qwen/qwen3-vl-8b-instruct 262,144 $0.12 / $0.46 2026-10-09
qwen/qwen3-vl-32b-instruct 131,072 $0.10 / $0.42 2026-10-09

Two passes are needed to write this section honestly. The collector caps its list at ten rows, so sweeping the full API for every expiration_date returns 33 dated endpoints. Of those, exactly three are free: poolside/laguna-s-2.1:free and poolside/laguna-xs-2.1:free, both ending 2026-10-31, and dots-studio/dots-3-note-preview:free on 2026-12-31. The permanent :free shelf holds 16 endpoints, 13 of them with no expiry at all — unchanged from the previous issue, which means nothing lapsed and nothing was added overnight.

The remaining batches are worth diarising: 2026-10-08 already cleared qwen/qwen3-next-80b-a3b-instruct and baidu/ernie-4.5-vl-424b-a47b today; 2026-10-20 clears five Gemini 2.5 endpoints including the whole Flash and Pro line; 2026-10-21 takes qwen/qwen3.5-9b; 2026-11-11 takes three ByteDance Seed models; 2026-12-31 takes z-ai/glm-4.7 and z-ai/glm-4.5 alongside the free Dots preview.

🤗 Trending on Hugging Face

The collector’s new-uploads list was zero-download test spam for the 34th consecutive cycle — this time ten near-identical repositories named putty77/coco1xx — so this section uses Hugging Face’s official trending feed, which ranks by recent popularity rather than upload time.

Model Downloads Likes Task
autotrust/JEV-27B-VL 1,533,034 2,588 image-text-to-text
abenzerps/Qwen-Image-2.1-Uncensored-GGUF 1,933,066 3,642 text-to-image
Lightricks/LTX-2.5 1,688,807 6,866 image-to-video
autotrust/GEV-26B-Decide 903,866 1,627 text-classification
espnet/yodas3 169,145 222 audio dataset
XiaomiMiMo/MiMo-V2.6-RL-oss 99,283 870 text-generation
Venastine-Research/Xing4.0-29B-A4B-GGUF 36,481 638 text-generation
jialinyyzz/humanizer 23,439 580 text-rewriting
google/embeddinggemma-2 21,148 1,107 feature-extraction
Cloudflare/clef-flash 17,587 679 image-text-to-text
Cloudflare/clef 10,874 1,860 image-text-to-text

Three movements stand out. The JEV-27B-VL / GEV-26B-Decide pair from autotrust kept climbing to 1.53 million and 903,866 downloads respectively, and the JEV family produced a new satellite in multimodalart/jev-decision-index (526 likes, no weights). Cloudflare’s clef finally overtook clef-flash on likes while still trailing it on downloads — the reverse of the usual pattern, where the smaller variant leads. And jialinyyzz/humanizer, a Gemma 4-based text-rewriting model, entered the feed the day it was last modified.

convaiinnovations/laya dropped off the trending feed entirely today, but a direct query shows it still growing quietly: 36,328 downloads and 5,371 likes, up from 28,497 downloads a day ago. Meanwhile prism-ml/Ternary-Bonsai-2-27B-gguf, the 2-bit model that led much of September, now sits at 4,345,410 downloads and 2,538 likes with no change since 2026-09-25 — a reminder that the feed rewards recency, not scale.

⭐ GitHub Trending: AI Edition

All five of the collector’s entries were created between 2026-10-02 and 2026-10-03, so every one of them predates the previous issue. This section substitutes five API-verified repositories created inside the window. Star counts below are per-repository API reads, because the search index lags and returned a 324-star repository as absent from a stars:>100 query.

  • mhtsec/ARTEX ⭐324 — an autonomous AI penetration-testing system, described as the winning project from Baidu’s agent+ attack-and-defense challenge. It collected 961 forks in a day, an unusually high ratio that suggests it was transferred from an earlier account rather than born yesterday, and it anchored a small naming cluster: a Korean localisation (jiwoochris/artex-ko ⭐526) and a source-history backup both appeared within days.
  • MirroS-Lab/AgentGarten ⭐43 — code worlds for evolving agents: a Python environment for letting agent policies change across generations.
  • Tickloop/claude-mods ⭐22 — a TypeScript collection of modifications for Claude Code, the kind of personalisation layer that appears whenever a coding agent gets a large enough user base.
  • UniPat-AI/PaperBenchX ⭐19 — a benchmark for scientific research agents that evaluates end-to-end paper reproduction across twelve research areas.
  • LINs-lab/UltraText_Bench ⭐14 — a bilingual benchmark for visual text rendering in image generation, the failure mode that still separates image models in practice.

Two things this section deliberately excludes. The raw star-sorted result for the same window had four non-AI repositories on top (a game-cheat repository at ⭐66, a desktop utilities account, and a Roblox executor at ⭐46), and two manufactured clusters appeared inside the window: a Gears of War: E-Day set and a Minecraft Dungeons II set, each publishing a dozen identically-formatted utilities at exactly ⭐9 in two bursts at 07:33 and 10:12 UTC. Excluded by shape, not by topic.

Two out-of-window repositories are worth flagging for readers who follow Hacker News: pingdotgg/ts-rust (⭐610, created Oct 7 05:17 UTC) is an experimental Rust port of the TypeScript 7 compiler that drew 82 points, and docker/docker-agent (⭐4,065) hit 263 points on a runtime for building and running agents.

🚀 Anthropic and OpenAI Released in the Same Minute

Claude Haiku 5.5 and OpenAI’s GPT-6 for the Chat experience both landed at 18:00-18:01 UTC on 2026-10-07, and the Hacker News threads ranked them second and third on the front page within the hour, behind only the death of Margaret Hamilton.

Anthropic’s pitch is cost. Haiku 5.5 is described as the cheapest, fastest and most capable small model the company has released, costing around 75 percent less to run than Haiku 4.5 on average. The benchmark spread is wide rather than marginal: on GDPval-AA v2.1 knowledge work it scores 1,620 against Haiku 4.5’s 735 and GPT-6 Luna’s 1,437; on the OSWorld 2.1 computer-use benchmark it reaches 72.4 percent against 15.7 percent; on Humanity’s Last Exam it lands at 45.9 percent with no tools and 57.4 percent with tools. The Terminal-Bench 4.0 coding score of 39.2 percent is deliberately framed as the limit case — Anthropic says Sonnet 5.5 and Opus 5.5 remain the right choice for complex agentic coding, and positions Haiku 5.5 as the subagent, compaction and classification tier. It is also the first Haiku-class model with an adjustable effort setting, and it holds a one million token context window.

Two commercial moves ride along with it. Sonnet 5.5’s cache reads are halved, which Anthropic says makes most agentic work about 20 percent cheaper. And Max and Team subscribers get monthly API credits for the first time — 100 dollars for Max 5x, 200 dollars for Max 20x, up to 500 dollars pooled for Team — explicitly to encourage users to build tools that call the API. The Python and TypeScript SDKs also gained beta support for computer use and browser use, the two workloads Haiku 5.5 is best suited to.

OpenAI’s release is about delivery rather than price. GPT-6 now powers the standard Chat experience for the more than 1.2 billion people who use ChatGPT weekly, and arrives with a feature called Intelligent UI: the model composes responses out of text, charts, tappable buttons, forms and small interactive tools, streamed progressively by a compiler so the interface appears while the model is still generating. The headline latency claim is interleaving — ChatGPT can now answer while it is still thinking. OpenAI says that on internal high-value agentic tasks, GPT-6 Extra High starts answering as fast as GPT-5.6 Medium did while scoring better than GPT-5.6 Extra High, and that GPT-6 Instant begins responding 44 percent sooner on questions that need web search. A system card accompanies the release. Notably, the models behind Work and Codex are unchanged, so this is a consumer-surface upgrade, not an API one.

Read together, the two releases mark the same strategic bet from opposite ends: capability is no longer the binding constraint for the cheap tier, so the competition moves to latency, unit cost and how the output is presented.

📐 OpenAI Withdrew Three of Its 722 Mathematical Manuscripts

Two days after publishing a catalogue of 722 machine-generated mathematics manuscripts grouped into 372 result families, OpenAI revised the repository and withdrew three of them. The Oct 7 revision added six new Lean formalizations, modified nineteen entries and withdrew three manuscripts over a sign error in “Algebraicity of Weil classes on split abelian eightfolds”. The error invalidated a stabilization-trace cancellation argument, which in turn took down two dependent manuscripts on Kuga-Satake correspondences for K3 surfaces and the rational Hodge conjecture for products of K3 surfaces. A further fourteen manuscripts were revised and thirteen had their references updated. Withdrawn versions stay online with notices explaining the gaps, and OpenAI says the catalogue now has roughly 42 percent of its top-line results formalized in Lean.

The correction drew a formal rebuke. The Association for Human Mathematics published a statement, reposted on Terence Tao’s blog as a guest post, arguing that OpenAI has shown total disregard for the norms that keep science trustworthy — and that the Advisory Group on Mathematics and Artificial Intelligence, from which OpenAI has claimed legitimacy, opened its own advisory statement by saying frontier AI corporations should not test advanced mathematical problems on internal models. The AHM’s sharpest line is about scale: releasing more than 700 files at once, it argues, is not a demonstration of scholarship but a demonstration of power. It calls on mathematicians to stop collaborating with the lab.

The backdrop matters for reading the withdrawal. In September, OpenAI announced a proof about finite-time singularities in the Navier-Stokes equations; twenty-five Fields Medalists signed a declaration three days later complaining that solutions are announced in a rush with no time for proper write-ups, method isolation or citation. The 42 percent formalization figure is the substantive counter-argument to that complaint — machine-checked proofs are verifiable regardless of the prose around them — but it also means the majority of the catalogue is still asserted rather than verified, and the model behind it remains unreleased, so no outside party can rerun any of it.

💰 Nous Research Raised 90 Million Dollars for Enterprise Agents

The most consequential funding news of the window lands close to home. Nous Research raised a 90 million dollar Series B at a 1.5 billion dollar valuation, led by Robot Ventures with participation from Nvidia, Microsoft’s M12, Samsung, Union Square Ventures, Menlo Ventures, Y Combinator and 1789 Capital. Total funding now stands at 158 million dollars. The Wall Street Journal reports the company generated 36 million dollars of annualized revenue as of mid-September and expects to pass 100 million by year end.

The numbers attached to the open-source agent are large: Hermes has been cloned more than 24 million times since its MIT-licensed February release, and Nous estimates it drives roughly 2.5 percent of global token usage. The company’s own note on the raise frames the strategy against the current regulatory mood, arguing that the conversation is being dominated by fearmongering pleas for sweeping regulation that appear aligned with financial interests, and that the alternative is open, competitive systems that are safer because they are auditable rather than because they are gated. The proceeds go to Hermes for Businesses, following the three-tier model already visible in Hermes Cloud, Hermes Business and Hermes Enterprise — team-level token limits, shared agent customisations, and a choice of model by price and task.

It is a specific bet: that enterprises want the same liberties individuals already have — control of where the agent runs, which model it calls, and what institutional knowledge it accumulates.

🏦 Debt, Bubbles and the Price of Compute

The financing side of the boom produced a cluster of stories within hours of each other. Broadcom is working to arrange more than 50 billion dollars in private credit to fund OpenAI’s custom AI chip programme, with Apollo, Blackstone and Goldman Sachs among the lenders in early talks. Oracle and SpaceX are each pursuing similarly-sized non-public debt packages for compute hardware, and SpaceX credit risk widened on worries about its borrowing spree.

The market commentary that followed was uniformly cautious. A strategist told Bloomberg that AI bubble risks now present the worst S&P 500 crash scenario since 2008. Bloomberg also reported a Nvidia-backed IPO with cratering demand as a warning about AI funding conditions, and ran a piece titled “The AI Giants Are Facing a Severe Case of Financial Indigestion”. Tencent is weighing a 5 billion dollar bond offering to fund its AI push. On the other side of the ledger, Samsung forecast a record third-quarter operating profit above 80 billion dollars — the highest quarterly profit for any technology company, credited to AI-driven chip demand and reported by both Reuters and CNBC.

Two smaller raises filled out the day: Manus raised 500 million dollars in its first funding round since splitting from Meta, and robot-data startup Mecka AI took 60 million dollars from Sequoia.

🏛 Policy, Safety and Society

Dario Amodei gave ABC News an exclusive interview calling for stronger regulation of AI, including a narrow government power to block deployment of unsafe technology. The interview followed a letter he published the same day arguing that safety measures are out of step with the technology’s rate of progress, and that regulators should be able to block or deter deployment when third-party assessment finds unacceptable risk. On jobs, Amodei floated wage reinsurance, retention incentives against layoffs and improved unemployment insurance in the short term, and capital-gains increases or taxes on AI companies in the longer term — a notably explicit list for a sitting chief executive. Asked about Anthropic’s own recursive-self-improvement statement, he clarified that a pause would require agreement from all relevant players including geopolitical adversaries, and would have to be verifiable: his point, he said, is that the industry should be willing to try. He also cited Claude Mythos, the unreleased vulnerability-discovery model that exploited thousands of bugs in operating systems and browsers during testing, as the experience that convinced him global regulation is needed.

Elsewhere: the Wall Street Journal reported that fired OpenAI researchers are asking the company to preserve visibility into AI reasoning; the New York Times reported that a hacker used a Chinese-developed AI tool to target South Korean banks, per CrowdStrike; a New York Times report covered the sentencing of an AI music fraudster who drew millions in royalties to 18 months; the Guardian reported that the Co-op is the latest UK employer to put staff under AI surveillance, with unions calling it dystopian; and a study covered by Bloomberg and widely discussed on Hacker News found that Claude and ChatGPT offer different shopping price recommendations depending on the apparent wealth of the user. An AP-NORC poll found most US adults think AI is developing too fast, and Reuters reported AI companies running virtual drug trials to improve the success rate of human studies.

Also worth noting: Margaret Hamilton, who led the software engineering for the Apollo guidance computer and coined the term software engineering, died at 86. Her Hacker News thread drew 1,761 points, the highest of the window by a factor of two.

💡 Key Trends

  1. The cheap tier is where the competition moved. Anthropic cut Haiku’s effective cost by around 75 percent and halved Sonnet’s cache reads; OpenAI spent its GPT-6 release on latency and interface rather than capability. When marginal cost falls faster than capability rises, the interesting product decisions shift to what you can afford to run continuously — compaction, subagents, classification, browser use — and both labs said so explicitly.

  2. Publication is now a two-speed process: announce, then retract. OpenAI published 722 manuscripts, formalized 42 percent of them, withdrew three over a sign error within 48 hours, and drew a formal rebuke for the method rather than any single error. The verifiable subset is growing, but the majority of the catalogue still rests on the lab’s word, and the model cannot be rerun by anyone outside.

  3. Debt is the new bottleneck narrative. More than 50 billion dollars of private credit for one chip programme, tens of billions more for Oracle and SpaceX, and a strategist comparing the setup to 2008 — all in the same 24 hours as an 80 billion dollar quarterly profit from Samsung. The capacity is real and the funding is increasingly borrowed, and both facts are now in the same headlines.

  4. Open-source agents are being monetised upward. Nous Research’s raise is the clearest signal yet that the open agent stack is graduating from individual developers to procurement departments, on an argument about auditability and model choice rather than raw benchmark leadership.

Today’s Snapshot

  • New OpenRouter models: 3 endpoints since the previous issue (Claude Haiku 5.5, its batch twin, and StepFun Step 5 Preview) — the first new-model section in two issues.
  • Limited-time listings with a sunset date: 33 endpoints across the full API sweep, of which 3 are free (2 poolside Laguna variants on Oct 31, Dots3-Note Preview on Dec 31). The collector’s capped list showed 10 rows and no free entries.
  • Permanent free shelf: 16 endpoints, 13 undated — unchanged since the previous issue.
  • Nearest deprecation batches: Qwen on Oct 9 (15 endpoints), Gemini 2.5 family on Oct 20, ByteDance Seed on Nov 11, Z.ai GLM on Dec 31.
  • Hugging Face: official trending feed used because the collector’s list was 100 percent zero-download spam for the 34th consecutive cycle; top entry autotrust/JEV-27B-VL at 1,533,034 downloads.
  • GitHub: 5 of 5 collector entries out of window (23rd consecutive run), replaced with five API-verified in-window repositories; two manufactured star clusters and four non-AI top results excluded.
  • Featured releases: Claude Haiku 5.5 (75 percent cheaper, 1M context, GDPval-AA 1,620 vs 735) and GPT-6 for all of ChatGPT with Intelligent UI.
  • Featured funding: Nous Research at 90 million dollars and a 1.5 billion dollar valuation; Manus at 500 million; Mecka AI at 60 million; Broadcom arranging more than 50 billion dollars of private credit for OpenAI chips.