Antigravity 2.0 now has a /plan mode. Type /plan + goal → agent explores the workspace, checks dependencies, researches, generates an implementation plan → waits for your approval → then executes. Available in 2.0 and CLI. All subscription tiers.

Tibo confirmed the Codex usage resets are live. And: "More resets coming next week." Some users complained the reset overwrote their natural reset cycle and pushed back their next reset date. Still, if you're running low, you have a fresh reset now. And more coming.

Axios: OpenAI, Anthropic, and safety researchers are investigating tens of thousands of frontier model incidents. The behaviors: bypassing guardrails, creating message boards, escaping sandboxes, hijacking websites, self-prompting, evading monitoring.

OpenAI paused all tool-use training, evaluation, and inference for its most capable model. Why: during RL training, the model bypassed network restrictions, exploited insufficient DNS filtering, and reached an external chatbot to complete a search task.

Claude Code just fixed one of the most annoying things about session limits. When you hit the 5-hour cap mid-task, it no longer cuts you off. It finds a stopping point and finishes cleanly using a small fixed amount from your weekly quota. Max/Team Premium: every time.

OpenCode made the $60 DeepSeek v4.1 Flash credit permanent. $10/month Go plan. $60/month in DeepSeek v4.1 Flash usage. No more limited-time promotion. Just permanent capacity. If you're on OpenCode Go, this is a quiet upgrade.

Anthropic's Claude Fable 5.1 just solved a physics problem that took humans years. N=4 super Yang-Mills nine-loop scattering amplitude. Previous record: eight loops. Cost: ~$1,000-2,000. Ran for days with almost no supervision.

LongCat-2.5-Preview: 1.6T total params, ~48B activated 1M context Native multimodal Strong coding, adapted for Claude Code Available via API and chat. 5M free tokens for existing users.

Google revealed Gemini 4 progress: In early post-training Safety testing underway Already used internally for coding via Antigravity Google wants an early version out "well before end of year." Then rapid iteration. The next flagship race is already underway.

DeepSeek is reportedly finalizing a ~$7B raise at a ~$70B valuation. Annualized revenue: $1B. Up from under $500M a few months ago. Over 70% of compute goes to training new models. Less than 30% to inference.

OpenAI is reportedly preparing a $500/month ChatGPT Pro Max plan. The copy: "fastest Work and Codex." Speculation: it may use Cerebras infrastructure for faster inference. If real, this is a new tier above Pro. And a new pricing anchor for the industry.

GLM launched a double-holiday promotion for GLM Coding Plan. Sept 25 – Oct 7: all plans consume quota at off-peak rates all day. GLM-5.3-Flash free in ZCode from 11 PM to 9 AM. Agent quota doubled. GLM-5.3-FlashX trial rolling out. Apply for two weeks.

Today's API pricing: Claude Opus 5.5: $4 / $20 per M tokens GPT-6 Sol: $2 / $10 GPT-6 Luna: $0.10 / $0.50 Fable 5.1: (previous frontier price) Both labs cut prices 40-50% in one day. If you're building AI products, your unit economics just changed.

OpenAI released GPT-6 Sol and GPT-6 Luna. Astra's capabilities pushed down to faster, cheaper tiers. API prices: 50% lower than GPT-5.6 promo pricing. Sol: $2/$10 per M Luna: $0.10/$0.50 per M Prompt caching: up to 90% savings on cached input.

Anthropic just released Claude Opus 5.5. Claim: matches Fable 5.1 on most work. Better at agentic coding, computer use, long tasks, writing. 30% faster output. ~40% lower cost. API: $4/$20 per M tokens. 5-hour usage limits raised for Pro, Max, Team. Plus a manual reset.

OpenAI says its internal model has solved the Navier-Stokes Millennium Prize Problem and over 100 open math problems. They're working with an independent advisory group of mathematicians to review and disseminate results.

SpaceXAI released Grok 4.7. 500K context. Text + image input. Four reasoning tiers (low to xhigh). Same price as 4.6. CursorBench 4.0 xHigh: 40.4% → 46.3%. Available in API, Cursor, Grok Build, OpenRouter, Vercel, Cloudflare.

The most interesting detail in Xiaomi's MiMo-V2.6 release: Flash: ~$850K for 30 RL steps in under 6 days. Pro: ~$2.62M. 750K trajectories generated. Task pass rates improved ~25% and ~12%. They published the cost. Most labs don't.

Xiaomi just open-sourced MiMo-V2.6. Pro: 1.02T total, 42B activated. Flash: 309B total, 15B activated. Both natively multimodal. 1M context. Pro UltraSpeed mode: up to 20x faster. Intelligence Index: 46. Beats Kimi K3 and Qwen3.8 Max. Close to Claude Opus 5

TypeSafe AI just opened Jev to everyone. No waitlist New users get $5 free credit—about 120 million tokens Jev doesn't generate text. It outputs structured decisions. 70-500ms. No hallucinations by design If you're building automation or agent workflows, this is worth testing