Fable 5.1 lands: same token price, much cheaper cache reads
Anthropic shipped Fable 5.1 to Claude and Claude Code. Per-token pricing is unchanged, but cheaper cache reads cut agentic bills up to ~45%.
LLMs
Model releases, benchmarks and capabilities. 106 stories, page 2 of 5.
Tencent drops a 770B open-weight model under Apache 2.0
Hy4 preview: 770B params, 49B active, 1M context, Apache 2.0 — and it undercuts GLM-5.3 by ~40% on price.
Grok Bot is coming to the Grok app — behind a $300 paywall
Leaked strings point to Grok Bot landing in Grok's mobile apps, plus a hidden nav tab for custom agents — Heavy tier only.
OpenAI finally ships transparent PNGs for GPT-Image-2
GPT-Image-2 now accepts background=transparent in preview — alpha-channel assets straight from the API, no cutout step.
Anthropic's 'Hub Mode' spotted — same icon as the old 'Orbit'
A 'Hub Mode' string surfaced in Anthropic's code carrying Orbit's old icon — likely a multi-agent dashboard, still unannounced.
Nous turns Hermes Desktop into a roster of named AI bots
Bot Mode ships default-on in Hermes Agent v0.20.3: each bot gets its own model, memory, skills — and can @mention the others.
Qwen hits 3B downloads — Meta and Google aren't close
Alibaba's Qwen passed 3 billion downloads and now accounts for more than half of all open-model downloads worldwide.
GLM-5.3 tops the West at finding bugs, not exploiting them
Zhipu's 743B GLM-5.3 beat Mythos 5 and GPT-5.6 Sol on CyberGym at 84.5% — then trailed both by 24 points on ExploitBench.
Gemini 3.7 Flash lands in the Gemini app, at half the price
Google's new workhorse model is now default in the Gemini app — 4 points smarter than 3.6, 40% faster than GPT-5.6 Terra.
Three labs, one week: AI pricing just split in two directions
Google halved Gemini 3.7 Flash to $0.75/M, OpenAI shipped a 14x-faster tier, and DeepSeek hiked prices 4.5x.
Grok 4.6 doesn't beat DeepSeek V4 Pro — and it costs 7x more
Both shipped Aug 12-13. Grok wins knowledge work, DeepSeek wins agents and security — at $0.87 vs $6 per million output tokens.
Grok 4.6 is cheap — until your agent crosses 200K tokens
SpaceXAI's new frontier model matches GPT-5.6 Sol on benchmarks at a fraction of the price. The fine print doubles it.
OpenAI's Codex app finally lands on Linux — with one catch
Codex now ships native builds for Ubuntu, Debian 13 and Fedora — but Computer Use is still missing from the Linux version.
Meta goes Apache 2.0 with Muse Glimmer, a 30B that fits 24GB
Meta's first no-strings open-weight model in years: 30B, agentic, beats Gemma 4 31B on coding, runs on one consumer GPU.
Meta open-sourced a 30B agent model that fits your GPU
Muse Glimmer is Apache 2.0, runs offline on a 24GB card, and beats Qwen and Gemma on multi-step tool use.
Meta ships Muse Code, and Opus 5 still beats it on paper
Meta's new terminal coding agent hits 59.3% on DeepSWE 1.1 — solid, but Claude Opus 5 leads on Meta's own charts.
Alibaba's 2.4T-param Qwen3.8-Max goes after Fable 5
Alibaba shipped its biggest model yet — 2.4T params, 1M context, frontier-tier coding scores. Open weights land next week.
DeepSeek drops V4-Flash beta — beats its own Pro on agents
DeepSeek-V4-Flash-0731 is live in API beta, outscoring V4-Pro-Preview on agentic benchmarks and adding Responses API support.
OpenAI Previews 'Astra': A Model Family of Teamed-Up Agents
Altman demoed Astra to US senators — multiple agents grinding on one hard problem for hours, not one model answering fast.
One dev vibe-coded a Homeworld clone with Opus 5 for $632
A Homeworld-style space RTS — code, 3D models and all — generated end-to-end by Claude Opus 5 for 4.6M output tokens.
Kimi K3's weights are free. The 64 GPUs to run them aren't.
Moonshot open-sourced a 2.8T-parameter model on July 27. Self-hosting it needs 64+ accelerators, so most devs will just pay the API.
GPT-Live finally lands in ChatGPT's work plans, globally
Three weeks after its consumer debut, OpenAI's full-duplex voice model is live for Business, Enterprise and Edu seats.
Moonshot Open-Sources Kimi K3 — 2.8T Params, Free to Grab
Kimi K3's weights are public: 2.8T params, 104B active, 1M context. Largest open-weight model ever shipped.
ChatGPT Pets go social: custom pets now ship as a link
OpenAI added shareable links for custom ChatGPT Pets — copy from Settings Personalisation, and anyone can adopt yours. Web only for now.