AI Models

MiniMax M3.1 Flash went GA on 19 Oct 2026 at a flat $0.15/$0.60 per 1M tokens, half of M3's price. A 428B/23B MoE with 1M context and always-on reasoning, it scores 43 on Artificial Analysis vs M3's 29. Covers license terms, vLLM/SGLang serving and M Plan tiers.

Meta's 816B/42B-active open-weight MoE (Oct 15, 2026): what the Muse Model License allows, hardware from two Mac Studios to 2TB GPU clusters, how to run it with llama.cpp or vLLM, OpenRouter pricing (Oct 2026) and model-card benchmarks.

Kimi K3.1 Pro launched Oct 23, 2026 at $2 input and $10 output per 1M tokens, with a 1M-token context and thinking that cannot be turned off. It scores 55 on Artificial Analysis at $2.11 per task vs $5.98 for Claude Opus 5.5. Weights are promised by Nov 4 but not yet released.

OpenAI shuts down GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini snapshots on Jan 14, 2027, fine-tunes included; new fine-tuning jobs stop Jan 6. The replacement, gpt-6-luna ($0.10/$0.50 per 1M), reasons by default and can't be fine-tuned.

Gemini 4 Argon (gemini-4-argon) went GA Oct 13, 2026. $2/$10 per 1M tokens through Nov 30, $4/$20 from Dec 1. 1M unified window, 65,536 default output, paid tier only, implicit caching. Leads DeepSWE v1.1 (77.9%), trails Opus 5.5 on Terminal-Bench 4.0.

Claude Haiku 5.5 (claude-haiku-5-5, Oct 7, 2026): $0.10/$0.50 per MTok up to 100K prompt tokens, $0.50/$2.50 above, with cache counted. 1M context, 128K output, effort parameter, 400-error breaking changes, AA index 43 vs GPT-6 Luna's 38.

claude-sonnet-4-5-20250929 retires Nov 30, 2026. Sonnet 5.5 costs $2/$10/$0.10 cache read vs $3/$15/$0.30, but turns the same text into ~30% more tokens and rejects sampling parameters, thinking budgets, prefill and forced tool choice. Cost math and a 5-week migration plan.

The anonymous Stealth model on OpenRouter had a 1M context, a 524,288-token output cap found on no other model, and no lab behind it. Here is the evidence for MiniMax M3.1-Flash-Preview, the counter-case, and how to fingerprint a stealth model yourself.

Meta published the Muse Gadget SDK on 2 October 2026: open-source firmware that turns a $12 ESP32 board or a spare Raspberry Pi into a device your Muse agent can drive. The 24 supported boards, real costs, and the clause that caps it at 50 free units.

Alibaba added three Qwen3.8 SKUs in September with no launch post. Max-Prime is a higher-throughput serving tier at double the price, not new weights. Omni-Flash accepts audio and video but returns text only, at Flash pricing. Full lineup, pricing and licence tables.

Z.ai added GLM-5.3 Prime and GLM-5.3 FlashX in September 2026. Neither is a new model — both are accelerated serving tiers of weights that already shipped. Here is the full GLM-5.3-era lineup, with verified prices, licences and a decision rule.

Verified Grok 4.7 API pricing: $2 input and $6 output per 1M tokens, doubling at 200K prompt tokens, plus server-side tool fees and worked monthly cost examples.

GPT-6 Luna costs $0.10/1M input, $0.01 cached and $0.50 output with a 1.05M-token context. It is the only model behind the Decisions API (GA Oct 22, 400 questions/request), is rolling out to ChatGPT Free and Go, and replaces GPT-4o/4.1 snapshots retiring Jan 14, 2027.

Everything Meta announced for Muse at Connect 2026: the Muse Charm keychain, the Muse Realtime Avatar model (448x768 at 25fps, ~870ms), Realtime Voice, Mac computer use, glasses, connectors, Muse for Small Business and the Meta Enterprise Platform.

Run Qwen-Image-2.1 locally on 8-24GB GPUs: official ComfyUI files, GGUF quants for low VRAM, and the Diffusers pipeline. Includes a VRAM table, transparent PNG and multi-reference editing, the vendor benchmark, and the non-commercial license caveat.

Claude Sonnet 5.5 and GPT-6.1 Sol launched a day apart at the same $2/$10 price. Sonnet leads Artificial Analysis's Intelligence Index v4.3.2 56.00 to 51.83; Sol costs $0.72 per task vs $7.67. Benchmarks, caching, long-context pricing and a worked cost example.

Sonnet 5.5 costs half of Opus 5.5 per token and nearly matches it on benchmarks, but uses ~60% more output tokens — $7.67 vs $5.98 per task at max effort on Artificial Analysis v4.3.2. Benchmarks, a per-task cost worked example, and a task-by-task pick table.

MiMo-V2.6-Flash is Xiaomi's MIT-licensed 309B/15B-active MoE at $0.14/$0.28 per 1M tokens. The Distill-Qwen-9B runs on an 8 GB GPU. Specs, benchmarks, VRAM by quant, and Ollama / LM Studio / llama.cpp commands.

GPT-6.1 Sol costs $2/$0.10/$10 per 1M tokens. Ultrafast went live Oct 8 at $12/$0.60/$60, 12h cache writes cost $5 per 1M, it is the Codex CLI default since 0.161.0, and it is in Google Cloud private preview. Benchmarks, Codex multipliers, cost math.

Claude Sonnet 5.5 (claude-sonnet-5-5): $2/$10 per 1M tokens, cache reads $0.10 since Oct 7 (half of Opus 5.5), 1M context, 128K output. Worked cache math, Artificial Analysis's $5.46 vs $5.98 per task, Haiku 5.5, API 2026-10-21 changes and the Nov 30 Sonnet 4.5 retirement.

Grok 4.7 explained: the grok-4.7 model ID, 500k context, $2/$6 pricing with the 200k tier and Fast variant, xAI vs Artificial Analysis benchmarks, API quickstart, availability and who should use it.

Grok 4.7 vs Grok 4.6: same $2/$6 price and 500K context, a new larger base model, +17 points on Terminal-Bench, but ~2.25x the tokens per task. Specs, benchmark deltas, real cost per task, API migration, and the Grok 4.8 outlook.

Claude Opus 5.5 and GPT-6 Sol launched 90 minutes apart on 22 Sept 2026. Opus 5.5 wins on benchmarks; Sol costs half per token. Pricing, cost per task, shared benchmarks, Claude Code vs Codex, and which to pick.

Migrating from Claude Opus 5 to Opus 5.5: thinking can't be disabled, forced tool_choice returns 400, thinking blocks are bound to the conversation, computer_20251124 is rejected, and effort defaults to medium. Exact errors, before/after code, costs, and a rollback plan.