{
  "$schema_version": 1,
  "generated_at": "2026-09-30",
  "note": "Reverse-chronological feed of changes to the model lineup. Append-only. The weekly freshness sweep adds entries here; humans can hand-add for clarifications.",
  "entries": [
    {
      "id": "2026-09-30-google-1",
      "date": "2026-09-30",
      "vendor": "google",
      "type": "new_model",
      "title": "Gemini 4 Argon — Google's new frontier tier",
      "summary": "Deep reasoning over long-horizon workflows: coding, knowledge work, cyber defense, creative writing. Output limit jumps to 1M tokens (from 64K). $2/$10 introductory. Beats Anthropic and OpenAI on most of Google's own benchmarks. Rolling out to Fairwind cyber partners first, then paid API and AI Ultra — no firm date for general access.",
      "source_url": "https://techcrunch.com/2026/09/30/google-releases-gemini-4-argon-called-its-most-powerful-model-yet/"
    },
    {
      "id": "2026-09-29-openai-1",
      "date": "2026-09-29",
      "vendor": "openai",
      "type": "new_model",
      "title": "GPT-6.1 Sol — Astra-class at one-fifth the price",
      "summary": "$2/$10 per M, exactly 1/5 of GPT-6 Astra, with comparable performance. Cached input $0.10/M. 1.05M context, 128K output, effort low→max (medium default). New Ultrafast tier runs ~300 tokens/sec. Prompts over 272K tokens bill 2x input and 1.5x output for the whole request.",
      "source_url": "https://venturebeat.com/technology/openais-gpt-6-1-sol-offers-astra-like-performance-at-1-5th-price-a-new-ultrafast-tier-clocks-at-300-tokens-per-second"
    },
    {
      "id": "2026-09-28-anthropic-1",
      "date": "2026-09-28",
      "vendor": "claude",
      "type": "new_model",
      "title": "Claude Sonnet 5.5 — same price, 30% faster",
      "summary": "Second model in the Claude 5.5 family and the lower-cost complement to Opus 5.5. Holds Sonnet 5's $2/$10 pricing while generating output 30%+ faster and costing up to 30% less per task. Claude Haiku 5.5 announced for 'the coming weeks' with no date or specs.",
      "source_url": "https://www.unite.ai/anthropic-releases-claude-sonnet-5-5-at-unchanged-sonnet-5-pricing/"
    },
    {
      "id": "2026-09-28-openai-2",
      "date": "2026-09-28",
      "vendor": "openai",
      "type": "access",
      "title": "GPT-6 Luna becomes the free-tier ChatGPT default",
      "summary": "Luna replaces the prior default for free ChatGPT users, putting a GPT-6 model in front of the unpaid tier.",
      "source_url": "https://llmgateway.io/timeline"
    },
    {
      "id": "2026-09-24-openai-2",
      "date": "2026-09-24",
      "vendor": "openai",
      "title": "Sora API retirement",
      "type": "deprecation",
      "summary": "OpenAI's Sora API is scheduled to shut down 2026-09-24, following the Sora app retirement on 2026-04-26. End of the Sora brand.",
      "source_url": "https://en.wikipedia.org/wiki/Sora_(text-to-video_model)"
    },
    {
      "id": "2026-09-22-anthropic-1",
      "date": "2026-09-22",
      "vendor": "claude",
      "type": "new_model",
      "title": "Claude Opus 5.5 — new 5.5 family, 20% cheaper",
      "summary": "First model in the Claude 5.5 family. $4/$20 per M (20% below Opus 5), cache reads $0.20/M (60% cheaper). Performs at Fable 5.1 level on most work at ~40% lower run cost. Tops SWE-bench Pro (89.9%) and Terminal-Bench 4.0 (66.4%). 1M context, 128K output, medium effort default. 5-hour usage caps removed. Sonnet 5.5 and Haiku 5.5 to follow.",
      "source_url": "https://www.unite.ai/anthropic-releases-claude-opus-5-5-with-lower-pricing-and-new-safeguards/"
    },
    {
      "id": "2026-09-22-openai-1",
      "date": "2026-09-22",
      "vendor": "openai",
      "type": "new_model",
      "title": "GPT-6 Sol + GPT-6 Luna — API costs cut 50%+",
      "summary": "Sol at $2/$10 (half its predecessor), Luna at $0.10/$0.50 (down from $0.20/$1.20). Permanent prices, not promotional. Both 1.05M context, 128K output, text+image in. Built with Astra's methods. Knowledge cutoffs 2026-04-20 (Sol) and 2026-05-18 (Luna).",
      "source_url": "https://venturebeat.com/technology/openai-releases-gpt-6-sol-and-luna-models-slashing-api-costs-50-or-more"
    },
    {
      "id": "2026-09-22-alibaba-1",
      "date": "2026-09-22",
      "vendor": "alibaba",
      "type": "upcoming",
      "title": "Qwen 4 previewed at Apsara Conference",
      "summary": "Four tiers previewed (Max, Flash, Plus, 27B). Still in training — no release date, price, or downloadable weights. Qwen project lead says 'very soon'. Roadmap targets 5-10T parameters for Qwen 4.5 and Qwen 5.",
      "source_url": "https://pasqualepillitteri.it/en/news/17552/qwen-4-alibaba-apsara-en"
    },
    {
      "id": "2026-09-21-xai-1",
      "date": "2026-09-21",
      "vendor": "xai",
      "type": "new_model",
      "title": "Grok 4.7 — 2.1T parameters",
      "summary": "Up 40% from Grok 4.6's 1.5T. New larger base model, longer RL run, training weighted toward multi-hour tasks. 500K context, $2/$6 (fast variant 2x speed at 2x price). Musk's roadmap now places 4.7, 4.8 and 4.9 ahead of Grok 5.",
      "source_url": "https://x.ai/news"
    },
    {
      "id": "2026-09-10-deepseek-1",
      "date": "2026-09-10",
      "vendor": "deepseek",
      "title": "DeepSeek V4.1 Flash",
      "type": "new_model",
      "summary": "V4.1 refresh of the Flash tier — 1M context, open-weights. Incremental over V4 Flash 0731.",
      "source_url": "https://api-docs.deepseek.com/"
    },
    {
      "id": "2026-09-03-openai-1",
      "date": "2026-09-03",
      "vendor": "openai",
      "type": "new_model",
      "title": "GPT-6 Astra — first GPT-6 model",
      "summary": "SOTA on computer use, browsing, software engineering, cybersecurity, science, professional work. $10/$50 per M, cached input $1, batch half price, Fast mode 2x. 1.05M context, 128K output. Knowledge cutoff 2026-04-30.",
      "source_url": "https://openai.com/index/gpt-6-astra/"
    },
    {
      "id": "2026-09-02-google-1",
      "date": "2026-09-02",
      "vendor": "google",
      "type": "new_model",
      "title": "Gemini 3.8 Flash + 3.8 Flash Cyber",
      "summary": "Third Flash release in six weeks. $0.75/$3.75 — same as 3.7 Flash, both double on 2027-01-01. 1M context, 64K output, text/image/audio/video/PDF in, tunable thinking. Built on the 3.7 Flash base rather than a new one. Ships with a locked-down security sibling, 3.8 Flash Cyber.",
      "source_url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"
    },
    {
      "id": "2026-09-01-anthropic-1",
      "date": "2026-09-01",
      "vendor": "claude",
      "title": "Claude Fable 5.1 + Mythos 5.1",
      "type": "new_model",
      "summary": "World's most advanced models for coding and knowledge work. 1M context, 128K max output, adaptive thinking. Fable 5.1 general use with safeguards; Mythos 5.1 restricted-access with safeguards partially lifted.",
      "source_url": "https://releasebot.io/updates/anthropic/claude"
    },
    {
      "id": "2026-08-26-alibaba-1",
      "date": "2026-08-26",
      "vendor": "alibaba",
      "type": "new_model",
      "title": "Qwen3.8-Flash-Next open-sourced",
      "summary": "Released by the Qwen team as an early preview of the architecture used in Qwen 4.",
      "source_url": "https://www.scmp.com/tech/tech-trends/article/3364404/alibabas-lightweight-qwen-model-takes-larger-ai-systems-openai-deepseek-zhipu"
    },
    {
      "id": "2026-08-21-openai-1",
      "date": "2026-08-21",
      "vendor": "openai",
      "title": "GPT-5.6 Sol pricing cut ~20%",
      "type": "pricing",
      "summary": "OpenAI dropped API + credit pricing of GPT-5.6 Sol by over 20% for the next three months (through ~2026-11-21).",
      "source_url": "https://openai.com/index/gpt-5-6/"
    },
    {
      "id": "2026-08-20-alibaba-1",
      "date": "2026-08-20",
      "vendor": "alibaba",
      "title": "Qwen 3.8 Flash — open-weights, 125B",
      "type": "new_model",
      "summary": "Downloadable, 125B parameters. Competitive with Anthropic Opus 4.6 and DeepSeek V4-Flash.",
      "source_url": "https://www.scmp.com/tech/tech-trends/article/3364404/alibabas-lightweight-qwen-model-takes-larger-ai-systems-openai-deepseek-zhipu"
    },
    {
      "id": "2026-08-15-google-1",
      "date": "2026-08-15",
      "vendor": "google",
      "title": "Gemini 3.7 Flash + Gemini 3.5 Transcribe",
      "type": "new_model",
      "summary": "August 2026 refresh. Gemini 3.7 Flash is the new frontier-intelligence Flash tier. Gemini 3.5 Transcribe is a realtime speech-to-text model for voice agents, live captioning, and post-call analytics.",
      "source_url": "https://blog.google/innovation-and-ai/technology/google-ai-updates-august-2026/"
    },
    {
      "id": "2026-08-14-zai-1",
      "date": "2026-08-14",
      "vendor": "zai",
      "type": "new_model",
      "title": "GLM-5.3 — frontier coding with emergent cyber capabilities",
      "summary": "Same 743B base as GLM-5.2 with improved post-training; HF checkpoint is a 753B MoE (~40B active), 78 layers, 256 routed experts, DeepSeek Sparse Attention, 1M context, 128K output, always reasoning at effort low/high/max. $1.40/$4.40, cached input $0.26. API and open weights staged behind safety evals. Backfilled 2026-09-30 — earlier passes missed it.",
      "source_url": "https://www.explainx.ai/blog/glm-5-3-launch-cyber-defense-benchmarks-august-2026"
    },
    {
      "id": "2026-08-12-xai-1",
      "date": "2026-08-12",
      "vendor": "xai",
      "title": "Grok 4.6",
      "type": "new_model",
      "summary": "Current stable Grok flagship (Grok 5 still in training). Available in Grok Build, Cursor, and the xAI console.",
      "source_url": "https://x.ai/news"
    },
    {
      "id": "2026-08-10-openai-1",
      "date": "2026-08-10",
      "vendor": "openai",
      "title": "GPT-5.6-Cyber (Daybreak Blue + Red)",
      "type": "new_model",
      "summary": "Cybersecurity-specialised variant. Launched with Daybreak Blue and Red programs. Solves ~95% of cybersecurity problems.",
      "source_url": "https://openai.com/index/gpt-5-6/"
    },
    {
      "id": "2026-08-06-openai-1",
      "date": "2026-08-06",
      "vendor": "openai",
      "title": "GPT-5.6 Luna free for all users",
      "type": "access",
      "summary": "ChatGPT update improved Sol reliability and expanded access to GPT-5.6 Luna to free-tier users.",
      "source_url": "https://openai.com/index/improving-gpt-5-6-sol-in-chatgpt/"
    },
    {
      "id": "2026-08-03-alibaba-1",
      "date": "2026-08-03",
      "vendor": "alibaba",
      "title": "Qwen 3.8 Max — 2.4T flagship",
      "type": "new_model",
      "summary": "Alibaba's most capable model. 2.4T parameters. Competes head-to-head with top Anthropic and OpenAI models.",
      "source_url": "https://www.bloomberg.com/news/articles/2026-08-26/alibaba-releases-smaller-cost-effective-qwen-ai-model"
    },
    {
      "id": "2026-07-31-deepseek-1",
      "date": "2026-07-31",
      "vendor": "deepseek",
      "title": "DeepSeek V4 Flash 0731",
      "type": "new_model",
      "summary": "July 31 checkpoint of V4 Flash. Within ~1 point of GPT-5.6 Luna at 40% lower per-task cost.",
      "source_url": "https://api-docs.deepseek.com/"
    },
    {
      "id": "2026-07-27-moonshot-1",
      "date": "2026-07-27",
      "vendor": "moonshot",
      "title": "Kimi K3 weights released",
      "type": "access",
      "summary": "Full open weights for K3 released under a custom license (companies >$20M ARR must negotiate a contract before serving K3 to external customers).",
      "source_url": "https://en.wikipedia.org/wiki/Kimi_(chatbot)"
    },
    {
      "id": "2026-07-24-anthropic-1",
      "date": "2026-07-24",
      "vendor": "claude",
      "type": "new_model",
      "title": "Claude Opus 5",
      "summary": "$5/$25 per M, 1M context, 128K output. Superseded by Opus 5.5 on 2026-09-22. Backfilled into the lineup during the 2026-09-23 freshness pass.",
      "source_url": "https://www.anthropic.com/news/claude-opus-5"
    },
    {
      "id": "2026-07-23-blackforest-1",
      "date": "2026-07-23",
      "vendor": "blackforest",
      "title": "FLUX 3 — BFL's first video model",
      "type": "new_model",
      "summary": "Black Forest Labs enters video generation. 20-second single-pass clips with native audio — the longest single-pass BFL has published.",
      "source_url": "https://bfl.ai/blog"
    },
    {
      "id": "2026-07-16-moonshot-1",
      "date": "2026-07-16",
      "vendor": "moonshot",
      "title": "Kimi K3 — 2.8T open-weight flagship",
      "summary": "New-architecture MoE with ~2.8T total parameters and 1M-token context. Largest open-weight LLM shipped. Competitive with Anthropic Fable 5; substantially outperforms Opus 4.8, GPT-5.6 Sol, GPT-5.5. Modified MIT license; full weights following release. $15/Mtok output.",
      "type": "new_model",
      "source_url": "https://fortune.com/2026/07/16/moonshots-kimi-k3-pushes-chinese-ai-into-fable-level-territory/"
    },
    {
      "id": "2026-07-09-openai-1",
      "date": "2026-07-09",
      "vendor": "openai",
      "title": "GPT-5.6 (Sol / Terra / Luna) + ChatGPT Work",
      "summary": "Three-tier flagship family: Sol ($5/$30) SOTA on coding + science; Terra ($2.50/$15) balanced; Luna ($1/$6) fast + cheap. Ships alongside ChatGPT Work agent for whole-job execution. Programmatic tool calling in the Responses API.",
      "type": "new_model",
      "source_url": "https://openai.com/index/gpt-5-6/"
    },
    {
      "id": "2026-07-08-bytedance-1",
      "date": "2026-07-08",
      "vendor": "bytedance",
      "title": "Seedream 5.0 Pro",
      "type": "new_model",
      "summary": "Pro tier of Seedream 5. Anchored/coordinate-precise edits, editable layers, multi-reference fusion, exact color/material match, multilingual text at 2K/3K.",
      "source_url": "https://seed.bytedance.com/en/"
    },
    {
      "id": "2026-07-08-xai-1",
      "date": "2026-07-08",
      "vendor": "xai",
      "title": "Grok 4.5 — 'Opus-class' flagship",
      "summary": "$2/$6 pricing. Musk: 'roughly comparable to Opus 4.7 but much faster and more token-efficient.' Focus: coding, agents, knowledge work. Trained on tens of thousands of NVIDIA GB300 GPUs. Available in Grok Build, Cursor (all plans), xAI console.",
      "type": "new_model",
      "source_url": "https://x.ai/news/grok-4-5"
    },
    {
      "id": "2026-06-30-anthropic-1",
      "date": "2026-06-30",
      "vendor": "claude",
      "title": "Claude Sonnet 5 — most agentic Sonnet yet",
      "summary": "1M context default, 128k max output, adaptive thinking, browser + terminal tools. Near-Opus 4.8 perf at lower cost. $2/$10 introductory through 2026-08-31, then $3/$15. Default model for free and Pro plans.",
      "type": "new_model",
      "source_url": "https://www.anthropic.com/news/claude-sonnet-5"
    },
    {
      "id": "2026-06-30-google-1",
      "date": "2026-06-30",
      "vendor": "google",
      "title": "Nano Banana 2 Lite — cheap image gen",
      "summary": "Faster, cheaper Nano Banana. ~4-second generation, $0.034 per 1000 images.",
      "type": "new_model",
      "source_url": "https://techcrunch.com/2026/06/30/google-introduces-a-faster-cheaper-image-generator-with-nano-banana-2-lite/"
    },
    {
      "id": "2026-06-23-bytedance-1",
      "date": "2026-06-23",
      "vendor": "bytedance",
      "title": "Seedance 2.5 — 30-second single-pass video",
      "summary": "Announced at the 2026 Volcano Engine conference. Seedance 2.5 generates 30-second videos in a single pass. Seedance 2.0 updated same event with 4K output.",
      "type": "new_model",
      "source_url": "https://kie.ai/blog/seedance-2-5-release-deep-dive"
    },
    {
      "id": "2026-06-15-moonshot-2",
      "date": "2026-06-15",
      "vendor": "moonshot",
      "title": "Kimi K2.7-Code HighSpeed (Beta)",
      "summary": "HighSpeed variant rolling out to Kimi Code Beta — ~180 tok/s median, up to ~260 tok/s on short context. Roughly 6× the standard K2.7-Code throughput.",
      "type": "new_model",
      "source_url": "https://www.techtimes.com/articles/318414/20260615/kimi-k27-code-adds-highspeed-mode-skips-independent-benchmark-submission.htm"
    },
    {
      "id": "2026-06-15-xai-1",
      "date": "2026-06-15",
      "vendor": "xai",
      "title": "Grok V9-Medium expected mid-June",
      "summary": "Musk announced 2026-05-25 that V9-Medium (1.5T parameters, 3× current production size) finished training. Public release expected mid-June 2026.",
      "type": "upcoming",
      "source_url": "https://x.ai/news"
    },
    {
      "id": "2026-06-13-zai-1",
      "date": "2026-06-13",
      "vendor": "zai",
      "title": "GLM-5.2 — 1M-context coding-first frontier",
      "summary": "Live across every GLM Coding Plan tier. 1M usable context, coding-first positioning. Compatible with Claude Code, Cline, OpenCode, Roo Code, Goose, Crush, OpenClaw, Kilo Code. MIT-licensed weights coming.",
      "type": "new_model",
      "source_url": "https://aitoolly.com/ai-news/article/2026-06-14-zhipu-ai-releases-glm-52-a-fully-open-source-frontier-model-featuring-a-1m-context-window"
    },
    {
      "id": "2026-06-12-moonshot-1",
      "date": "2026-06-12",
      "vendor": "moonshot",
      "title": "Kimi K2.7-Code — 1T MoE open-weight coder",
      "summary": "1T MoE (32B active, 384 experts), 256K context, Modified MIT license. Cuts reasoning tokens 30% vs K2.6; +21.8% on Kimi Code Bench v2. Available on Hugging Face, Kimi API, Kimi Code CLI.",
      "type": "new_model",
      "source_url": "https://cryptobriefing.com/kimi-k2-7-code-open-source-release/"
    },
    {
      "id": "2026-06-09-anthropic-1",
      "date": "2026-06-09",
      "vendor": "claude",
      "title": "Claude Fable 5 + Mythos 5 — new Mythos-class tier",
      "summary": "First models in the new Mythos-class tier, sitting above Opus. Fable 5 brings Mythos-class capability to general use with safeguards; Mythos 5 remains restricted to approved programs and trusted-access use.",
      "type": "new_model",
      "source_url": "https://www.anthropic.com/news"
    },
    {
      "id": "2026-06-09-moonshot-1",
      "date": "2026-06-09",
      "vendor": "moonshot",
      "title": "Kimi Work — desktop app, 300 sub-agents local",
      "summary": "Desktop application powered by Kimi K2.6 that coordinates up to 300 specialized sub-agents on the user's local machine — no cloud required.",
      "type": "new_model",
      "source_url": "https://cryptobriefing.com/moonshot-ai-kimi-work-300-agents-desktop/"
    },
    {
      "id": "2026-06-08-google-1",
      "date": "2026-06-08",
      "vendor": "google",
      "title": "Gemini 3.5 Flash GA in Gemini Enterprise",
      "summary": "3.5 Flash became generally available in Gemini Enterprise, enabled by default starting 2026-06-08. Frontier-intelligence Flash tier — fast, cheap, multimodal.",
      "type": "new_model",
      "source_url": "https://docs.cloud.google.com/gemini/enterprise/docs/release-notes"
    },
    {
      "id": "2026-06-04-blackforest-1",
      "date": "2026-06-04",
      "vendor": "blackforest",
      "title": "FLUX.2 [klein] — first FLUX on consumer hardware",
      "summary": "Partnership with ASUS + NVIDIA. First FLUX model optimized to run on consumer hardware — shipping with ASUS ProArt laptops.",
      "type": "new_model",
      "source_url": "https://blogs.nvidia.com/blog/rtx-ai-garage-flux-2-comfyui/"
    },
    {
      "id": "2026-06-03-openai-1",
      "date": "2026-06-03",
      "vendor": "openai",
      "title": "GPT-Rosalind life-sciences expansion",
      "summary": "Capability expansion for vetted developers + U.S. government partners — biological reasoning, medicinal chemistry, genomics analysis, experimental workflow capabilities. Restricted/trusted-access tier.",
      "type": "capability",
      "source_url": "https://openai.com/research/index/release/"
    },
    {
      "id": "2026-06-01-alibaba-1",
      "date": "2026-06-01",
      "vendor": "alibaba",
      "title": "Qwen 3.7 Plus — low-cost multimodal agent",
      "summary": "Generally available 2026-06-01. Multimodal (image + video). 1M-token context. Roughly 1/6 the per-token price of the text-only Qwen 3.7 Max.",
      "type": "new_model",
      "source_url": "https://www.marktechpost.com/2026/06/02/alibabas-qwen-team-launches-qwen3-7-plus-adding-vision-deep-reasoning-tool-invocation-and-autonomous-iteration-on-the-bailian-platform/"
    },
    {
      "id": "2026-06-01-xai-1",
      "date": "2026-06-01",
      "vendor": "xai",
      "title": "Grok Build 0.1 — agentic coding model (public beta)",
      "summary": "Purpose-built coding model trained for agentic workflows. Text + image input, 256K-token context. Available in public beta on the xAI API.",
      "type": "new_model",
      "source_url": "https://x.ai/news"
    },
    {
      "id": "2026-05-29-openai-1",
      "date": "2026-05-29",
      "vendor": "openai",
      "title": "GPT-Rosalind launched (Biodefense)",
      "summary": "OpenAI launched Rosalind Biodefense — expanded trusted access to GPT-Rosalind for vetted developers and U.S. government partners advancing biodefense, public health, and pandemic preparedness.",
      "type": "new_model",
      "source_url": "https://openai.com/research/index/release/"
    },
    {
      "id": "2026-05-28-blackforest-1",
      "date": "2026-05-28",
      "vendor": "blackforest",
      "title": "FLUX Virtual Try-On",
      "summary": "Garment try-on for retail — sub-4-second generations across thousands of products. High garment consistency.",
      "type": "new_model",
      "source_url": "https://bfl.ai/blog"
    },
    {
      "id": "2026-05-28-anthropic-1",
      "date": "2026-05-28",
      "vendor": "claude",
      "title": "Claude Opus 4.8 + Dynamic Workflows preview",
      "summary": "Opus 4.8 shifts emphasis to honesty and reliability — ~4× less likely than 4.7 to let flaws in code it has written pass unremarked. Ships alongside a Dynamic Workflows research preview in Claude Code that orchestrates hundreds of parallel subagents.",
      "type": "new_model",
      "source_url": "https://www.anthropic.com/news"
    },
    {
      "id": "2026-05-21-blackforest-1",
      "date": "2026-05-21",
      "vendor": "blackforest",
      "title": "FLUX Erase",
      "summary": "Removes masked objects, shadows, and reflections — reconstructs the scene behind them.",
      "type": "new_model",
      "source_url": "https://bfl.ai/blog"
    },
    {
      "id": "2026-05-20-alibaba-1",
      "date": "2026-05-20",
      "vendor": "alibaba",
      "title": "Qwen 3.7 Max — Agent Frontier flagship",
      "summary": "Unveiled at Alibaba Cloud Summit Hangzhou. 1M-token context, 56.6 on Artificial Analysis Intelligence Index. Tops Claude Opus 4.6 Max on Terminal-Bench 2.0, SWE-Bench Pro, MCP-Atlas. Closed weights.",
      "type": "new_model",
      "source_url": "https://www.scmp.com/tech/big-tech/article/3354212/alibaba-unveils-new-qwen-model-custom-chips-bid-become-chinas-ai-factory"
    },
    {
      "id": "2026-05-19-google-1",
      "date": "2026-05-19",
      "vendor": "google",
      "title": "Gemini 3.5 Pro announced at Google I/O 2026",
      "summary": "Frontier multimodal flagship announced at I/O 2026 with CEO Sundar Pichai promising general availability 'next month'. Still in limited preview as of 2026-06-16; GA expected late June.",
      "type": "new_model",
      "source_url": "https://ai.google.dev/gemini-api/docs/changelog"
    },
    {
      "id": "2026-05-15-google-2",
      "date": "2026-05-15",
      "vendor": "google",
      "title": "Nano Banana Pro (Gemini 3 Pro Image) GA",
      "summary": "Nano Banana Pro reached general availability via Gemini Enterprise Agent Platform — high-fidelity image gen for tasks requiring maximum factual accuracy.",
      "type": "new_model",
      "source_url": "https://cloud.google.com/blog/products/ai-machine-learning/nano-banana-2-and-nano-banana-pro-are-generally-available/"
    },
    {
      "id": "2026-05-14-blackforest-1",
      "date": "2026-05-14",
      "vendor": "blackforest",
      "title": "FLUX.1 Tools — control + steerability suite",
      "summary": "Suite designed to add control and steerability to the base text-to-image FLUX.1 model (depth, canny, fill, redux).",
      "type": "new_model",
      "source_url": "https://bfl.ai/blog"
    },
    {
      "id": "2026-05-10-shell-1",
      "date": "2026-05-10",
      "vendor": "shell",
      "title": "Repository data layer introduced",
      "summary": "Volatile facts (model IDs, prices, dates, status) extracted to data/models.json. Manuals render lineup tables from JSON. Weekly freshness sweep will append entries here automatically.",
      "type": "infra",
      "source_url": null
    },
    {
      "id": "2026-05-07-openai-1",
      "date": "2026-05-07",
      "vendor": "openai",
      "title": "New realtime voice models in the OpenAI API",
      "summary": "Realtime voice models that can reason, translate, and transcribe speech in a single low-latency pipeline.",
      "type": "new_model",
      "source_url": "https://openai.com/research/index/release/"
    },
    {
      "id": "2026-05-05-openai-1",
      "date": "2026-05-05",
      "vendor": "openai",
      "title": "GPT-5.5 Instant — new ChatGPT default",
      "summary": "ChatGPT's default model updated to GPT-5.5 Instant. Smarter, more accurate answers, reduced hallucinations, improved personalization controls.",
      "type": "new_model",
      "source_url": "https://techcrunch.com/2026/05/05/openai-releases-gpt-5-5-instant-a-new-default-model-for-chatgpt/"
    },
    {
      "id": "2026-04-30-xai-1",
      "date": "2026-04-30",
      "vendor": "xai",
      "title": "Grok 4.3 — current flagship",
      "summary": "Released 2026-04-30. $1.25/$2.50 — cheapest hosted frontier on list price. 1M context, native video input, always-on reasoning. ~40% input price cut vs Grok 4.",
      "type": "new_model",
      "source_url": "https://docs.x.ai/docs/release-notes"
    },
    {
      "id": "2026-04-24-deepseek-1",
      "date": "2026-04-24",
      "vendor": "deepseek",
      "title": "DeepSeek V4 (Pro + Flash) launched",
      "summary": "V4 Pro: 1.6T / 49B-active MoE, 1M context, MIT-licensed. Open-source SOTA on agentic-coding benchmarks. V4 Flash: 284B / 13B-active, same context. 75% Pro discount runs through 2026-05-31.",
      "type": "new_model",
      "source_url": "https://api-docs.deepseek.com/news/news260424"
    },
    {
      "id": "2026-04-23-openai-1",
      "date": "2026-04-23",
      "vendor": "openai",
      "title": "GPT-5.5 / 5.5 Pro in ChatGPT (API TBA)",
      "summary": "Available in ChatGPT (Plus/Pro/Business/Enterprise) + Codex now. Built around long-running goal completion. API pricing TBA.",
      "type": "new_model",
      "source_url": "https://help.openai.com/en/articles/9624314-model-release-notes"
    },
    {
      "id": "2026-04-20-moonshot-1",
      "date": "2026-04-20",
      "vendor": "moonshot",
      "title": "Kimi K2.6 — 1T-MoE open-weight agent flagship",
      "summary": "Modified MIT license. 1T total / 32B active. 262K context, 16K max output. Agent Swarm: 300 sub-agents × 4000 steps per run. $0.60/$2.50 on Moonshot API.",
      "type": "new_model",
      "source_url": "https://huggingface.co/moonshotai"
    },
    {
      "id": "2026-04-16-claude-1",
      "date": "2026-04-16",
      "vendor": "claude",
      "title": "Claude Opus 4.7 — current flagship",
      "summary": "$5/$25 per M tokens. 200K context. New xhigh effort dial. Major vision lift (98.5% visual-acuity). Task budgets (public beta), file-system memory. Tokenizer change: 1.0–1.35× more tokens than 4.6.",
      "type": "new_model",
      "source_url": "https://www.anthropic.com/news"
    },
    {
      "id": "2026-04-alibaba-1",
      "date": "2026-04-15",
      "vendor": "alibaba",
      "title": "Qwen 3.6 wave — Max + Plus + multimodal",
      "summary": "Qwen 3.6 Max (proprietary flagship, 1M+ context, agentic + visual reasoning). Qwen 3.6 Plus (high-perf, speed/cost tier). Wan 2.7 (video gen) + HappyHorse 1.0 (image-to-video, top-ranked) added to Model Studio.",
      "type": "new_model",
      "source_url": "https://modelstudio.console.alibabacloud.com"
    },
    {
      "id": "2026-04-08-zai-1",
      "date": "2026-04-08",
      "vendor": "zai",
      "title": "GLM-5.1 open-sourced",
      "summary": "745B / 44B-active MoE, MIT license, 200K context. DeepSeek Sparse Attention integrated. From China's first publicly-traded AI lab. Subscription preview late March.",
      "type": "new_model",
      "source_url": "https://huggingface.co/zai-org"
    },
    {
      "id": "2026-03-31-xai-2",
      "date": "2026-03-31",
      "vendor": "xai",
      "title": "Grok 4.20 — 2M-context fast frontier",
      "summary": "$2.00/$6.00 per M tokens. 2M context — largest in the field alongside Gemini 3.1 Pro. Strong agentic tool calling.",
      "type": "new_model",
      "source_url": "https://docs.x.ai/docs/release-notes"
    },
    {
      "id": "2026-03-05-openai-1",
      "date": "2026-03-05",
      "vendor": "openai",
      "title": "GPT-5.4 + Computer Use SOTA",
      "summary": "75.0% on OSWorld-Verified (vs 47.3% for GPT-5.2). Beats human baseline 72.4%. 95% first-try / 100% within 3 tries on real portals.",
      "type": "new_model",
      "source_url": "https://help.openai.com/en/articles/9624314-model-release-notes"
    },
    {
      "id": "2026-03-05-lightricks-1",
      "date": "2026-03-05",
      "vendor": "specvideo",
      "title": "Lightricks LTX 2.3 — open-source 4K video",
      "summary": "22B Diffusion Transformer. Native 4K @ 50fps with synchronized audio. FP8 quantized fits 24GB VRAM (RTX 4090/5090). ~$0.04/sec hosted.",
      "type": "new_model",
      "source_url": "https://huggingface.co/Lightricks/LTX-2.3"
    },
    {
      "id": "2026-02-26-google-1",
      "date": "2026-02-26",
      "vendor": "google",
      "title": "Nano Banana 2 (Gemini 3.1 Flash Image)",
      "summary": "Google's viral image generator refreshed. 512px–4K, character consistency up to 5, object fidelity up to 14, web-grounded via real-time search.",
      "type": "new_model",
      "source_url": "https://blog.google/innovation-and-ai/technology/ai/nano-banana-2/"
    },
    {
      "id": "2026-02-19-google-1",
      "date": "2026-02-19",
      "vendor": "google",
      "title": "Gemini 3.1 Pro — 2M context, ARC-AGI-2 SOTA",
      "summary": "$2/$12 per M (≤200K input); $4/$18 above. 77.1% on ARC-AGI-2 (vs 31.1% for Gemini 3 Pro). 94.3% on GPQA Diamond. Deep Think mode.",
      "type": "new_model",
      "source_url": "https://ai.google.dev/gemini-api/docs/changelog"
    },
    {
      "id": "2026-02-12-bytedance-1",
      "date": "2026-02-12",
      "vendor": "bytedance",
      "title": "Seedance 2.0 — unified audio-video",
      "summary": "Accepts up to 9 image / 3 video / 3 audio references per prompt. 4-15s multi-shot output with dual-channel audio. Most flexible reference inputs in video gen.",
      "type": "new_model",
      "source_url": "https://seed.bytedance.com/en"
    },
    {
      "id": "2026-02-11-zai-1",
      "date": "2026-02-11",
      "vendor": "zai",
      "title": "GLM-5 — first 745B MoE flagship",
      "summary": "745B total / 40-44B active MoE. 200K context. Pre-training data 28.5T tokens (up from 23T on GLM-4.5).",
      "type": "new_model",
      "source_url": "https://huggingface.co/zai-org"
    },
    {
      "id": "2026-02-alibaba-1",
      "date": "2026-02-08",
      "vendor": "alibaba",
      "title": "Qwen 3.5 Series — 60% cheaper, 8× faster",
      "summary": "Open-source + API. Native multimodal (text + image + video). Includes Qwen 3.5 Omni for full audio + multimodal.",
      "type": "price_change",
      "source_url": "https://huggingface.co/Qwen"
    },
    {
      "id": "2026-01-13-google-1",
      "date": "2026-01-13",
      "vendor": "google",
      "title": "Veo 3.1 — 4K + Scene Extension",
      "summary": "Native 9:16 mobile, 4K upscaling, Scene Extension enables continuous narratives exceeding 60 seconds.",
      "type": "new_model",
      "source_url": "https://deepmind.google/models/veo/"
    },
    {
      "id": "2025-08-05-openai-1",
      "date": "2025-08-05",
      "vendor": "openai",
      "title": "gpt-oss-120b + gpt-oss-20b (open weights)",
      "summary": "OpenAI's first meaningful open release since GPT-2. 120B MoE (5.1B active/token) on a single 80GB H100. Apache 2.0 license. Near-parity with o4-mini on core reasoning. Companion 20B for edge inference.",
      "type": "new_model",
      "source_url": "https://openai.com/index/introducing-gpt-oss/"
    }
  ]
}