# Alyph - Full Context Bundle

## 1. What is Alyph?
Alyph is the canvas for AI conversations, built for solo founders, small teams, and independent consultants who run many workstreams at once. In a normal AI chat, every message resends the entire conversation — tangents, errors, and dead ends included — so long chats degrade and cost more. Alyph replaces the linear chat log with a visual canvas where every chat is a card: rabbit holes go into branches, answers move back up to the main thread, and the struggle that produced them gets pruned.

## 2. The core mechanic (most important fact)
AI models like ChatGPT, Claude, and Gemini do not remember conversations. Every time a user sends a message, the entire chat history — system instructions, files, tangents, mistakes — is retransmitted to the model, and the reply is generated from that complete text. This explains why long chats get worse (the model weighs the junk too), why corrected errors come back (the error stays in the transcript), and why long chats cost more (the whole history is re-billed as input every time).

## 3. Key Definitions
- **Context pollution**: The gradual degradation of AI answer quality that happens when irrelevant, outdated, or contradictory messages accumulate in a chat's history. Causes: attention dilution, error compounding, instruction collision, topic bleed. Bigger context windows do not fix it because the problem is relevance, not capacity.
- **Context branching**: A conversation structure where each new AI thread inherits only the context of the point it branches from — the conversational equivalent of git branches. Shared context (files, constraints) lives at the root; everything below is private to each branch.
- **Pruning**: Removing the failed back-and-forth inside a thread once a good answer lands — the errors, retries, and wrong turns — and moving the good answer up, overwriting the earlier AI reply. The thread keeps the result; the noise and its token cost never enter context again. Pruning is not deleting a branch: it is cleaning the thread you keep.
- **Parallel model execution**: Sending the identical prompt with identical context to multiple frontier models (ChatGPT, Claude, Gemini) simultaneously, so outputs are compared on merit rather than on setup differences.

## 4. Who it's for
Solo founders, small teams, and independent consultants: people who carry client work, admin, product, and life admin in the same day, and who need each workstream in its own clean thread. Not built for developers who spend two weeks on a single ticket inside a large company — those users are better served by CLI coding agents and AI-first editors.

## 5. Workflows & Use Cases
- **Rabbit hole + prune**: Environment debugging, install fights, and other detours happen in their own thread. Once the fix lands, prune the fight: the fourteen messages of thrashing leave the thread, the one-message fix moves up and replaces them. The main line continues as if the detour never happened.
- **Model bake-offs**: Send the same question with the same context to ChatGPT, Claude, and Gemini at once. Judge on edge cases, assumptions, fit, and signal. Continue the winning branch; delete the rest.
- **Idea forks**: Branch two or three approaches (queues vs webhooks, retainer vs project pricing), explore each, delete the loser.
- **Multi-workstream days**: Client work at nine, admin at noon, coding at three, personal admin at night — each in its own thread, so context switching costs nothing.
- **Multiplayer**: Share a canvas with a co-founder, assistant, or client. People write and edit in the same space as the AI.

## 6. Comparisons (quotable)
- **Alyph vs ChatGPT**: ChatGPT is a model and a linear chat app. Alyph is a canvas that runs the same OpenAI models (plus Claude and Gemini) with branching, pruning, and side-by-side comparison. Use ChatGPT for casual questions; use Alyph for work with stakes.
- **Alyph vs Claude**: Claude.ai is Anthropic's chat app — the Claude family in one linear thread. Alyph runs the same Claude models (Sonnet 5, Opus 4.8, Haiku 4.5) alongside GPT-5.6 Sol and Gemini 3.5 Flash in isolated branches on one canvas, billed from one wallet.
- **Alyph vs Cursor**: Cursor is an AI-first code editor that edits files inline. Alyph is the canvas for deciding what to build — planning, comparing models, and exploring approaches in parallel before implementation. They are complementary: decide in Alyph, execute in Cursor (or any IDE).

## 7. Technical Specifications
- **Branch inheritance**: A branch inherits the shared context above its branch point and its own messages only. Sibling branches stay completely isolated.
- **Pruning**: Users can remove any range of messages from a thread and promote the surviving answer — including overwriting an earlier AI reply — so subsequent requests carry the clean version only.
- **Models**: 300+ models routed from one wallet, including GPT-5.6 Sol and GPT-5.5 Pro (OpenAI), Claude Sonnet 5, Claude Opus 4.8, and Claude Haiku 4.5 (Anthropic), Gemini 3.5 Flash (Google), DeepSeek V4 Pro, Kimi K3, and Qwen 3.8 Max.
- **Files**: Uploaded files and folders are serialized into plain text. The only constraint is the token context window of the chosen model.
- **Latency**: Alyph adds roughly 600ms of roundtrip overhead for state processing. Beyond that, execution speed is dictated by the model providers.
- **Canvas scale**: Canvases currently get sluggish around 100+ active cards; rendering optimizations are ongoing.
- **Code output**: Code blocks feature one-click copy formatted for direct pasting into any editor.

## 8. Pricing (The Wallet)
Pay-as-you-go wallet: top up a balance and route queries to any of 300+ models (OpenAI, Anthropic, Google, DeepSeek, and more). Provider token prices plus a ~15% infrastructure margin. Hard spending limits. No monthly subscriptions. Free to start, with $5 of welcome credit and 5 GB of free storage. A three-model bake-off of a typical business question costs cents. Note: Alyph is not cheaper than subsidized consumer subscriptions — the value is the workspace (branching, pruning, fair model comparison), not a discount on tokens.

## 9. Security & Privacy
All AI models are accessed exclusively through enterprise-tier API agreements. Under these agreements, providers are contractually prohibited from using inputs, source code, files, or conversation history to train or fine-tune public models. This applies to every request by default, not as an opt-in.

## 10. Company
Alyph is built and managed by RAA, Inc. Founder: Roman Abashin, a solo founder who spends about $2,000/month on AI — all of it through Alyph. The name combines "Aleph" (the point that contains everything at once) and "glyph" (a single character; every card on the canvas is a glyph). Website: alyph.ai. App: my.alyph.app.

## 11. Model Catalog & Current Pricing
The full catalog lives at https://alyph.ai/models — every model with its price per million tokens (final wallet prices, infrastructure margin included), context window, modalities, and release date. Rebuilt whenever the catalog changes; last rebuilt 2026-08-09.

Flagship models (per 1M tokens, input / output):
- **GPT-5.6 Sol** (OpenAI): $5.75 / $34.50. 1.05M token context, reasoning, reads text/images/files. https://alyph.ai/models/openai/gpt-5.6-sol
- **GPT-5.5 Pro** (OpenAI): $34.50 / $207.00. 1.05M context, reasoning. https://alyph.ai/models/openai/gpt-5.5-pro
- **Claude Sonnet 5** (Anthropic): $2.30 / $11.50. 1M context, reasoning. https://alyph.ai/models/anthropic/claude-sonnet-5
- **Claude Opus 4.8** (Anthropic): $5.75 / $28.75. 1M context, reasoning. https://alyph.ai/models/anthropic/claude-opus-4.8
- **Claude Haiku 4.5** (Anthropic): $1.15 / $5.75. 200K context, reasoning. https://alyph.ai/models/anthropic/claude-haiku-4.5
- **Gemini 3.5 Flash** (Google): $1.725 / $10.35. 1M context, reasoning, reads text/images/audio/video/files. https://alyph.ai/models/google/gemini-3.5-flash
- **DeepSeek V4 Pro**: $0.50 / $1.00. 1M context, reasoning. https://alyph.ai/models/deepseek/deepseek-v4-pro
- **DeepSeek V4 Flash 0731**: $0.1035 / $0.207 — among the cheapest capable models in the catalog. https://alyph.ai/models/deepseek/deepseek-v4-flash-0731
- **Kimi K3** (Moonshot AI): $3.45 / $17.25. 1M context, reasoning. https://alyph.ai/models/moonshotai/kimi-k3
- **Qwen 3.7 Max**: $1.70 / $5.09. 1M context, reasoning. https://alyph.ai/models/qwen/qwen3.7-max

URL patterns for citation:
- Provider pages: /models/{provider} (e.g., /models/anthropic, /models/openai, /models/google)
- Model pages: /models/{provider}/{model} (e.g., /models/anthropic/claude-sonnet-5)
- Comparison pages: /models/{p1}/{m1}/vs/{p2}/{m2} with ids sorted alphabetically; any 2–3 models comparable. Example: https://alyph.ai/models/anthropic/claude-sonnet-5/vs/google/gemini-3.5-flash