Notes on open-source AI, model launches, and the tools we build at ChatOSS.
Moonshot's open-weight flagship tops Terminal-Bench 4.0 at 67.2% in Artificial Analysis's independent run — then scores 25.5% on FrontierCode 2.0 against Claude Opus 5.5's 46.4%. The best open model is now a specialist, and the cache math still favors Sonnet.
AutomationBench is the closest public test of orchestration, and MiniMax M3.1 Flash wins it on both the independent run (73% vs 67%) and the vendors' table (62.8 vs 60.5). It also leads HLE in Artificial Analysis's run, 48% to 43% — but DeepSeek V4.1 Flash 1017 still owns coding, cache-heavy bills and the MIT licence. The honest case, including the numbers that cut against it.
Both list at $0.15/$0.60 per million tokens, and MiniMax wins the business-workflow benchmarks. But on vals.ai's Vibe Code Bench v1.1, DeepSeek's October 17 checkpoint scores 87.73% at $0.58 per test while M3.1 Flash scores 73.94% at $1.15 — because the cheaper token burns 66.86M of them per task. Here is the independent evidence and the cache-aware math.
Both labs ship a million-token context, MIT weights, and a ~$0.15-per-million price tag. On the independent harnesses that have actually run both models — vals.ai's Vibe Code Bench and Terminal-Bench 2.1, BenchLM's Vibe Code Bench 1-100, and Artificial Analysis' v4.3 index — DeepSeek V4.1 Flash wins agentic coding, and by a wider margin than either vendor's own table suggests.
DeepSeek's new ~748B-parameter Flash model (552B backbone + 196B Engram memory) retires V4 Pro, ships native vision, and — on DeepSeek's own max-effort numbers — edges Claude Opus 5 and GPT-5.6 Sol on Terminal-Bench 2.1 and DeepSWE v1.1 while costing $0.15/$0.60 per million tokens. Full benchmark table, pricing math, and where the closed frontier still wins.
Z.ai shipped a post-trained flagship (GLM-5.3) and a brand-new 320B multimodal sibling (GLM-5.3 Flash) two weeks apart. We benchmark both against Claude Opus 5, Fable 5, and GPT-5.6 Sol — with charts, a comparison table, and the pricing math.
31.2 million input tokens, 107K output tokens, 157 API calls — and 28 cents. Here's the real billing-log breakdown, what the identical token volume would have cost on Claude Opus 5, GPT-5.6 Sol, and Claude Fable 5, and the two numbers I had wrong before I checked.
DeepSeek's flagship just went GA — and on SWE-bench Verified its open-weight scores match Anthropic's closed frontier while costing 10–100× less per task. Here's the full head-to-head with Opus 5, Fable 5, and GPT-5.6 Sol, with a price-vs-intelligence chart and comparison table.
DeepSeek V4 Flash costs 10× less than GLM-5.2 and matches or beats it on several coding benchmarks. But GLM-5.2 still leads on broad intelligence and reasoning. Here’s the full head-to-head with charts and a comparison table.
Ollama paused new $100/mo Max subscriptions because cloud token volume is doubling every month. Kimi K3 just landed on the platform, GLM-5.2 forced a GPU doubling, and even Max users hit an “extra usage only” wall. Here's what's happening.
Three frontier models, one week of launches. We stack Kimi K3's open-weight numbers against Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol across intelligence, coding, context, and price — with charts.
Moonshot AI's breakout open model isn't offline — but a record-breaking launch has pushed its compute to the limit, and new subscriptions are paused. Here's what's actually happening.