DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///
DARC.DEV

[ REF — The field glossary ]

Understand the term. Avoid the trap.

Thirty compact definitions for people using and building AI. Each makes a practical distinction—not a claim that the terminology is standardized across providers.

30 terms · definitions for practical use

Token

A model-specific unit of input or output representation. It is not a universal word count; measure with the selected tokenizer or usage report.

Context window

The capacity available to a request. Some providers quote a total window, others separate input and output limits.

Inference

Running a trained model on inputs to produce outputs. It has latency, memory and compute costs even when the weights are downloadable.

Reasoning effort

A provider-specific control over reasoning behavior or computation. Labels are not standardized across models.

Prompt cache

Reuse of computed prefix state. It is not a saved final answer and does not guarantee deterministic output.

Cache write

Input work stored for possible reuse. Depending on the provider it has a distinct billing rate, not necessarily an additive fee.

KV cache

Attention-related runtime state associated with processed context. It consumes memory and is distinct from model weights.

RAG

Retrieval-augmented generation: selecting external evidence and supplying it to a model at inference time.

Embedding

A numeric representation used for similarity and retrieval. Similarity does not prove truth or authorization.

Reranking

A second-stage relevance assessment over retrieved candidates. It adds its own cost and latency.

Hybrid search

Combining retrieval signals such as lexical matching and semantic vectors. It must be evaluated on real query types.

Grounding

Connecting an answer to supplied or retrieved evidence. The evidence itself still needs quality and permission checks.

Structured output

Output constrained to a schema or format. It can be syntactically valid and semantically incorrect.

Tool call

A model-proposed request for an external operation. The application must validate and authorize it before execution.

Agent

A system that repeatedly uses model decisions, tools and observations toward a goal, with explicit limits and stop states.

Harness

The software surrounding a model: prompts, tools, execution loop, retries, context management and controls.

MCP

Model Context Protocol: a standardized connection between hosts/clients and tool or resource servers. It does not replace application authorization.

Prompt injection

An attempt to make untrusted content influence behavior outside its intended role, often by posing as instructions.

Sandbox

An execution boundary restricting resources or actions. It limits reach, not necessarily business-logic mistakes.

Idempotency

A property that repeated execution of the same logical operation does not repeat its external effect.

Checkpoint

Persisted workflow state used for recovery or human review. It does not prove that an external write happened exactly once.

Eval

A defined test of behavior against a rubric or expected result. Its value depends on the task set, judge and run conditions.

LLM-as-judge

Using a model to grade outputs. The judge needs calibration, bias checks and its own observability.

Pass rate

Accepted cases divided by attempted cases under a stated rubric. It is not a universal measure of safety or intelligence.

Cost per accepted task

Total cost of all evaluated attempts divided by accepted results. Undefined when no results are accepted.

Open weights

A released model checkpoint under a specific license. Not a promise of unrestricted use, open training data or free serving.

Quantization

Representing weights or runtime state with reduced precision to change memory and performance trade-offs.

MoE

Mixture of Experts: an architecture that activates part of a larger model per token. Active parameters are not the full storage requirement.

TTFT

Time to first token. It is only one part of end-to-end user latency; tools, queueing and speech output can dominate.

Observability

Traceable evidence of what a system did, how long it took and what it cost. It does not by itself judge correctness.