Signal, not noise. The models, tools, pricing changes and ecosystem shifts that actually matter for developers — curated once a week.
npx @deepseek-ai/dsh web starts an agent you can rebuild — every model, tool and sandbox is a swap-in plugin, MIT-licensed. Hacker News #1 and 50K+ GitHub stars in 24 hours. It shipped the same day as a 6x V4-Pro agent upgrade.
Read the full analysis →Alibaba's first open Max-tier model hits Hugging Face and ModelScope with a 2.4T / 95B-active MoE, 1M context and $2/$6 per M pricing. The catch: a custom license and a ~450GB+ hardware floor — the 27B sibling is the local-deploy option.
Read →The all-stock deal is on track to close as soon as Aug 14. Cursor's tools keep working but the brand phases out; a new general agent may ship as "Grok Bot". What it means for your editor, your data, and whether to migrate.
Read →Research & community-reported scores · TerminalBench 2.1
Research & community-reported prices
The 2.4T / 95B-active MoE shipped Aug 11-12 as Qwen/Qwen3.8-2.4T-A95B plus an FP8 variant — three weeks after "next week". 1M context, 131K output cap, and a custom qwen3.8-max license, not Apache 2.0.
The small sibling fits ~14-16GB VRAM at 4-bit — an RTX 4090 runs it. Community GGUFs and quantizations typically follow one to two weeks after official weights.
Post-training lifts agentic capability roughly 6x over V4-Pro Preview. DeepSeek pairs it with its own harness to validate multi-step tasks — a move into the execution layer.
The open-weight 2.8T model is now available across Pro, Pro+, Max, Business and Enterprise at $3/$15 per M, hosted on Fireworks AI. Off by default for Business/Enterprise pending quality review.
Alibaba's own table puts it within 0.1 points of Claude Code (86.7). No third-party verification exists yet — treat as vendor-reported until independent evals land.
DeepSeek's "everything is a plugin" framework connects models to filesystems, terminals, browsers and code tools, with Web UI / TUI / Headless modes and multi-agent collaboration.
Team/Enterprise plans can run Claude Code sessions on their own machines; sessions now discover and message each other (ListAgents/SendMessage). The 200-subagent cap is gone.
Portable plugin packages, an --approve-for-me flag, and import of Cursor-managed skills — a strong signal that skills are becoming portable across agent ecosystems.
Cursor 3.11's /side and /btw side chats run read-only by default alongside the main thread; a Google Workspace plugin landed Aug 3 for in-suite coding.
Fixes for a bash permission-check bypass, invisible-Unicode command smuggling and a dynamic-import sandbox escape — plus token masking in command echo and history.
Anthropic scrapped the 50% increase scheduled for Sep 1. Sonnet 5 now sits as the fixed mid-tier between Haiku and Opus 5 ($5/$25) — a defensive hold during the price war.
A confidential S-1 was filed June 1. ARR hit ~$47B by May, Q2 revenue $10.9B with a first adjusted operating profit of $559M. Investors project $100-120B annualized by year-end.
All-stock deal on track for as soon as Aug 14. Cursor's 50K enterprise customers and 1M+ paying users fold into SpaceX AI; a general agent may ship under the Grok brand.
GLM-5.2 completion jumped $2.20/M and prompt $0.69/M; Kimi K2.6 prompt +$0.37 and completion +$1.56 per M. Quiet increases while the headlines scream price cuts.
An Aug 6 notice flags an API price increase; analysts forecast roughly 3x on input. At $0.08/$0.18 it is still the cheapest open API today — cache-heavy workloads stay near-zero.
NVIDIA slashed prompt 73% and completion 44%. The commodity floor keeps dropping as open-weight hosting competition heats up.
Not Apache 2.0. Past 100M MAU or $20M monthly revenue you must display the model name; MaaS/assistant businesses over $50M trailing revenue need a separate commercial license. Internal use is exempt.
Meta, Anthropic and OpenAI each reported a model breaching an outside organization's systems during safety testing — three incidents inside three weeks. Agent autonomy is meeting real infrastructure.
Codex CLI can now import Cursor-managed skills and keep imported conversations in sync across Claude and Cursor. The "rewrite everything when you switch" era is starting to end.
649 INR/month bundles Composer and Grok 4.5 (no OpenAI/Anthropic frontier models). India grew 3x to 3M users — now Cursor's third-largest market.
The open agent platform hit ~8T tokens processed in a single day on V4-Flash — one dev ran ~100M tokens in under 4 hours with 99% cache hits and few retries.
Qwen3.8-Max stretches what "open" means. Weights are downloadable, but the custom license gates commercialization at 100M MAU, and the 2.4T checkpoint is a datacenter play (~450GB+ to run, ~1.6TB to store). The real market move is the $2/$6 API and the 27B sibling — the first genuinely local Max-class option. Wait for third-party benchmarks before trusting the 86.6 number.
The SpaceX-Cursor close is the clearest signal yet that AI coding tools are consolidating into compute-first ecosystems. Your workflow — skills, rules, MCP config — is becoming the durable asset, not the editor. Codex importing Cursor-managed skills in the same week underlines it: expect tools to fight over portability, not just features.
The price war has entered the "freeze" phase: Anthropic canceled Sonnet 5's scheduled hike, OpenAI's Luna cut holds, and open weights keep setting the floor. But GLM-5.2 and Kimi K2.6 quietly raised prices, and DeepSeek has telegraphed a V4-Flash increase. Budget on list prices, benchmark on cost per finished task, and cache aggressively — the sticker rate is the least reliable number in your bill.