Tutorials, deep dives and product notes — built for developers.
Hy3 (295B MoE, Apache 2.0, $0.80/1M) vs GPT-5.5 (proprietary, $30/1M). GPT-5.5 leads coding (+11-42 pts), but Hy3 fights back on agents: wins MCP Atlas (+3.8), edges HLE w/tools (+1.0), near-ties BrowseComp (84.2 vs 84.4). All at 1/37th the cost. 5 charts, full breakdown.
Hy3 (295B MoE, Apache 2.0, $0.80/1M) vs Claude Sonnet 5 (proprietary, $10/1M). Sonnet leads every shared benchmark (+0.5 to +8.7 pts). But Hy3 ties on BrowseComp (84.2 vs 84.7), leads MCP Atlas (79.1%), costs 12.5x less. Open-weight agent vs proprietary coder — 5 charts, 10-point verdict.
Tencent's 295B MoE Hy3 just took the fight to DeepSeek's 1.6T V4 Pro — and won on 12 of 18 shared benchmarks. Pricing is close: Hy3 cheaper on fresh input/output, V4 Pro's disk caching is 16.5× cheaper on repeated contexts. Full breakdown.
Hy3 (295B MoE, Apache 2.0, $0.80/1M) vs GLM 5.2 (753B MoE, MIT, $4.40/1M). GLM 5.2 wins every coding benchmark by 4-18 points. Hy3 counters with MCP Atlas #1 open-weight (79.1%), BrowseComp 84.2%, DeepSearchQA 91.0%, 47% fewer tokens, and 5.5× cheaper. Full comparison with 5 charts and a 10-point verdict.
Claude Sonnet 5 vs DeepSeek V4 Pro: Sonnet leads every coding benchmark (+7.8 Pro, +9.2 HLE tools). DeepSeek is #1 global on LiveCodeBench (93.5%), MIT open-weight, and 7.8× cheaper per task ($0.12 vs $0.90). Is 7.8 more Pro points worth 7.8× the cost?
Claude Sonnet 5 vs GLM 5.2: near-ties on every benchmark (±0.6-2.7 pts). GLM 3.4x cheaper on output, MIT open-weight, self-hostable. Sonnet has OSWorld, BrowseComp, Anthropic safety ecosystem. Proprietary premium vs open-weight value.
GLM-5.2 (62.1% Pro, $4.40/1M) vs DeepSeek V4 Pro (55.4%, $0.87/1M). GLM leads all shared benchmarks (+6.7 Pro, +6.5 HLE, +3.4 MCP). But DeepSeek dominates competitive coding: LiveCodeBench 93.5% (#1 global), Codeforces 3206, GPQA 90.1%. Both MIT, both 1M context. Full comparison.
GLM-5.2 (62.1% Pro, MIT, $4.40/1M) vs MiniMax M3 (59.0%, open-weight, $1.20/1M). GLM leads all shared benchmarks (+3.1 Pro, +15.0 TB 2.1, +2.8 MCP Atlas). But M3 is 3.7× cheaper, multimodal (video+image+desktop), and leads BrowseComp (83.5%). Text-only powerhouse vs the Swiss Army knife. Full comparison.
Claude Opus 4.8 leads every benchmark — but GLM-5.2 is within 0.7 pts on FrontierSWE and 0.8 pts on MCP Atlas. At $4.40 vs $25 per 1M (5.7× cheaper) with MIT open weights, GLM-5.2 is the first open-weight model that makes Opus look expensive. Full 8-benchmark comparison from Z.AI & LLM Stats data.
GLM-5.2 (62.1% Pro, MIT open-weight, $4.40/1M) beats GPT-5.5 (58.6%, $30/1M) on SWE-bench Pro by 3.5 points at 1/7 the cost. Also leads HLE w/tools (+2.5), FrontierSWE (+1.8), MCP Atlas (+1.7). GPT-5.5 counters with DeepSWE (+23.8), TB 2.1 (+3.0). Full comparison with 12 shared benchmarks from Z.AI/VentureBeat data.
Claude Opus 4.8 (69.2% Pro, $25/1M) dominates every benchmark vs Kimi K2.6 (58.6%, $4/1M) by 3-11 pts. But Kimi fights back on BrowseComp (-3.9), Agent Swarm (300 sub-agents), DeepSearchQA (92.5%), and is 6.25× cheaper. Full comparison with real benchmark data, 10-point verdict.
GPT-5.5 and Kimi K2.6 are tied at 58.6% SWE-bench Pro. But Kimi leads HLE w/tools (54.0%), DeepSearchQA (+13.9), and Agent Swarm (300 sub-agents). GPT counters with OSWorld (+1.9), BrowseComp, Terminal-Bench (Codex CLI 82.7%), and 7.5× higher cost. The most evenly matched comparison of 2026.