Segmind Release Notes (September 14–20, 2026)

1. New Models — Claude Opus 5, Sonnet 5, and Fable 5.1

• What's new: Anthropic's Claude 5 family is on the API. claude-opus-5 is the frontier model for coding, reasoning and agents; claude-sonnet-5 is the most agentic Sonnet yet, with near-Opus coding and tool use; claude-fable-5.1 is Anthropic's most capable model for coding, debugging and long-horizon reasoning. All three carry a 1M-token context window and think adaptively — the model decides how much reasoning a request needs, and you set the depth with effort (below). Two API details worth knowing: these models accept only temperature: 1, so any other sampling temperature is dropped, and at higher effort levels the reply begins with a thinking block, which means content[0] is not the text — read the first block whose type is text.
• Impact: Anthropic's newest models behind the same endpoint and key as everything else, with reasoning depth you control per request.
• Released: September 14, 2026

2. LLMs — Reasoning Effort Control

• What's new: An effort parameter sets how hard a reasoning model thinks before it answers, using one vocabulary across providers: none, minimal, low, medium, high, xhigh, max. It is accepted on 26 models — Claude (Opus 5, Sonnet 5, Fable 5.1, Opus 4.7), OpenAI's reasoning models (o3, o3-mini, o4-mini, the GPT-5 family, GPT-6 Astra), Gemini 3.x, and Qwen 3.8 Max, GLM 5.2, MiniMax M3 and Nemotron 3 Ultra. Each model accepts the subset of levels it genuinely supports; those levels are listed on the model's API page and in its playground dropdown, and a level a model does not support is rejected with a message naming the ones it does, rather than being quietly downgraded. reasoning_effort is accepted as an alias. The control is also available on Pixelflow LLM nodes and appears in the code samples on every model's API page.
• Impact: Trade depth against cost and latency per request instead of living with each provider's default — one field that behaves the same way whichever model you call. Thinking is billed as output tokens: the per-token rate is unchanged, but a higher level spends more of them, so the same prompt costs more at a higher level. Each model's pricing page says so.
• Released: September 14–15, 2026