LongCat-2.5: a free 1.6T, 1M-context agent model on OpenCode
A frontier-scale MoE you can drive for free this week — plus Anthropic resumes billing some refusals and Opus 5.5 gains mid-run inline tools.

Copy markdown
LongCat-2.5-Preview is free to run on OpenCode
Meituan quietly dropped LongCat-2.5-Preview: a ~1.6T-parameter mixture-of-experts model (~48B active per token) with a native 1M-token context, tool calling, and interleaved reasoning — now free to run on OpenCode's Zen and Go tiers, with no end date announced. Pay-as-you-go promo pricing is $0.30 in / $1.20 out per million tokens. Weights are closed and there are no published benchmarks yet, but it's a frontier-scale long-context agent you can test-drive at zero cost this weekend.
Anthropic starts billing some pre-output refusals again
As of Sept 24, Claude's API resumed charging for refusals that arrive before any output — but only when stop_details.category is bio, frontier_llm, or reasoning_extraction, buckets Anthropic says carry low false-positive rates. Refusals in every other category stay unbilled, and mid-stream refusals were always billed. If your app trips those three categories, expect them on your invoice at standard model rates.
Opus 5.5 adds inline tools you can swap mid-conversation
A new beta (header inline-tools-2026-09-15) lets you define, add, or remove tools inside mid-conversation system messages via a tool_addition block — without editing the top-level tools array and, crucially, without invalidating your prompt cache. It also works with the MCP connector, so a mid-run message can attach an entire MCP toolset. Useful for agents that should unlock capabilities as a task progresses.
Cursor now auto-reviews every PR for security holes
Cursor's Security Review (Teams and Enterprise) scans each pull request for exploitable bugs — SQL, command, and template injection, auth and authorization bypasses, leaked secrets, SSRF, unsafe deserialization, and vulnerable dependencies — and flags them inline. Plans ship trial credits worth roughly 50 changes (Teams) or 500 (Enterprise) to try it before it counts against usage.