Rycode

Concept map

Concepts on this site, and the articles that carry each one.

  • Macro & rates
  • Build-out & energy
  • Models & safety
  • Ring: links reach more than one cluster
11 Open weights 14 Frontier pricing 7 Agent swarms 9 Evals and safety 1 Force majeure 7 Power and energy 1 Data-center financing 1 Private credit 1 AI capex 3 Treasury yields 2 Fed policy 2 Debt stock 2 Mortgages 1 Oil prices 1 Jobs 2 Equity prices 47 Model releases 20 Hardware 3 Training loop

All concepts

Open weights

Published model weights you can download and run yourself, versus renting a frontier API.

  • Frontier price meets open-weight, and the IPO clock Open-weight cost curves are colliding with frontier listing calendars, so IPO timing and API pricing get squeezed together.
  • GLM-5.3 GLM-5.3 is Z.AI's flagship coding and agent model, with public weights on Hugging Face.
  • Gemma 4 Gemma 4 is Google's open-weight family, separate from the hosted Gemini API.
  • Kimi K3 Kimi K3 is Moonshot's open-weight flagship for long-horizon coding and knowledge work.
  • MiMo-V2.6-Pro MiMo-V2.6-Pro is Xiaomi's flagship open-weight model, announced 22 September 2026 with downloadable weights.
  • Qwen3.8 Max Qwen3.8 Max is Alibaba's lead hosted Qwen model, with an open checkpoint beside it on Hugging Face.
  • Flash DeepSeek-V4.1-Flash is DeepSeek's current recommended API model, id deepseek-flash, with open weights on Hugging Face.
  • Llama 4 Llama 4 is Meta's current open-weight generation, centered on Scout and Maverick.
  • Muse Glimmer Muse Glimmer is Meta's 30-billion-parameter open-weight model, distilled from Muse Spark, under the Apache 2.0 license.
  • R1 DeepSeek-R1 is the January 2025 open reasoner, released under an MIT license, and this page keeps it as an archive.
  • AI glossary Twenty plain-English AI terms so a non-expert can follow the rest of the /ai door without hunting definitions elsewhere.

Frontier pricing

What a closed lab charges for tokens, and how an open-weight price pulls that number down.

  • Frontier price meets open-weight, and the IPO clock Open-weight cost curves are colliding with frontier listing calendars, so IPO timing and API pricing get squeezed together.
  • Claude Opus 5.5 Claude Opus 5.5 is the first Claude 5.5 model, released 22 September 2026, and Anthropic prices it below Opus 5 for similar work.
  • Claude Sonnet 5.5 Claude Sonnet 5.5 is the second Claude 5.5 model, released 28 September 2026, at $2 per million input tokens and $10 per million output tokens.
  • Claude Haiku 5.5 Claude Haiku 5.5 is the third Claude 5.5 model, released 7 October 2026, and Anthropic's fastest, priced from $0.10 per million input tokens and $0.50 per million output tokens.
  • Claude Opus 5 Claude Opus 5 is Anthropic's default model for most agentic coding and enterprise work, at $5 per million input tokens and $25 per million output tokens.
  • Claude Sonnet 5 Claude Sonnet 5 is Anthropic's faster everyday model for coding and agents, at $2 per million input tokens and $10 per million output tokens.
  • Claude Haiku 4.5 Claude Haiku 4.5 is Anthropic's fastest near-frontier model, priced for cost-sensitive work and sub-agents.
  • Claude Fable 5.1 Claude Fable 5.1 is Anthropic's most capable widely released model, built for long-running agents and document work.
  • GPT-6 Sol GPT-6 Sol is OpenAI's mid-tier GPT-6 model for complex coding and agent work, released 22 September 2026 at $2 and $10 per million tokens.
  • GPT-6.1 Sol GPT-6.1 Sol is an upgrade to GPT-6 Sol, released 29 September 2026, at $2 input, $0.10 cached input, and $10 output per million tokens.
  • GPT-6 Luna GPT-6 Luna is the cheaper, faster GPT-6 tier released with Sol on 22 September 2026, at $0.10 and $0.50 per million tokens.
  • GPT-6 Astra GPT-6 Astra is OpenAI's flagship for hard end-to-end work: reasoning, coding, computer use, research, and documents.
  • Gemini 4 Argon Gemini 4 Argon was announced 30 September 2026 for trusted cyber defenders, not the public, at an intro price of $2 input and $10 output per million tokens.
  • Grok 4.7 Grok 4.7 is xAI's current model for coding, agents, and knowledge work, API id grok-4.7, shipped 21 September 2026.

Agent swarms

Many agents on one job, spending tokens and messages to reach a result one model does not.

Evals and safety

Tests that measure raw capability, and the scores labs publish when a model ships.

  • Gemini 3.1 Pro Preview Gemini 3.1 Pro Preview is Google's Pro-line model for harder reasoning and tool use.
  • Gemini 4 Argon Gemini 4 Argon was announced 30 September 2026 for trusted cyber defenders, not the public, at an intro price of $2 input and $10 output per million tokens.
  • GPT-6 Astra GPT-6 Astra is OpenAI's flagship for hard end-to-end work: reasoning, coding, computer use, research, and documents.
  • Claude Opus 5.5 Claude Opus 5.5 is the first Claude 5.5 model, released 22 September 2026, and Anthropic prices it below Opus 5 for similar work.
  • Claude Sonnet 5.5 Claude Sonnet 5.5 is the second Claude 5.5 model, released 28 September 2026, at $2 per million input tokens and $10 per million output tokens.
  • Claude Haiku 5.5 Claude Haiku 5.5 is the third Claude 5.5 model, released 7 October 2026, and Anthropic's fastest, priced from $0.10 per million input tokens and $0.50 per million output tokens.
  • Muse Spark Muse Spark is Meta's flagship multimodal model for agents, coding, and Meta AI, with 1.3 as the current developer version.
  • Jev Jev returns a typed decision a program can branch on.
  • AI glossary Twenty plain-English AI terms so a non-expert can follow the rest of the /ai door without hunting definitions elsewhere.

Force majeure

Contract outs that let a committed AI campus slip when power, permits, or fuel fail.

  • AI build-out contagion A slip in AI campus building can travel to chip stocks and then to the whole market, and three live numbers track each step.

Power and energy

The electricity, gas, and permits a training campus has to hold before the models run.

  • AI build-out contagion A slip in AI campus building can travel to chip stocks and then to the whole market, and three live numbers track each step.
  • Memphis / Colossus Memphis Colossus is a hyperscale training campus. The company page lists 170 PB/s of aggregate memory bandwidth and 2.8 Tb/s per server.
  • What xAI has running SpaceX told investors on 12 June 2026 that Colossus and Colossus II together provide about 1 gigawatt of compute power, and the Colossus page lists 170 PB/s and 2.8 Tb/s.
  • Meta stack Meta's stack is MTIA, a Broadcom 2nm accelerator through 2029, millions of Nvidia Blackwell and Rubin GPUs, and up to 6 gigawatts of AMD Instinct.
  • Prometheus Prometheus in New Albany, Ohio is over 1 gigawatt and tens of thousands of graphics chips once complete.
  • Hyperion Hyperion in Richland Parish, Louisiana is published at 5 gigawatts and more than $50 billion.
  • Meta nuclear package Meta's January 2026 nuclear package is up to 6.6 gigawatts by 2035.

Data-center financing

How AI campuses raise money through leases and private placements before the revenue arrives.

  • AI build-out contagion A slip in AI campus building can travel to chip stocks and then to the whole market, and three live numbers track each step.

Private credit

Loans and private placements that fund data-center builds outside the public bond market.

  • AI build-out contagion A slip in AI campus building can travel to chip stocks and then to the whole market, and three live numbers track each step.

AI capex

Capital spending on AI campuses that has to be financed before the revenue shows up.

  • AI build-out contagion A slip in AI campus building can travel to chip stocks and then to the whole market, and three live numbers track each step.

Treasury yields

What the Treasury curve charges to borrow for 2, 10, and 30 years.

  • Oil and the $100 line Oil was above $100 at the Fed's first hike on 16 Sep 2026, and this brief tracks WTI against that line and the 10-year yield against 5% with live numbers from the watch-list feed.
  • The 30-year yield holds near 6% The 30-year Treasury yield sits below 6%, and this brief tracks whether it reaches that level and holds through a long-term borrowing date, with live numbers from the watch-list feed.
  • Debt refinancing Treasury borrowing, 10-year auction demand, and mortgage rates form one chain, and this brief tracks each step with live numbers from the watch-list feed.

Fed policy

The short-term rate the Fed sets, and the inflation and geopolitics it names when it moves.

  • Debt refinancing Treasury borrowing, 10-year auction demand, and mortgage rates form one chain, and this brief tracks each step with live numbers from the watch-list feed.
  • Oil and the $100 line Oil was above $100 at the Fed's first hike on 16 Sep 2026, and this brief tracks WTI against that line and the 10-year yield against 5% with live numbers from the watch-list feed.

Debt stock

The stock of federal debt, the pile that comes due, and what inflation does to old fixed claims.

  • Debt refinancing Treasury borrowing, 10-year auction demand, and mortgage rates form one chain, and this brief tracks each step with live numbers from the watch-list feed.
  • The 30-year yield holds near 6% The 30-year Treasury yield sits below 6%, and this brief tracks whether it reaches that level and holds through a long-term borrowing date, with live numbers from the watch-list feed.

Mortgages

The 30-year mortgage rate buyers face when the long Treasury yield is high.

  • Debt refinancing Treasury borrowing, 10-year auction demand, and mortgage rates form one chain, and this brief tracks each step with live numbers from the watch-list feed.
  • The 30-year yield holds near 6% The 30-year Treasury yield sits below 6%, and this brief tracks whether it reaches that level and holds through a long-term borrowing date, with live numbers from the watch-list feed.

Oil prices

Where crude sits when the Fed moves, because that price is already in the room.

  • Oil and the $100 line Oil was above $100 at the Fed's first hike on 16 Sep 2026, and this brief tracks WTI against that line and the 10-year yield against 5% with live numbers from the watch-list feed.

Jobs

Who got hired, whether the pace is slowing, and whether the jobs are for doing the work with AI or only managing it.

  • Payrolls and the jobs mix Payrolls and unemployment show whether hiring is slowing, and the hiring that matters is people who do the work with AI rather than only people who manage that work.

Equity prices

What investors will pay for a dollar of revenue, how far the S&P sits from its peak, and whether this boom already has earnings.

  • Paying up for the boom This boom already has revenue and profit, and the open question is the price investors will pay, tracked against a 10% correction and a 20% bear market from the S&P peak.
  • AI build-out contagion A slip in AI campus building can travel to chip stocks and then to the whole market, and three live numbers track each step.

Model releases

The models and product surfaces the labs have shipped, and what each one is for.

  • Claude Opus 5.5 Claude Opus 5.5 is the first Claude 5.5 model, released 22 September 2026, and Anthropic prices it below Opus 5 for similar work.
  • Claude Sonnet 5.5 Claude Sonnet 5.5 is the second Claude 5.5 model, released 28 September 2026, at $2 per million input tokens and $10 per million output tokens.
  • Claude Haiku 5.5 Claude Haiku 5.5 is the third Claude 5.5 model, released 7 October 2026, and Anthropic's fastest, priced from $0.10 per million input tokens and $0.50 per million output tokens.
  • GPT-6 Luna GPT-6 Luna is the cheaper, faster GPT-6 tier released with Sol on 22 September 2026, at $0.10 and $0.50 per million tokens.
  • GPT-6 Sol GPT-6 Sol is OpenAI's mid-tier GPT-6 model for complex coding and agent work, released 22 September 2026 at $2 and $10 per million tokens.
  • GLM-5.3 GLM-5.3 is Z.AI's flagship coding and agent model, with public weights on Hugging Face.
  • Gemini 3.1 Pro Preview Gemini 3.1 Pro Preview is Google's Pro-line model for harder reasoning and tool use.
  • Gemini 3.5 Flash-Lite Gemini 3.5 Flash-Lite is Google's low-latency model for high-volume parsing, translation, and simple extraction.
  • Gemini 4 Argon Gemini 4 Argon was announced 30 September 2026 for trusted cyber defenders, not the public, at an intro price of $2 input and $10 output per million tokens.
  • Gemini 3.8 Flash Gemini 3.8 Flash is Google's current Flash model for long-horizon software work and agents at Flash speed and cost.
  • Gemma 4 Gemma 4 is Google's open-weight family, separate from the hosted Gemini API.
  • Kimi K3 Kimi K3 is Moonshot's open-weight flagship for long-horizon coding and knowledge work.
  • MiMo-V2.6-Pro MiMo-V2.6-Pro is Xiaomi's flagship open-weight model, announced 22 September 2026 with downloadable weights.
  • Qwen3.8 Max Qwen3.8 Max is Alibaba's lead hosted Qwen model, with an open checkpoint beside it on Hugging Face.
  • Step 5 Preview Step 5 Preview is StepFun's agent model on the API now, and the weights are dated 15 October 2026.
  • API DeepSeek's API speaks the OpenAI and Anthropic request shapes, and the live model ids are deepseek-flash and deepseek-v4-pro.
  • Flash DeepSeek-V4.1-Flash is DeepSeek's current recommended API model, id deepseek-flash, with open weights on Hugging Face.
  • Grok 4.7 Grok 4.7 is xAI's current model for coding, agents, and knowledge work, API id grok-4.7, shipped 21 September 2026.
  • Llama 4 Llama 4 is Meta's current open-weight generation, centered on Scout and Maverick.
  • Muse Code Muse Code is Meta's terminal coding product built for Muse Spark, and it is a product surface rather than a weights release.
  • Muse Glimmer Muse Glimmer is Meta's 30-billion-parameter open-weight model, distilled from Muse Spark, under the Apache 2.0 license.
  • Muse Image Muse Image is Meta's image generation model, wired to Muse Spark and available on Meta AI.
  • Muse Spark Muse Spark is Meta's flagship multimodal model for agents, coding, and Meta AI, with 1.3 as the current developer version.
  • Muse Video Muse Video is Meta's video generation model in the Muse family, still described publicly as a preview.
  • R1 DeepSeek-R1 is the January 2025 open reasoner, released under an MIT license, and this page keeps it as an archive.
  • Thinking DeepSeek thinking mode is on by default for Flash and Pro, and a caller can turn it off or change the effort.
  • V3 DeepSeek-V3 was the prior chat generation, and the old deepseek-chat name is retired.
  • V4 Pro DeepSeek V4 Pro remains on the API as deepseek-v4-pro after 14 September 2026, with billing unchanged until further notice.
  • Chat models write strings. System One returns typed decisions. Chat models are trained to write text people read, and Jev returns a typed decision software can branch on.
  • Jev Jev returns a typed decision a program can branch on.
  • Grok 4.20 Grok 4.20 is listed as three API ids, for reasoning, non-reasoning, and multi-agent use, on the same token prices as Grok 4.3.
  • Grok 4.3 Grok 4.3 is an xAI API model with a 1 million token context window and batch pricing.
  • Grok 4.5 Grok 4.5 is an xAI API model with a 500k context window.
  • Grok 4.6 Grok 4.6 is the prior xAI flagship and is still on the API, with a knowledge cutoff of 1 February 2026.
  • Grok Bot Grok Bot is xAI's product of named bots on a cloud computer, with jobs and context that persist.
  • Grok Build Grok Build is xAI's coding agent for the terminal, headless scripts, and the Agent Client Protocol.
  • Imagine Imagine is xAI's image and video generation family on the API and in consumer Grok.
  • Voice Voice is xAI's spoken interface: speech to speech, speech to text, and text to speech.
  • grok-build-0.1 grok-build-0.1 is a dedicated coding model on the xAI API, separate from the Grok Build product.
  • Claude Fable 5.1 Claude Fable 5.1 is Anthropic's most capable widely released model, built for long-running agents and document work.
  • Claude Haiku 4.5 Claude Haiku 4.5 is Anthropic's fastest near-frontier model, priced for cost-sensitive work and sub-agents.
  • Claude Opus 5 Claude Opus 5 is Anthropic's default model for most agentic coding and enterprise work, at $5 per million input tokens and $25 per million output tokens.
  • Claude Sonnet 5 Claude Sonnet 5 is Anthropic's faster everyday model for coding and agents, at $2 per million input tokens and $10 per million output tokens.
  • GPT-6 Astra GPT-6 Astra is OpenAI's flagship for hard end-to-end work: reasoning, coding, computer use, research, and documents.
  • GPT-6.1 Sol GPT-6.1 Sol is an upgrade to GPT-6 Sol, released 29 September 2026, at $2 input, $0.10 cached input, and $10 output per million tokens.
  • Dots Dots is an always-on ChatGPT agent powered by GPT-6 Astra, with its own cloud computer and browser.
  • GPT-6 and Intelligent UI in ChatGPT Intelligent UI lets GPT-6 answers in ChatGPT include visuals and controls you can tap, rolling out in the Chat tab from 7 October 2026.

Hardware

Where the chips are fabricated, and where the training campus consumes them.

  • Memphis / Colossus Memphis Colossus is a hyperscale training campus. The company page lists 170 PB/s of aggregate memory bandwidth and 2.8 Tb/s per server.
  • Terafab Terafab is a planned chip factory from SpaceX and Tesla in Grimes County, Texas, with a Phase 1 figure of $16.8B.
  • What xAI has running SpaceX told investors on 12 June 2026 that Colossus and Colossus II together provide about 1 gigawatt of compute power, and the Colossus page lists 170 PB/s and 2.8 Tb/s.
  • MTIA MTIA 300 is in production with 216 GB HBM3E, MTIA 400 is a 72-accelerator rack, and compute FLOPS from the 300 to the 500 is 25x.
  • Blackwell GB300 NVL72 is 72 Blackwell Ultra GPUs and 36 Grace CPUs, with 20 TB of GPU memory, 130 TB/s NVLink, and 1,440 PFLOPS of FP4 Tensor Core with sparsity.
  • Rubin The Rubin GPU is 50 PFLOPS of NVFP4 inference and 288 GB of HBM4, and Vera Rubin NVL72 is 72 Rubin GPUs and 36 Vera CPUs.
  • Instinct The custom Instinct GPU is based on the MI450 architecture, and MI300 and MI350 are the series Meta already runs.
  • Ironwood Ironwood is Google's seventh-generation TPU, listed at $12 per chip-hour on demand and $5.40 on a three-year commitment in us-central1, and Trillium (TPU v6e) is the generation before it.
  • Trainium2 Trainium2 is the AI chip Amazon's Annapurna Labs designs for AWS: 1.3 PFLOPS of dense FP8 and 96 GiB of HBM per chip, 64 chips to a Trn2 UltraServer.
  • Maia 200 Maia 200 is Microsoft's own inference chip: TSMC 3nm, over 10 PFLOPS of FP4, 216 GB of HBM3e at 7 TB/s, in a 750 W envelope.
  • Cerebras WSE-3 The Cerebras WSE-3 is one chip the size of a silicon wafer: 4 trillion transistors, 900,000 AI cores, 44 GB of on-chip SRAM, and 125 PFLOPS of peak AI compute.
  • Ascend 910C The Ascend 910C is Huawei's current AI chip. Up to 384 of them form the Atlas 900 A3 SuperPoD, rated by Huawei at up to 307.2 PFLOPS of FP16, with 128 GB of memory per chip.
  • Meta and Nvidia Nvidia says the Meta partnership deploys millions of Blackwell and Rubin GPUs, including GB300 systems, plus Grace CPUs.
  • Meta and AMD AMD will supply up to 6 gigawatts of Instinct GPUs. The first gigawatt is a custom MI450, with Helios racks. Meta already runs MI300 and MI350.
  • Meta and Broadcom Broadcom will deliver custom accelerator technology for Meta's MTIA program through 2029. The first phase is more than 1 gigawatt.
  • Meta stack Meta's stack is MTIA, a Broadcom 2nm accelerator through 2029, millions of Nvidia Blackwell and Rubin GPUs, and up to 6 gigawatts of AMD Instinct.
  • Prometheus Prometheus in New Albany, Ohio is over 1 gigawatt and tens of thousands of graphics chips once complete.
  • Hyperion Hyperion in Richland Parish, Louisiana is published at 5 gigawatts and more than $50 billion.
  • Eagle Mountain Eagle Mountain in Utah was raised to more than $3 billion on 14 September 2026.
  • Meta nuclear package Meta's January 2026 nuclear package is up to 6.6 gigawatts by 2035.

Training loop

A language model as a next-token predictor, shaped by staged training and a tool loop.