Gemini 3.1 Flash Lite Image and Claude Sonnet 5 are now on CometAPI — fast, cost-effective image generation and editing, plus high-performance coding and agent workflows. Try them now
GLM 5.2
Z

GLM 5.2

Entrada:$1.12/M
Saída:$3.528/M
O GLM-5.2 é uma atualização significativa da Zhipu nas áreas de modelos de grande porte de código aberto e programação de IA.
GLM 5.1
Z

GLM 5.1

Entrada:$1.12/M
Saída:$3.528/M
GLM-5.1 (lançado em abril de 2026), projetado especificamente para tarefas autônomas de longo prazo. Ao contrário dos modelos tradicionais otimizados para interações curtas, o GLM-5.1 se destaca em manter o alinhamento com os objetivos, reduzir o desvio estratégico e entregar resultados em nível de produção ao longo de períodos prolongados — até 8 horas de trabalho autônomo contínuo em uma única tarefa complexa. Ele representa um grande salto na engenharia de agentes, deslocando a avaliação da inteligência de uma única interação para a execução sustentada no mundo real.
GLM 5 Turbo
Z

GLM 5 Turbo

Contexto:200k
Entrada:$0.96/M
Saída:$3.264/M
GLM-5 Turbo é um novo modelo da Z.ai, projetado para inferência rápida e desempenho robusto em ambientes orientados por agentes, como cenários OpenClaw.
GLM 5
Z

GLM 5

Entrada:$0.8/M
Saída:$3.2/M
Want this tightened or tailored for a landing page/README/press? I can adapt it. Here are sharper versions plus what info to add for credibility. Options - One‑liner: GLM‑5 is Z.ai’s open-source foundation model for end‑to‑end systems design and long‑horizon agent workflows—built for experts, production‑ready, and competitive with top closed models on large‑scale coding tasks. - 3 value bullets: - Agentic planning + iterative self‑correction for multi‑step, repo‑wide changes - Strong backend reasoning and tool use; ships for production workloads - Open weights, competitive benchmarks vs. closed leaders on code+agent tasks - 3‑sentence blurb: GLM‑5 is Z.ai’s flagship open‑source model for complex systems design and long‑horizon agent workflows. Built for expert developers, it pairs advanced planning with deep backend reasoning and iterative self‑correction, enabling full‑system construction—not just code snippets. In benchmarks and real deployments, it delivers production‑grade performance on large programming tasks, rivaling leading closed models. Make it concrete (replace placeholders and I’ll plug them in) - Model sizes/context: {N}B params, {K} tokens context, {toolformer/routers?} - Inference: {tok/s} on A100/H100, {batch size}, {throughput/latency P50/P95} - Benchmarks: SWE‑bench verified {X}%, HumanEval+ {Y}%, MBPP {Z}%, AgentBench {A}%, Repo‑level tasks {B}% (with eval setup) - Production: pass@k for PR generation, test coverage deltas, rollback rate, success on multi‑repo tasks - Ecosystem: supports {tools} (Git, shell, HTTP, DB, code indexers), {frameworks} (LangChain, LlamaIndex), license {Apache‑2.0/MIT}, safety/guardrails Copy tweaks to consider - Replace “autonomous execution” with “safe autonomous execution with guardrails (approval gates, sandboxes, timeouts)” - Avoid vague “rivaling” without numbers; pair every claim with a metric and hardware spec - Add one concrete example: “Upgraded a 120‑service monorepo from Django 3.2→4.2 in 3.1 hours wall‑clock with 92% tests passing on first run” Tell me the target (website hero, README, press note, tweet/thread), and any real metrics you can share—I’ll produce the final copy.