Text-to-image models are now bought like any other API primitive: metered per image or per megapixel, in the $0.01–$0.15 band, with editing and multi-reference composition as table stakes rather than separate products. The 2026 shape of the market is a top tier of reasoning-based multimodal models from OpenAI, Google, Microsoft and Meta that plan a layout before rendering and therefore get text and instruction-following right; a middle tier of specialist models (Ideogram for typography, Recraft for vector/brand systems, FLUX for open weights and self-hosting); and a shrinking tail of standalone image startups.
19 entries
Text-to-video and image-to-video models turn a prompt, a still frame, or reference footage into a short clip — now routinely with a synchronized audio track generated in the same pass. As of September 2026 the practical unit of work is still a 5–15 second shot, with Alibaba's Wan 3.0 and ByteDance's Seedance 2.5 pushing to 30 seconds; nobody generates a coherent multi-minute scene in one call.
18 entries
This category covers three related but commercially distinct markets: text-to-speech (TTS), speech-to-text (STT/ASR), and generative music. TTS and STT are now commodity-priced infrastructure — batch transcription has fallen to roughly $0.18–$0.30 per audio hour and streaming TTS to $0.03–$0.10 per 1,000 characters — so the differentiators have shifted to time-to-first-audio, end-of-turn detection, and language coverage rather than raw accuracy or MOS.
20 entries
Embedding models turn text (and increasingly images, audio and video) into dense vectors so that semantic similarity becomes a distance computation; rerankers are cross-encoders that take a query plus a candidate list and re-score them jointly, trading throughput for precision. Nearly every RAG or semantic-search stack uses both: a cheap embedding model to retrieve 50-200 candidates from a vector index, then a reranker to cut that to the 5-10 chunks that go in the prompt.
20 entries
Hosted inference vendors run open-weight models on their own hardware and sell them as a metered API, so you get open-model economics and licence freedom without operating GPUs. It splits into three genuinely different products that get lumped together: per-token serverless APIs where you pay for tokens and share capacity with everyone else; dedicated endpoints where you rent reserved GPU capacity but still get a managed serving stack; and raw GPU/serverless-container clouds where you bring your own vLLM or SGLang.
18 entries
Cloud AI platforms are the managed layers that hyperscalers and data platforms put in front of foundation models: one API endpoint, one billing relationship, and the cloud's existing identity, networking and audit controls. They matter less for model access — most frontier models are reachable directly from their labs — and more for the surrounding plumbing: IAM instead of bearer tokens, VPC/PrivateLink instead of the public internet, regional pinning for GDPR and sector regulators, and reserved capacity so a production workload does not sit behind a shared rate limit.
17 entries
An LLM gateway sits between your application and one or more model providers, exposing a single OpenAI-compatible (and increasingly Anthropic-compatible) endpoint while handling key management, provider failover, retries, rate limits, caching, spend attribution and audit logs. The category has split into three distinct shapes: credit-based marketplaces that also sell you the tokens (OpenRouter, Vercel AI Gateway, Requesty), self-hosted proxies you run yourself and point at your own provider keys (LiteLLM, Bifrost, Envoy AI Gateway, APISIX, Helicone AI Gateway), and API-management platforms that added LLM plugins to an existing enterprise gateway (Kong, Azure API Management, Databricks).
20 entries
A vector database stores high-dimensional embeddings and answers approximate-nearest-neighbour (ANN) queries over them, almost always combined with metadata filters and increasingly with lexical BM25 scoring in the same query. The category has bifurcated since 2024: purpose-built engines (Pinecone, Qdrant, Weaviate, Milvus, Turbopuffer) on one side, and vector indexes bolted into systems you already run (Postgres/pgvector, Elasticsearch, MongoDB Atlas, Redis, OpenSearch) on the other.
19 entries
Agent frameworks sit between your application code and a model API, supplying the tool-calling loop, conversation state, streaming plumbing and multi-agent orchestration you would otherwise hand-write. In 2026 the category has visibly split into three shapes: low-level orchestration runtimes (LangGraph, LlamaIndex Workflows, Microsoft Agent Framework workflows) that give you explicit control over execution and persistence; batteries-included "harnesses" (Deep Agents, Claude Agent SDK, Strands, Agent Framework's Harness Agent, AI SDK's HarnessAgent) that ship a pre-tuned long-horizon agent with planning, a filesystem and context compaction; and thin typed loops (Pydantic AI, OpenAI Agents SDK, smolagents) that stay close to the raw API.
20 entries
This category covers tools that write, edit, run and review code on a developer's behalf — from terminal agents and IDE forks to browser app-builders and cloud agents that open pull requests unattended. As of September 2026 the market has consolidated hard: Cognition folded Windsurf into Devin Desktop, Cursor acquired Continue and was itself acquired by SpaceX, Roo Code's extension shut down and was archived, and Aider has not shipped a release in over a year.
20 entries
LLM observability and evaluation tools capture what an LLM app or agent actually did — the prompts, tool calls, retrieved documents, token counts, latency and cost of every span — and then let you score that behaviour, either offline against a dataset or online against live traffic. They exist because the usual software signals (exit codes, HTTP status, p99 latency) say nothing about whether an answer was correct, grounded, or safe, and because a one-line prompt change can silently degrade quality across thousands of requests.
20 entries
This category covers the general-purpose chat assistants and answer engines that ordinary people and knowledge workers use directly, as opposed to APIs or coding agents. As of September 2026 the market has consolidated around five subscription ladders (OpenAI, Anthropic, Google, Microsoft, xAI) priced from roughly $5 to $300 per month, plus a second tier of answer engines (Perplexity), agentic "do the work" products (Manus, Genspark, Kimi), and cheap or free open-weight-backed challengers (DeepSeek, Mistral).
19 entries
Every model launch cites benchmark numbers, and almost every one of those numbers is less informative than it looks. A benchmark is a fixed sample of tasks with a scoring rule; a leaderboard adds a harness, a prompt, a scaffold, a sampling budget and a submitter's incentive to look good.
18 entries
The AI stack spent 2023-2024 with no interop layer at all: every tool integration, every agent handoff, every trace format was bespoke. Since late 2024 a set of standards has emerged to fix that, and by 2026 most of the important ones sit under the Linux Foundation rather than a single vendor — MCP is now "a Series of LF Projects, LLC", A2A joined the Agentic AI Foundation in August 2026, AGENTS.md is stewarded there too, and x402 has its own LF foundation.
20 entries
This category covers the three services that sit between raw sources and an LLM's context window: document parsers that turn PDFs, scans and Office files into markdown or JSON; web scrapers that turn URLs into clean text while handling JavaScript and anti-bot defences; and search APIs that return ranked, LLM-ready results or grounded answers. They are the highest-variance part of most RAG and agent stacks — a parser that mangles a table or a search API with a stale index degrades everything downstream, and neither failure shows up in your model evals.
19 entries
Fine-tuning is how you get a smaller, cheaper, or more obedient model to beat a large general model on one narrow task — format adherence, domain jargon, classification, tool-call shape, latency and cost per request. The landscape splits three ways: managed API services that take a JSONL file and hand back a served model, open-source trainer frameworks you run on GPUs you rent, and the raw GPU clouds underneath them.
20 entries
This category covers the runtime controls that sit between users, LLMs and tools — classifiers that score prompts and completions for harm, prompt-injection and jailbreak detectors, PII redaction, groundedness/hallucination checks and topic restriction — plus the governance frameworks that auditors and regulators actually ask about. The market has split into three tiers: free or near-free classifier models you host yourself (Llama Guard, Granite Guardian, ShieldGemma), metered cloud moderation APIs bundled with the hyperscaler you already use (Bedrock Guardrails, Azure AI Content Safety, Model Armor, OpenAI's free moderation endpoint), and enterprise AI-security platforms that combine runtime guardrails with red teaming and model-supply-chain scanning.
19 entries
Agent frameworks produce code; something has to run it. A sandbox is the isolated compute environment where model-generated shell commands, Python, and whole dev servers execute without touching your production hosts, your other tenants, or your cloud credentials.
18 entries
This is the software that turns a folder of open weights into an HTTP endpoint. It is a different category from hosted "inference platforms": here you own the GPUs, and the server you pick sets tokens/sec per GPU, which is the entire cost model for self-hosting.
18 entries
These are the products you buy instead of assembling a parser, an embedding model, a vector store, a reranker and a connector fleet yourself. They split into three shapes that are frequently confused: developer retrieval APIs you call from your own app (Vectara, Contextual AI, Ragie, LlamaCloud, Bedrock Knowledge Bases), end-user work assistants with a search and chat UI already built (Glean, Gemini Enterprise, Microsoft 365 Copilot, Onyx, Dashworks), and hyperscaler index services that sit halfway between (Azure AI Search, Google Agent Search).
18 entries
A speech model list stops where the hard part starts. Shipping a voice agent means solving voice activity detection, semantic endpointing, barge-in, sub-second round trips, WebRTC or WebSocket transport, SIP trunking and PSTN bridging, plus reconnects, jitter and concurrency — none of which an STT or TTS API gives you.
18 entries
Agent memory tooling answers two questions a raw vector database does not: what is worth writing down from a conversation, and what small subset of that gets injected into the next prompt. Products in this category typically do some combination of fact extraction from message streams, conflict resolution when a stored fact goes stale, per-user/per-agent/per-session scoping, and a retrieval step that returns a bounded context block rather than a pile of chunks.
17 entries
This category covers the companies that supply the human side of model training: annotated datasets, supervised fine-tuning demonstrations, preference and reward signal for RLHF, RL environments, red-teaming, and human evaluation panels. It splits into three structurally different things that buyers routinely conflate — managed expert-data services that deliver finished datasets (Surge, Scale, Snorkel, Toloka, Invisible), talent marketplaces that rent you vetted humans while you run the pipeline (Mercor, Handshake AI, Prolific, Alignerr), and annotation software you staff yourself (Label Studio, Encord, SuperAnnotate, Argilla).
18 entries
AI red teaming is the offensive counterpart to runtime guardrails: before an LLM application ships, someone deliberately attempts prompt injection, jailbreaks, tool abuse, authorisation bypass and data exfiltration against the real system, and writes down what worked. The tooling splits into three shapes — open-source frameworks you script and run in CI, commercial SaaS platforms that maintain an attack library and produce management reporting, and human-led penetration tests that produce a signed, scoped report.
18 entries