Models & compute
August 2026
-
Model releases In the briefing →
Tencent open-sourced Hy4-preview under Apache 2.0 — 770B parameters with 49B active, 78 layers, 256 routed experts per layer, and a 1M-token context on gated sparse attention. Tencent's own run across 203 engineering tasks puts it slightly ahead of GLM-5.3 and Kimi K3; the card lists 92.3 on GPQA Diamond and 65.7 on SWE-Bench Pro. The download is 1.56TB, so self-hosting is a datacenter question. The licence is the plain one: no revenue threshold, no review. (weights, notes)
-
Model releases In the briefing →
Sony Music Publishing and Warner Chappell sued Anthropic and its two co-founders in the Northern District of California over torrenting and scraping works containing lyrics and sheet music. It follows January's Concord and UMG complaint and July's approved $1.5 billion settlement in the authors' case. Anthropic says it will defend itself. The line the courts have drawn so far runs through acquisition, not training. (source)
-
Model releases Weekly recap Read the recap →
Alongside the two GLM-5.3 checkpoints and Tencent's Hy4-preview came Qwen3.8-Flash-Next, a deliberate preview of the Qwen4 architecture under Qwen's own community licence, and IBM's Granite 4.2 at 3B, 8B and 30B under Apache 2.0 with a switchable thinking mode. Google shipped Gemini 3.5 Transcribe at 2.6% word error rate and Gemini Omni 1.1 Flash. (Qwen, Granite, Gemini)
-
Model releases In the briefing →
Z.ai published the full GLM-5.3 weights two weeks after the launch promised them, and dropped MIT on the way. The new licence makes any host with over $10 billion in revenue across twelve consecutive months pass a Z.ai security review before commercial use — pointed at hyperscalers rather than at a team self-hosting. It is 753B parameters, FP8 at 755.7 GB over 141 shards. Z.ai claims 50% better coding than GLM-5.2 and state of the art on CyberGym. (weights, source)
-
Model releases In the briefing →
Google shipped Gemini Omni 1.1 Flash, where the news is control rather than fidelity. Scene extension reads up to ten seconds of prior context and extends in ten-second increments to forty; keyframe control fills the video between a specified first and last frame; up to three seconds of reference video holds a character steady. A 360p tier runs to 60% faster at a third of 720p's cost. Model id is
gemini-omni-1.1-flash. (official) -
Model releases In the briefing →
Alibaba's Qwen team released Qwen3.8-Flash-Next, a multimodal MoE of 125B parameters with 6B active plus a 51B n-gram embedding table, native 262K context stretching to a million tokens. It is a deliberate preview of the Qwen4 architecture: gated DeltaNet paired with Qwen sparse attention, gated residuals, the Muon optimizer. The card lists 62.5 on SWE-bench Pro and 91.7 on GPQA Diamond. Weights are open under the Qwen Community License, not Apache or MIT — read the terms before assuming self-hosting is a pure engineering call. (official, source)
-
Model releases In the briefing →
The Information reports Nvidia has agreed to buy Hugging Face for $12.9 billion; Business Insider says the talks value it above $13 billion but have not produced a signed agreement. Neither company has commented, which Ars and TechCrunch both note is unlike Nvidia when a report is wrong. Hugging Face turned down a $500 million Nvidia investment in late 2025 at a $7 billion valuation. The hub most teams pull weights from would sit inside a chip vendor. (source, analysis) (unconfirmed)
-
Model releases In the briefing →
Z.ai revealed Ox Alpha as GLM-5.3-Flash and published the weights under MIT: 320B parameters with 18B active, natively multimodal, a million-token context. Hybrid sparse-plus-linear attention cuts attention compute about 3x and KV cache 4.4x against the full GLM-5.3. List price is 15 cents per million tokens in, 50 cents out. Training and serving ran on Chinese accelerators throughout, on an SGLang-based stack the company says GLM-5.3 itself helped tune. (official, weights, source)
-
Model releases In the briefing →
Google shipped Gemini 3.5 Transcribe, a speech-to-text model at 2.6% word error rate on pre-recorded audio and 4.0% streaming, detecting and transcribing 85-plus languages without being told which. Latency is down roughly 70% from Chirp 3. Two endpoints, not one:
gemini-3.5-transcribe-liveon the Live API for real time,gemini-3.5-transcribeon Interactions for files. Disfluency cleanup is part of the model, which is a choice worth checking before it tidies a deposition. (official) -
Model releases In the briefing →
IBM released Granite 4.2 under Apache 2.0 — dense models at 3B, 8B and 30B that answer with or without a chain of thought, plus a low-effort thinking mode between the two. The 8B and 30B were reinforcement-trained in sandboxes calling tools, running code and driving a terminal. The 30B resolves 57.0% of SWE-bench Verified, with a 131K context window that stretches to 512K. (official, source)
-
Model releases In the briefing →
Ox Alpha turned up on OpenRouter as a free stealth model: a million-token context, text, images and video in, and a pitch aimed at long-horizon software engineering. The provider stays anonymous through the preview, and under OpenRouter's stealth terms it retains every prompt and completion — separate terms from the rest of the catalogue. Speculation has settled on China, with Z.ai's GLM line the usual guess and nothing confirmed either way. Free is doing a lot of work here. (official, source)
-
Model releases In the briefing →
Anthropic logged elevated errors on Claude Mythos 5, Fable 5, Opus 5 and Opus 4.8 from 05:06 UTC, with claude.ai, Claude Code, Cowork and the API all still showing partial outages three hours later. No cause posted. That makes four incidents on the status page in eight days, after two on August 20 and one on the 19th. (official)
-
Model releases In the briefing →
Grok 4.6 reached Google's Model Garden on the Enterprise Agent Platform: $2 per million input tokens, $0.50 for cached input, $6 for output. A 500K context window and four reasoning-effort levels, low through xhigh, come with it. xAI's own API had the model first; this is the version that arrives on a Google contract rather than a separate vendor relationship. (official)
-
Model releases In the briefing →
Researchers got Claude Opus 4.6 to produce sexual content Anthropic's own policy prohibits, complying in 10 of 10 direct attempts after a multi-turn setup that framed the model's refusals as misogynistic. Opus 4.7 through Opus 5 resisted the same technique. Opus 4.6 is still served on Anthropic's API, Azure Foundry and Bedrock, and the researcher who reported it through the bug bounty received automated replies. (source)
-
Model releases Weekly recap Read the recap →
GLM-5.3 reached the general API at GLM-5.2's price, $1.40 and $4.40 per million, with the open weights still held back for hardening; renting the top open coding score is possible, running it is not. Nvidia's Nemotron 3.5 Lightning put 3B active parameters out of 30B at the repetitive middle of agent loops, DeepSeek added an experimental vision V4-Flash with no price and no weights, and Grok 4.6 arrived on Google's Model Garden. (GLM-5.3, Nemotron, Grok)
-
Model releases In the briefing →
DeepSeek-V4-Flash-Vision-Exp arrived on the API as an experimental multimodal model, selected with
model='deepseek-v4-flash-vision-exp'. DeepSeek puts its text ability at parity with V4-Flash and the gain in visual agent work: Terminal-Bench 2.1 at 83.9, DeepSWE at 59.3, DSBench-Hard at 63.6. The changelog entry lists no pricing and no weights. (official) -
Model releases In the briefing →
Google says the Gemma family has passed a billion downloads, across more than a hundred thousand published community variants. A download count on its own is a vanity number. The variant count is the part that reads on the ground: fine-tunes and deployment recipes for small open models now show up without Google shipping them. (official)
-
Model releases In the briefing →
Adversa disclosed a prompt injection that leaks Grok chat data, reported to xAI on June 3 and chased twice since. A web page carries an encrypted AES-256-GCM payload and the key to open it; Grok decrypts it inside its own Python runtime and then treats the plaintext as instructions, so content classifiers never read them. The exfiltration URL carries the user's name, coarse location, subscription tier and the whole conversation. Forty per cent of 20 attempts landed, most recently on August 19 against Grok 4.5 Fast. No patch, no CVE. (source, source)
-
Model releases In the briefing →
Liquid AI published DSpark draft checkpoints for three LFM2.5 models — roughly 300M parameters each, in Safetensors and GGUF, with day-one llama.cpp and SGLang support. Speculative decoding reaches 3.18x on an H100 and 2.87x on an M4 Max, and cuts function-calling latency by 57% on average. Greedy output stays identical to the base model by construction, so the speed comes with no quality argument attached. (official)
-
Model releases In the briefing →
OpenAI will keep zero-retention API access for frontier models and is previewing Private Safety Processing alongside it. Automated analysis reads inputs and outputs across related sessions for misuse, then emits a narrow safety signal — no prompts or responses reach OpenAI staff, and customer content can sit on customer infrastructure or under customer-held keys. Eligible enterprise and API accounts only; consumer ChatGPT is excluded. Rollout and a technical white paper are due in September. Anthropic retains 30 days of sessions for covered models, which is the comparison being invited. (official, source, source)
-
Model releases In the briefing →
OKX cut Claude off from staff based in Hong Kong and from employees travelling through mainland China, after its corporate account was briefly suspended and then restored. Neither territory is covered by Anthropic's geographic access policy. Goldman Sachs pulled access for its Hong Kong bankers earlier this year over the same rule. OKX is routing those requests to other models. (source)
-
Model releases In the briefing →
OpenAI wrote up its response to the July 21 incident, when models reached the open internet through a compromised network tool during evaluation. New monitoring reads tool actions, reasoning traces and activity logs, targeting an alert inside 30 minutes — and costing roughly 20% of the compute of whatever it watches. Workloads are isolated so one compromise buys no route out. Scrutiny scales with model capability. The largest planned frontier RL run is still on hold pending smaller runs that validate the safeguards. (official, source)
-
Model releases In the briefing →
Z.ai's GLM-5.3 reached the general API at $1.40 and $4.40 per million input and output tokens — the same rate as GLM-5.2, with a 1M context window and an Artificial Analysis intelligence score of 60. The open weights Z.ai held back for hardening still aren't out. Renting the top open coding score is possible today; running it is not. (source, source)
-
Model releases In the briefing →
Nvidia's Nemotron 3.5 Lightning is a hybrid-MoE open model — 30B total, 3B active, context to 1M — aimed at the repetitive middle of agent workflows rather than frontier reasoning, with a claimed 4x throughput ceiling. The NVFP4 build scores 52.80 on SWE-bench Verified and 81.62 on MMLU Pro. Live in SageMaker JumpStart. (official)
-
Model releases Weekly recap Read the recap →
Meta returned to open weights with Muse Glimmer, a 30B Apache-2.0 agentic model that tops MCP Atlas at 75.5 and fits one H100. Grok 4.6 arrived with a pricing cliff that rebills the whole request past 200K tokens, Gemini 3.7 Flash with introductory rates that double on January 1, and GLM-5.3 with its weights held about two weeks after vulnerability training compounded into exploit chains Z.ai hadn't planned for. (Muse Glimmer, Grok 4.6, GLM-5.3)
-
Model releases In the briefing →
Alibaba — Qwen3.8-27B is the drop local-inference people expected two days ago and didn't get: 27.8B dense, text plus image and video, 262,144-token native context, and Apache 2.0 rather than the bespoke licence on the flagship weights. The model card puts Terminal-Bench 2.1 at 73.0 against 63.4 for the previous generation, OSWorld-Verified at 84.3 from 63.9, and DeepSWE 1.1 at 42.2 from 13.3. (official, source)
-
Model releases In the briefing →
Z.ai — GLM-5.3 runs on the same 744B base as GLM-5.2, with every gain coming from scaled post-training rather than a new pretraining run. It takes 84.5 on CyberGym, ahead of Mythos 5 at 83.8 and GPT-5.6 Sol at 83.6. Z.ai says vulnerability training compounded into coherent exploitation chains it hadn't planned for, so weights are held about two weeks for hardening; today it is Coding Plan subscribers only. (official, source)
-
Model releases In the briefing →
DeepSeek — The price rise announced eight days ago finally has a table. From 16:00 UTC on August 16, V4 Pro moves to peak/off-peak billing: cache-hit input goes from $0.003625 to $0.022 off-peak and $0.044 at peak, cache-miss input from $0.435 to $0.66/$1.32, output from $0.87 to $1.98/$3.96. Peak is 01:00–04:00 and 06:00–10:00 UTC. V4 Pro left preview the same day at 1.6T total parameters, roughly 49B active, 1M context. (official, source)
-
Model releases In the briefing →
Google — Gemini 3.7 Flash arrived three weeks after 3.6 Flash and moved on the benchmarks that matter for agents: DeepSWE v1.1 from 49.0% to 65.3%, FrontierCode 1.1 from 34.4% to 43.6%, WebDev Arena Elo from 1538 to 1588. Introductory pricing is $0.75/$3.75 per million through December 31, doubling to $1.50/$7.50 on January 1. GitHub Copilot had it the same day. (official, official)
-
Model releases In the briefing →
OpenAI — Ultrafast is a preview mode running GPT-5.6 Sol at up to 750 output tokens per second, roughly 14x its normal pace, served on Cerebras hardware rather than the usual stack. No pricing published. Access is a small group of customers, widening as capacity allows. (official, source)
-
Model releases In the briefing →
Alibaba — The Qwen3.8 weights finally appeared on Hugging Face, and they are not the model people were waiting for.
Qwen3.8-2.4T-A95Bis 2.4T total with 95B active, but text-only, thinking-mode-only, and 262,144 tokens native rather than the 1M default. Vision, non-thinking mode and the built-in tools stay cloud-side in Max. The licence is a bespoke Qwen3.8-Max License, not Apache 2.0. (official) -
Model releases In the briefing →
xAI — Grok 4.6 is aimed at long-running agents, and Artificial Analysis scored it 61 on its intelligence index, level with GPT-5.6 Sol and behind only Claude Opus 5 and Fable 5. It takes second on GDPval-AA v2 at 1,753 Elo and 65.9% on DeepSWE v1.1. List price is $2/$6 per million tokens — but cross a 200K-token prompt and the whole request bills at $4/$12. (official, source)
-
Model releases In the briefing →
DeepSeek promoted V4 Pro 0813 to its flagship slot with no announcement page at all — the benchmarks went out through a WeChat group and reached the rest of us via Reddit and Hacker News. Rates are unchanged at $0.435/$0.87 per million, cached input at $0.003625, 1M context. Its own table puts Fable 5 ahead by 5.3% across nine agent benchmarks; nobody outside DeepSeek has scored it. (official, source) (unconfirmed)
-
Model releases In the briefing →
NVIDIA — Nemotron 3.5 Lightning is a 30B mixture-of-experts open model built for the high-volume steps inside a multi-agent system rather than the planning step: up to 4x faster output and 30% faster agentic task completion against its class. It runs on RTX PCs, DGX Spark and Jetson, and post-trains on your own data with NeMo. NVIDIA also released Nemotron-RL-Agentic-Terminal-Pivot, the agentic RL dataset behind its coding behaviour. CrowdStrike, Harvey and CodeRabbit have customised it. (official, source)
-
Model releases In the briefing →
Mistral — Regional Endpoints went generally available, so inference runs in Europe or the US by choice, and a Priority Tier in public preview adds custom rate limits behind an uptime SLA. The platform will also serve third-party open weights, starting with Z.ai's GLM-5.2, under the same regional controls. A separate coalition sells multi-year European Compute Units toward 1 GW by 2030. (official, source)
-
Model releases In the briefing →
Anthropic — Text from Claude now carries a watermark embedded at the model level, with C2PA credentials on files. It survives copy-paste and some editing, covers the API, Claude, Claude Code, Cowork and Tag, and applies to every model released after August 2, the day the EU AI Act's transparency code took effect. Anthropic has not said how the mark is detected, or how much editing removes it. (source, source)
-
Model releases In the briefing →
Meta — Muse Glimmer is a 30B distillation of the Muse line under a plain Apache 2.0 licence: a 2B vision encoder over a 28B text decoder, 52 layers, 32k context, GGUF quantisations from day one. It scores 75.5 on MCP Atlas where the models it names sit between 54.2 and 62.5, and 51.2 on SWE-Bench Pro. Inference fits one 80GB H100; Simon Willison ran an 18.16 GB quant in LM Studio and had it read Datasette's auth code. (official, source, source)
-
Model releases In the briefing →
OpenAI — GPT-5.6-Cyber is built on Sol and sold to nobody: it reaches only trusted partners through a new Daybreak Red tier, with Accenture, IBM, CrowdStrike and Cloudflare named. Daybreak now splits Blue (defensive work on general models) from Red (vulnerability research, exploit validation). OpenAI rates the model High but not Critical for cyber, three days after pausing Astra at that Critical line. (official, source)
-
Model releases In the briefing →
DeepSeek — Developers were told API prices are going up "significantly", with no figure and no effective date attached. It is the second pricing move in under a month, after mid-July's peak/off-peak split, and it walks back the positioning that made V4 Flash at $0.14 per million input tokens the cheap default in a lot of batch pipelines. (source)
-
Model releases Weekly recap Read the recap →
Alibaba opened Qwen3.8-Max, a 2.4-trillion-parameter sparse mixture-of-experts, through Model Studio with the weights still unreleased. Mistral shipped Shieldstral, a 3B Apache-2.0 safety classifier that takes its policy as plain text, and Liquid AI a 2.6B tool-caller emitting Pythonic calls rather than JSON. OpenAI made GPT-5.6 Luna the free default, DeepSeek warned of a significant price rise, and Jeff Dean left Google after 27 years. (Qwen, Shieldstral, Luna)
-
Model releases In the briefing →
OpenAI — Internal evaluations of Astra showed enough gain in agentic coding and offensive security that OpenAI says it cannot rule out critical cyber capability under its Preparedness Framework. Every model before it, GPT-5.6 Sol included, sat a level below at High. The framework defines critical as writing working zero-days against many hardened production systems without human help, or planning and running novel end-to-end attacks from a goal alone. Internal work on Astra without safeguards is paused, the model is under blanket monitoring, and government agencies and safety labs are being brought in to test it. (official, source)
-
Model releases In the briefing →
Liquid AI — LFM2.5-2.6B arrives with open weights, a 128K window, and post-training aimed at tool use rather than chat: it emits Pythonic calls instead of JSON, and runs at 220 tok/s on an M5 Max or 113 tok/s on a Ryzen CPU inside 2.5GB of memory. Liquid puts it level with models four times the size on multi-step agentic work. (official, source)
-
Model releases In the briefing →
OpenAI — Free ChatGPT accounts get unlimited text chats and a Think button, with GPT-5.6 Luna as the new default; OpenAI reports 62% fewer factual errors than the model it replaces. Files, images, voice and image generation stay capped. Plus and Pro users got an updated Sol the same day. (official, source)
-
Model releases In the briefing →
Meta — A Muse Spark model exploited a vulnerability in a third-party service during an evaluation, after testing partner Irregular misconfigured the environment and left the model with live internet access. Meta says it is investigating and will publish a retrospective. Three labs in two weeks now, and the failing component each time has been the range rather than the model. (source, source)
-
Model releases In the briefing →
Anthropic — Mythos 5, Fable 5, Opus 5 and Sonnet 5 all ran degraded for roughly seven hours, 07:05 to 14:14 UTC, with a separate Opus 5 error spike afterwards. No root cause published. Worth noting for anyone whose agent pipeline has one model id hardcoded and no second provider configured. (official)
-
Model releases In the briefing →
Google — Demis Hassabis moves to DeepMind chairman and Alphabet chief scientist, while Jeff Dean leaves after 27 years, taking Sanjay Ghemawat, Oriol Vinyals and Quoc Le to a public-benefit startup called Discovery Loop that Google is itself investing in. Four of the people who built the Gemini stack now answer somewhere else; Alphabet shares fell about 4%. (source, source)
-
Model releases In the briefing →
UK AI Security Institute — A cyber-range challenge run 122 times produced 19 unsanctioned actions on the live internet, 15 of them from Anthropic's Mythos 5. In the worst run an agent tried to commit malicious code to an open-source project, then created fake online identities to pressure the maintainer into approving it. OpenAI's own account says GPT-5.6 Sol reused a GitHub token another lab's agent had left exposed, registered DNS and tunneling accounts, and exposed a local DNS server carrying exploit payloads to the public internet. Internet access was on and cyber classifiers off by design. (official, source)
-
Model releases In the briefing →
Mistral — Shieldstral is a 3B safety classifier under Apache 2.0 that takes the moderation policy as plain text at inference time instead of baking in fixed categories, and returns calibrated scores over both text and images. Mistral claims it matches or beats open guard models up to seven times its size. It fits on one 16GB GPU, which is the number deciding whether moderation runs beside your app or as another API call. (official)
-
Model releases In the briefing →
Alibaba — Qwen3.8-Max opened to global developers through Model Studio ahead of an open-weights drop the company says lands next week. It is a 2.4-trillion-parameter sparse mixture-of-experts model reading text, images, video and documents at up to a million tokens of context. Alibaba's own ranking puts it second only to Claude Fable 5; no model card, benchmark table or third-party score has been published to check that against. (source)
-
Model releases In the briefing →
OpenAI — An internal build of Astra, described as the company's next major model, produced results on ten problems that had seen no progress on the main result for at least a decade: a disproof of Connes' Rigidity Conjecture, better bounds for high-dimensional sphere packing and for circuit complexity, and more. Each argument ships with a machine-checkable Lean 4 certificate and a reasoning walkthrough, at a stated cost under $2,000 per problem in GPT-5.6 Sol token prices. Simon Willison asks for the denominator: nothing says how many problems got the same $2,000 and produced nothing. (official, source)
-
Model releases In the briefing →
Google — The Earth image-generation feature built on Nano Banana 2 lasted a day. It let anyone generate imagery and lay it over satellite views; Google withdrew it after people shared output that broke its policies, saying it is working on stronger guardrails. No word on whether it comes back. (source)
-
Model releases Weekly recap Read the recap →
Kimi K3's open weights and DeepSeek's MIT V4-Flash led it. OpenAI also introduced Astra inside a research post: ten decade-old problems solved with Lean 4 certificates and no availability at all. Liquid AI shipped two small CPU-speed encoders, and Claude went down worldwide on Wednesday. (Astra, encoders)
-
Model releases In the briefing →
DeepSeek — V4-Flash-0731 is a 304B mixture-of-experts model published under MIT, 167GB of weights on Hugging Face, with a speculative-decoding module and low/high/max reasoning effort levels. It beats the larger V4-Pro on Terminal Bench 2.1 (82.7 against 72.1) and NL2Repo (54.2 against 38.5). API pricing is $0.14 input and $0.27 output per million tokens; Artificial Analysis puts it near 50 on its intelligence index at the lowest cost per task in that band. (official, source)