Every story, at length

Claude Haiku 5.5 replaces Haiku 4.5, which was nearly a year old and priced at $1/million input and $5/million output — 10x the cost of OpenAI's GPT-6 Luna. The new model matches Luna's $0.10/$0.50 pricing up to 100,000 tokens, then jumps 5x to $0.50/$2.50. Luna's own price increase kicks in at 272,000 tokens, but only to $0.20/$0.75. There's also a hidden cost: Haiku 5.5 uses a less generous tokenizer, with the same prompt consuming roughly 1.25x as many tokens as Haiku 4.5.

Willison's testing shows the model no longer allows disabling reasoning and defaults to medium effort. A low-effort pelican SVG cost 0.0936 cents and took 7 seconds; a max-effort version took 5 minutes 9 seconds and cost 3.3826 cents. The reasoning trace for the max-effort run begins with "This is the classic pelican-on-bicycle SVG test..." — indicating awareness of the benchmark. For workloads under 100,000 tokens, Haiku 5.5 reports higher benchmark scores than Luna at the same price.

Separately, Anthropic announced it is halving cache read prices for Sonnet 5.5 and adding monthly API credits to subscription plans. Max 5x users get $100 per month, Max 20x users get $200, and Team subscribers receive up to $500 pooled across users. The credits match the subscription cost, do not roll over, and can be claimed via Settings → Billing. Anthropic also added the ability to disable API auto-reload, so requests stop when the balance runs out — useful for avoiding surprise charges while burning credits. OpenAI still allows Codex subscribers to use their subscription for personal API use, which remains a better deal for heavy API users.

NVIDIA reports two competition results built from Nemotron 3: a Nemotron-3-Ultra-CC specialist scored 535.4 out of 600 at IOI 2026, above the 361.12 gold threshold and the top human score of 498.27, while an IMO system combining Nemotron 3 Ultra general, SFT, and RL checkpoints scored 30 out of 42, above the official gold threshold of 29. The IOI run was live and prospective under the same time, internet-access, and submission constraints as human contestants, though unofficial and outside the official ranking. The IMO proofs were graded by official IMO graders.

The specialization recipe is deliberately unremarkable: start from a strong Nemotron base, curate domain problems and reasoning traces, apply SFT and RL, then pair the model with an inference loop that generates, evaluates, and refines candidates. For competitive programming, the team curated 22,000 problems and trained two specialists: Nemotron-3-Nano-CC (30B total, 3B active parameters) with SFT and RL, and Nemotron-3-Ultra-CC (550B total, 55B active parameters) with SFT only. On IOI 2025, Nano moved from 130 points pre-training to 280 after SFT and 291 after RL, then to 468 with GenCorrect, crossing the 438.3 gold threshold. Ultra-CC reached 502 points with the same test-time strategy. One SFT epoch on Ultra was enough to beat the fully post-trained Nano across IOI, ICPC, and LiveCodeBench Pro.

For IMO, the SFT corpus contained 414,890 quality-filtered examples across 15,818 unique proof problems, covering proof generation, refinement, verification, and meta-verification. The RL model trained on 9,597 proof problems near the model's capability frontier. The final system used both specialists plus the general model, generating candidate proofs, scoring them, producing critiques, and refining the most promising attempts, with a separate high-compute stage selecting the final submission. It worked entirely in natural language with no formal prover, external tools, or internet access, and earned full credit on four of six problems.

The key finding is that fine-tuning and test-time compute compound: better specialization gives the inference loop better candidates, critics, and refinements. At IMO, complementary SFT and RL checkpoints outperformed drawing more samples from a single checkpoint. Assets released include the Nemotron Labs IMO 2026 collection with SFT and RL checkpoints, both training datasets, Nemotron-IMO-Bench (200 olympiad-level problems), the Nemotron-3-Ultra-CC model, and inference pipelines in NeMo-Skills.

The release covers two models: d1-3B, trained from the decoder-only LFM2.5-VL-3B backbone with text and image inputs, and d1-omni-600M, built on the bidirectional LFM2.5-Encoder-350M with added vision and audio encoders for text-plus-image or text-plus-audio inputs. The 600M model is flagged as an early research release. Both are open-weight on Hugging Face and ship their own code, requiring transformers>=5.14 and trust_remote_code=True.

On the Decision Index 0.2.1, d1-3B scores 48.57, ahead of every 4B and 9B model and of Decider 35B-A3B at 47.11. Across seven public datasets—SQuAD 2.0, Civil Comments, MASSIVE intent, PubMedQA, BoolQ, XNLI, and PAWS-X—d1-3B posts a mean of 82.9, above Decider 4B at 81.1. d1-omni-600M reaches 78.4, surpassing Decider 2B's 77.1 with a quarter of the parameters. Liquid AI does not report vision or audio benchmarks, noting the Decision Index v0.3 includes only a private vision split and audio decision benchmarks remain an open problem.

Latency numbers, evaluated with NVIDIA, show d1-3B answering a single question in 16 ms on Jetson AGX Thor, 26 ms on Jetson AGX Orin 64 GB, and 50 ms on Jetson Orin Nano. Three questions take only 1.3x the time of one on the AGX Thor, going from 16 ms to 20 ms. On GPU, a single question runs in 8 ms on an RTX 4090 and 9 ms on an AMD MI325X. The API supports named questions over one text state, image-as-state, and packed batching with no padding via system_one and system_one_batch.

The paper targets a specific failure mode in conditional memory architectures such as DeepSeek Engram: different phrasings of the same fact activate different n-gram embeddings, while editing shared embeddings risks corrupting unrelated knowledge. EngramEdit addresses both by computing target memory representations that make the model predict the updated fact across multiple expressions, then jointly updating shared n-gram embeddings to match those targets. Updates to frequently reused embeddings are penalized more strongly, which preserves unrelated predictions.

On the evaluation side, the method achieves near-perfect editing success. Revised knowledge transfers to unseen expressions and to multi-hop reasoning, where EngramEdit reaches nearly three times the strongest baseline's accuracy under chain-of-thought prompting. The authors also report that unrelated knowledge and general capabilities remain largely preserved as factual updates accumulate, which is the property that makes this more than a one-off edit trick.

The contribution is architectural rather than a new model release. Conditional memory was previously framed as a capacity-scaling mechanism; EngramEdit extends that role to decoupled knowledge updates, keeping the Transformer backbone fixed. The key technical move is the joint optimization across expressions and edits with reuse-aware regularization, which directly targets the interference problem that made naive embedding edits unsafe.

No model sizes, training costs, or code release details appear in the abstract, so practitioners should treat the reported results as proof of concept until the full paper is available.

At a San Francisco Tech Week event on Wednesday, Microsoft detailed two RTX Spark-based machines: the Surface Laptop Ultra and the Surface RTX Spark Dev Box. The laptop comes in two base models, starting at $2,600 and $3,700 respectively, with memory and storage options pushing the price up to $5,900 — a configuration Microsoft says is already out of stock. The Dev Box workstation starts at $6,000 and ships with VS Code, GitHub Copilot CLI, WSL, and PowerShell 7.

The hardware pitch is local, free on-device inference: CPU, GPU, unified memory, and upgraded cooling tuned to run AI models without cloud metering. Both devices run a revamped Windows 11 that introduces Execution Containers, a feature Satya Nadella said will be available to all Windows 11 users, not just buyers of the new hardware. Nadella framed the shift as moving beyond models to orchestration: "You really do need to orchestrate, and you need to have memory outside of the model," with a harnessed layer combining multiple models, context, memory, and an action space — for third-party agents as well as Microsoft's own apps.

Microsoft is targeting developers directly, offering up to $1,000 off for MacBook Pro trade-ins, an explicit nod to Apple's foothold among AI developers in Silicon Valley. The company also claims the GPUs make the machines viable for content creation, video processing, and gaming.

Dell moved in parallel the same day, announcing specs and a $3,800 price for its previously revealed Dell XPS 16 Creator Edition, available for preorder now with late-October delivery at Best Buy.

The feature, called Intelligent UI, begins rolling out Wednesday and integrates visuals directly into conversation threads. OpenAI demonstrated the system generating diagrams for how an airplane wing produces lift, visuals for recipes, a bicycle mechanics diagram, a multi-day hiking map, and a personal savings calculator. Product manager Aarush Selvan framed the shift as a response to the limits of text: "ChatGPT has predominantly been a text-based interface," he said, but "the most helpful answers aren't just text."

What distinguishes this from earlier image generation is interactivity. Rather than static illustrations, the interface produces tappable buttons, task-specific calculators, interactive charts, and editable graphs that users can manipulate in place. OpenAI's stated goal is to make "learning complex topics easier," though the company did not explicitly claim the change is aimed at broadening its user base.

The rollout is tied to the GPT-6 model release. Intelligent UI goes live globally on Wednesday for Pro, Plus, Business, and Enterprise tiers, with free and Go tier users getting access Thursday. Users can dial back the visual density through the same customization mechanisms that control other ChatGPT personality settings, so teams that prefer text-dominant responses are not locked into the new format.

For practitioners, the notable shift is architectural rather than cosmetic: responses are becoming structured, interactive artifacts rather than prose. That has implications for how applications consume ChatGPT output, how users are onboarded to complex material, and how much UI logic now lives inside the model's generation layer instead of in the client.

The new site accepts a broad set of formats: JPG, JPEG, PNG, BMP, WEBP, AVIF, HEIC, HEIF, TIFF, TIF, and GIF for images; MP4, MOV, and WEBM for video; and WAV, MP3, OGG, FLAC, AAC, and M4A for audio. Google first introduced SynthID in 2023 and had previously restricted access to journalists, media professionals, and researchers at Google I/O last year. The public launch removes that gate.

SynthID watermarking is already embedded in Google's own generation stack: Nano Banana, Veo, and Lyria models, plus tools with generation capabilities including Gemini, Flow, ProducerAI, and Vids. OpenAI, Nvidia, and Kakao also support SynthID, and OpenAI maintains its own separate verification site. Apple is reported to be adding support soon. Google has also integrated verification directly into the Gemini app and Chrome.

Google states that people currently make 1 million verification requests per day. That figure predates the public launch and reflects usage through the existing Gemini and Chrome integrations. Microsoft and Meta run their own watermarking and verification standards, but the source notes these tools are not infallible and often fail to identify content created by their own makers' models.

The public site is an incremental distribution change rather than a new detection capability. The underlying SynthID technology has been in production since 2023, and the verification path was already available through Gemini and Chrome. What changes is friction: anyone can now check a file without using a Google product or holding a partner account.

Announced Wednesday from Google Labs, Playground lets users build browser-based games from natural-language prompts without coding. Creators pick a genre such as trivia or racing, or start from scratch, then specify 2D or 3D, single-player or multiplayer, and describe mechanics, gameplay, and visual style. The platform can also ingest uploaded visuals and convert them into game assets matched to the project's art style. Finished games run on mobile and desktop, can stay private, be shared by link, or be published to the Explore gallery, where leaderboards are supported and creators can keep modifying released projects.

Google says Unity Spark integration is planned for a future update, which would add 3D capabilities and what the company calls professional-level mechanics through Unity's technology. Access is currently limited to users 18 and older in the U.S., with expansion to other countries planned. Browsing and playing the catalog is free; game generation runs on a weekly token system, with a limited free tier and higher limits for Google One AI subscribers.

The launch places Playground in a small but growing field of AI-assisted game tools. Roblox announced its own Build feature in July, which also uses natural-language prompts for game creation. Google's move also revives its gaming ambitions after Stadia shut down in 2022; the company has since folded games into other surfaces, including YouTube Playables, a hub of instantly playable web-based games.

For practitioners, Playground is an experiment rather than a production toolchain. The token-gated generation, US-only availability, and browser-game scope suggest Google is testing demand and creative workflows more than courting professional developers. The planned Unity Spark integration is the detail worth watching, since it could extend the platform beyond casual browser games into more capable 3D territory.

The round was led by Robot Ventures, with participation from Nvidia, Union Square Ventures, Menlo Ventures, Samsung, and 1789 Capital, where Donald Trump Jr. is a partner. The financing brings the three-year-old startup's total funding to $158 million and confirms earlier TechCrunch reporting on the valuation.

The open source Hermes Agent has been cloned more than 24 million times and drives roughly 2.5% of global AI token usage, according to the startup's estimates. That developer traction is now being leveraged for a commercial push: the new capital will fund "Hermes for Businesses," which lets companies deploy customized AI agents that handle multi-step workflows while keeping data private and secure.

On the revenue side, Nous was at roughly $36 million in annualized revenue by mid-September 2026 and expects to pass $100 million before the end of 2026, The Wall Street Journal reported. The enterprise product is the mechanism for that jump.

For practitioners, the signal is less about the valuation than about distribution: an open source agent framework with significant token-share is being productized for enterprise deployment, with privacy and multi-step workflow execution as the stated differentiators.

Meta announced Wednesday that it took action against 33.2 million pieces of child sexual exploitation content on Facebook and Instagram in the first half of 2026. More than 97% of that content was found by its systems before users reported it. In India specifically, Meta acted on 5.3 million pieces during the same period, with over 98% detected proactively.

The core technical shift is a new large language model system trained to detect what Meta calls 'signposting' — ads that appear benign but direct users to illegal content hosted outside Meta's platforms. The company said this is a tactic bad actors have recently adopted as they adapt to evade detection. Because the ads themselves contain no illegal material, Meta is now evaluating where an ad sends users, not just what the ad contains. That destination signal lets Meta block offending websites and take action against the accounts behind them.

Meta is also deploying additional AI-driven scans to surface child exploitation content that earlier systems missed, and it will continue adding new signals as it learns how these networks operate. A separate 'red-teaming AI agent' probes Meta's own safety measures for weaknesses that bad actors could exploit, aiming to surface novel abuse methods before they scale. The company is also improving detection of recidivist users who return with new accounts after prior removals.

The announcement lands amid sustained legal and regulatory pressure. In August, Meta agreed to pay up to $18 billion to settle a child safety lawsuit involving 29 U.S. states. Earlier this year it shipped parental controls for Meta AI, preteen accounts on WhatsApp, and alerts for parents when children search Instagram for self-harm content. In September, WhatsApp added further parental controls over Channels, status visibility, group additions, and group activity notifications.

Speaking at MIT Future Fest, Fadell projected images of three discontinued devices — the Rabbit R1, the Humane Ai pin, and the Limitless pendant — and said their makers approached him for help. He declined. His diagnosis: none of them addressed a real pain point. 'It was just interesting technology for geeks,' he said, 'but it doesn't really apply to my life.'

The deeper problem is the framing. Fadell put the share of the world population that has ever employed a human assistant at less than 0.01%, so pitching an AI 'assistant' presumes a mental model most consumers do not have. Even for those who do, trust is built incrementally: he said it took him a couple of years to learn how to use an assistant and then to trust one with sensitive data, banking, and meeting scheduling. He contrasted that with Meta's Muse launch, where a security researcher quickly found a serious vulnerability and 404 Media reported employees discovered security issues that triggered a pre-launch 'mad dash' to fix them.

On architecture, Fadell is explicit: a successful agent will operate on-device only, for privacy and to keep the tech lightweight. He dismissed the idea that data centers will dominate, citing the compute already available on battery-powered devices. Apple, he said, is the only company he can see doing this today — 'maybe one other' — because it has the hardware, chips, and consumer trust built through features like Face ID, but it lacks a world-class proprietary AI model; the new Siri AI runs on custom-built versions of Google's Gemini. He theorizes Meta and OpenAI are moving into gadgets precisely because they lack access to phone sensors and want a device that can collect video, audio, GPS, and more without per-permission friction.

Fadell also drew a line between startups and incumbents on product risk. Asked about luck in finding product-market fit, he said a startup 'gets one shot,' adding: 'It's not like Apple with the Vision Pro.'

The financing splits into an $8 million seed co-led by Sequoia Capital and First Round Capital and a $30 million Series A led by Hummingbird Ventures, with no valuation disclosed. Founded in South Africa in 2022 by siblings Jemima and Josiah Meyer, the company began with a clinical nutrition tool for dietitians before pivoting to a general-purpose inpatient screening platform.

Healthleap plugs into a hospital's electronic health record system and applies language models to written notes, extracting affirmative or negated mentions of clinical concepts such as poor appetite, recent weight loss, muscle loss, and trouble swallowing. That output feeds risk models alongside structured data — labs, vitals, weights, medications, diet orders, and diagnoses — to produce a nightly risk score for every adult inpatient, delivered into the care team's existing workflow each morning. The software flags patients for review rather than diagnosing them.

The platform is deployed in more than 50 hospitals, up from three partners a year ago, with named customers including Penn Medicine, Cedars-Sinai, Intermountain, Houston Methodist, and Emory Healthcare. Revenue grew more than 10x over that period, though specifics were not disclosed. Contracts run three years, priced by licensed bed count, with an outcome-based component: the company claims every customer has seen at least 5x hard ROI, and in some cases over 20x annual total ROI. At the Hospital of the University of Pennsylvania, Healthleap attributes $23.8 million in annualized financial impact to its malnutrition program — $6.3 million from additional reimbursement and $17.5 million from shorter stays.

Beyond malnutrition and delirium, the startup has built programs for aspiration pneumonia, pressure ulcers, and congestive heart failure readmission risk, which are undergoing further clinical validation. The new capital will go toward engineering, product, sales, and customer success, with a stated goal of covering more than 40 major conditions and expanding into outpatient and home care.