DeepSeek-V4.1-Flash (Max) has overtaken Kimi K3 to become the #1 open model in Code Arena: WebDev at one-tenth the cost. The update gains 52 points over its predecessor and shifts the Pareto frontier with pricing of $0.3/$1.2 per million tokens, ranking within ~10 points of GPT-6 Sol.
Token Pricing
Token 计费
Business model calculating service fees based on the count of input and output text units
Curated Intelligence
TypeSafe AI closed an $870M Series A at a $7.5B valuation, driven by its product Jev which surged in value over 24 days. Jev redefines the LLM paradigm by stripping away text generation to focus solely on binary or multiple-choice decision-making, enabling free output tokens and ultra-low input costs ($0.042/M). This 'decision-only' architecture leverages extreme cost-efficiency to disrupt lightweight model markets.
TypeSafe AI secured an $870M Series A at a $7.5B valuation. Its flagship model, Jev, is a decision-only AI that strips away generative capabilities to drastically cut inference costs, offering free outputs and ultra-low input pricing ($0.042/M tokens). This architectural shift challenges existing lightweight models by focusing purely on binary or categorical judgment tasks.
Anthropic has launched a public beta of 'Dynamic Workflows' in Claude Managed Agents, enabling lead agents to orchestrate up to 1000 parallel sub-agents via programmatic execution plans. This architecture replaces session-bound sub-agent limits with server-side phased execution, dramatically improving consistency and coverage in complex tasks like large-scale code audits.
Anthropic launched the lightweight model Claude Haiku 5.5, ranking #30 in the Code Arena WebDev leaderboard with a score of 1587. Priced identically to GPT-6 Luna ($0.10/$0.50 per 1M tokens) but scoring 6 points higher, it offers a 90% cost reduction and a 257-point performance gain compared to its predecessor, Haiku 4.5.
Developer Simon Willison released the `llm-openai-decisions` plugin, enabling access to OpenAI's new Jev-style Decisions API via the LLM CLI. The tool supports text and image inputs using the `gpt-6-luna` model for binary, multiple-choice, or scoring tasks, charging only for input tokens.
OpenAI has launched GPT-6.1 Sol, priced at just 20% of its predecessor Astra ($2/$10 per M tokens), while achieving significant score improvements on key coding and automation benchmarks like DeepSWE and AutomationBench. Positioned as 'good, cheap AND fast,' the model entered the top 5 in Agent Arena with extremely low median cost per task, marking a major breakthrough in the cost-performance frontier for large models.
Replit introduces a 'Free Mode' powered by GPT-5.6 Luna, enabling users to turn ideas into working software without worrying about token costs. This feature aims to lower the barrier to entry for AI coding tools by offering a zero-cost onboarding experience.