Terminal Bench

Evidence (10)

• 7The New Stack·9/28/2026model_release
  Anthropic adds cyber safeguards and fallback routing to Claude Sonnet 5.5

  Anthropic released Claude Sonnet 5.5 with the same style of cyber safeguards and model fallback routing previously reserved for its most capable models. The company says the model is stronger at offensive security tasks than Sonnet 5, enough that blocked cyber requests may be rerouted to Sonnet 5, while API developers must opt in to fallback behavior.

• 7The New Stack·9/21/2026model_release
  Grok 4.7 improves long-horizon coding, but still lags top rivals

  SpaceXAI released Grok 4.7 with training focused on long-running, multi-hour tasks, emphasizing self-verification and long-context management. The model improved on endurance benchmarks such as Terminal-Bench 4.0 and CursorBench 4.0, but still trails Anthropic’s Claude Fable 5.1 on Terminal-Bench and remains unreliable in many long coding workflows.

• 6Hacker News·9/11/2026research
  Benchmark finds RTK token savings do not reliably reduce AI coding cost

  A benchmark analysis tested RTK, a terminal-output compression tool for AI coding agents, and found that large reported token reductions did not consistently translate into lower spend. Results varied by model and task: one setup saw a small cost improvement, while another became more expensive overall. The study concluded that less terminal output can help, do nothing, or even backfire depending on task structure and agent behavior.

• 6Hacker News·9/8/2026research
  Qwen3.8 27B 4-bit quantization keeps benchmark quality

  A benchmark comparison of Qwen3.8 27B quantizations found that the 4-bit Q4_K_M model performs essentially the same as BF16 on GPQA Diamond, IFBench, and Terminal-Bench 2.1 while using far less GPU memory. Performance drops sharply at very aggressive compression, with 1-bit quantization collapsing to near-random results on GPQA Diamond and degrading reasoning further on longer tasks.

• 6Hacker News·9/2/2026model_release
  Multiverse launches Quasar 438B reasoning model for agents

  Multiverse Computing has launched Quasar 438B, a compressed 438-billion-parameter reasoning model aimed at coding and enterprise agents. The company says the model is fast and efficient, but key details about its compression method, hardware requirements, and real-world agent performance remain unclear.

• 7Hacker News·8/12/2026model_release
  SpaceXAI releases Grok 4.6 with longer-task coding gains

  SpaceXAI released Grok 4.6 less than a month after Grok 4.5, highlighting training that emphasizes self-checking, error correction and sustained progress on longer coding and agent tasks. Benchmark results show clear gains over Grok 4.5, but Grok 4.6 still trails the top models on several coding and terminal tests.

• 7Hacker News·7/21/2026model_release
  Poolside releases Laguna S 2.1 coding model

  Poolside has released Laguna S 2.1, a 118B-parameter Mixture-of-Experts coding model with 8B active parameters per token and support for up to a 1M-token context window. The company says it is optimized for longer-horizon agentic coding work and reports strong results on benchmarks such as Terminal-Bench 2.1, SWE-Bench Multilingual, SWE-Bench Pro, DeepSWE, SWE Atlas, and Toolathlon Verified.

• 4Hacker News·7/19/2026research
  How one researcher cut AI agent token costs with shared subscriptions

  A Quesma researcher describes building a deep-research pipeline for studying AI agent economics after an initial run exhausted a Claude Max plan in 30 minutes. The revised setup uses Claude Code as the main harness, with Codex and Antigravity added as headless subagents sharing memory through claude-mem, and cheaper models assigned to specific roles to reduce cost and improve verification.

• 7The New Stack·7/9/2026framework_update
  DevRev releases an enterprise AI agent benchmark

  DevRev introduced an open Enterprise AI Agent Benchmark to better measure how AI agents perform on real enterprise work, especially across large, permissioned data contexts. The first release covers only the L1 and L2 tiers and includes the dataset, evaluation harness, judging criteria, results, and raw traces so others can reproduce the tests.

• 8Hugging Face Blog·6/17/2026model_release
  GLM-5.2 launches with 1M-token long-horizon coding support

  Z.AI introduced GLM-5.2, its latest flagship model for long-horizon tasks, with a stable 1M-token context and improved coding performance. The release also adds flexible effort levels, an IndexShare architecture to cut compute cost, and an MIT open-source license.