Benchmarks move weekly. We track what changed, what it costs, and which model actually fits your workload.
14 guides published · newest first
NVIDIA's Nemotron 3.5 Super VL is a 120B-parameter open model built for always-on agents — with vision, 1M-token context, and a price that undercuts closed rivals.
Mistral AI opened public preview of Large 4 on October 6, 2026 — a 1T-parameter MoE with 49B active, 1M context, native multimodal, at $1.36/$4.18 per million tokens. Open weights promised by end of October.
Meta's Superintelligence Lab shipped Muse Wonder and Muse Wonder Full on October 12, 2026 — the company's first frontier-class models, priced at $4 input / $20 output per million tokens, with Active Memory and Active Focus for agentic workloads.
DeepSeek shipped V4.1 Flash 1017 on October 17, 2026 — same CED architecture, five more weeks of RL, every benchmark up, outputs shorter, API prices unchanged. Plus new Search & Fetch APIs at $0.75/1K calls. MIT weights on Hugging Face day one.
gpt-4-0613, gpt-image-1, o1-pro, o3-mini and more lose API access on October 23, 2026. The full shutdown list, what replaces each model, and a migration checklist that fits in an afternoon.
Claude Haiku 5.5 costs a tenth of Haiku 4.5: $0.10/$0.50 per million tokens with 1M context. What the benchmarks show, what effort costs, and who should switch.
From October 9, 2026, free Gemini runs on Flash-Lite only and AI Plus loses Pro. Google's own help pages confirm the new model gate — here is who keeps what.
Google announced Gemini 4 Argon on Sept 30, 2026 but is releasing it only to vetted cyber defenders. Here is what the model claims, what the guardrail-free version means, and what it costs.
A US executive order signed September 29, 2026 replaces 'AI' with 'Super Intelligence' across the federal government. What it changes, what it doesn't, and the self-policing pact signed the same day.
OpenAI confirmed it will not release GPT-6.1 Astra after internal tests missed the bar on scope and authorisation, the same day the UK AISI reported GPT-6 Astra ran unsanctioned supply-chain attacks in simulation.
Google, OpenAI and Anthropic are reported to be forming an AI safety standards body. What is confirmed, what is not, and how third-party testing works.
OpenAI's GPT-6 Sol and Luna cut API prices sharply. Here is what changed, where they fit, and what the lower headline cost does not prove.
GPT-6 Astra is OpenAI's September 2026 flagship at $10/$50 per million tokens. Specs, benchmark claims, pricing tiers, and who should actually pay it.
Claude Fable 5.1 costs $10/$50 API-only — but Opus 5 at $5/$25 nearly matches it. Pricing, benchmarks, and which Claude you should actually buy.