Haiku 4.5

1. Show HN: TTFT benchmark: LLM Gateway vs. OpenRouter (Claude-haiku-4.5, 150 runs) (llmgateway.io via hn)

   +2 8w haiku

   We measured AI gateway performance with an open-source TTFT benchmark: 75 cold + 75 warm interleaved runs of claude-haiku-4.5 against LLM Gateway and OpenRouter, with phase-by-phase timings and raw data published. LLM Gateway reached first…

2. A free LLM passed every quality check in my production benchmark, then silently returned empty output 33% of the time (www.reddit.com via reddit)

   8w minimaxhaiku

   I benchmarked Claude Haiku 4.5 against 8 free/alt models (NVIDIA NIM plus a second free-tier provider) on a real task: writing outreach proposals from actual job postings, not a synthetic prompt. Same production prompts, same 3 real jobs,…

3. We need Haiku 5 pretty please! (www.reddit.com via reddit)

   8w haiku

   I think Haiku is pretty much underestimated and I feel it every day when the main agent spawns the Explore tool with Haiku 4.5 which hallucinates the F out of the codebase. Of course I could include in CLAUDE.md that a "better" model shoul…

4. A one-shot code-review benchmark: scoring restraint over recall across four Claude models (marcindudek.dev via reddit)

   9w haikusonnetopus

   I built a small benchmark to answer one question: is Claude Opus 4.6 actually worse than 4.8, or does it just feel older? It scores the half of code review that most evals ignore, which is restraint - not flagging correct code that looks s…

5. I dont understand the pricing anymore. (www.reddit.com via reddit)

   9w haikusonnet

   So I just checked my usage and I found this Session Total cost: $17.77 Total duration (API): 24m 43s Total duration (wall): 40m 54s Total code changes: 1513 lines added, 0 lines removed Usage by model: claude-haiku-4-5: 604 input, 18 outpu…

6. I benchmarked my own AXI (agent-cli) against the official ClickUp MCP server. Task success is a tie, but the context/cost gap is large. (data + code, incl. where mine loses) (www.reddit.com via reddit)

   9w haikusonnetmcp+1

   Setup: - 38 tasks - 2 Claude models (Haiku 4.5, Sonnet 5) × 5 reps, + a 1-rep Opus 4.8 probe, - same live workspace. - Deterministic state checks + an arm-blind LLM judge.

← all threads