# AI Model Reference Guide
URL: https://research-memex.org/docs/implementation/core-references/ai-model-reference-guide
Description: Compare current AI model families, understand reasoning and sampling controls, and choose models for research reasoning, writing, and analysis tasks
*The available models change fast; the task-fit question stays constant.*
## Overview: current, but not permanent [#overview-current-but-not-permanent]
AI model names, aliases, pricing, rate limits, and control parameters change quickly. We spot-checked this page against provider documentation on July 27, 2026. Before building a workflow around a specific model ID, check the provider's current model list.
The practical question is not "which model is best?" It is which model, access path, and control surface fit the task:
* Reasoning depth: Can it handle theoretical tension and multi-step analysis?
* Context capacity: Can it read the relevant corpus without destructive chunking?
* Writing behavior: Does it preserve academic nuance and voice?
* Cost and latency: Can you afford to iterate?
* Control surface: Does it want reasoning effort, thinking level, prompt constraints, or sampling parameters?
## Which AI Model Fits This Task? [#which-ai-model-fits-this-task]
Research Memex currently separates worker choice from coordination choice.
Good starting points:
* Claude Opus 5 or Claude Sonnet 5 for theory, critique, nuanced writing, and focused implementation.
* GPT-5.6 Sol or Kimi K3 when the same model may also need to coordinate later.
* Gemini 3.6 Flash or Gemini 3.1 Pro preview for large-context synthesis.
* GLM-5.2 for long-context coding-agent reasoning and Chinese-English synthesis.
Use cases: theory building, critical analysis, synthesis, methodology design, and one scoped coding task.
Good starting points:
* Claude Fable 5 for long-running coordination.
* GPT-5.6 Sol or Kimi K3 when the coordinator must also do substantive work itself.
Use cases: parallel reviews, specialist panels, multi-stage research workflows, and sessions that alternate between delegation and implementation.
Good starting points:
* Gemini 3.6 Flash or Gemini 3.5 Flash-Lite for high-volume Gemini API work.
* DeepSeek V4-era models for cheap screening and analysis.
* GPT-5.6 Terra or Luna for lower-cost OpenAI work.
* Qwen3.7-Plus or Qwen3.6-Flash where the deployment mode supports it.
Use cases: initial screening, pattern detection, summarization, and tag generation.
Good starting points:
* Kimi K2.7 Code for focused coding and adversarial review.
* Claude Haiku 4.5 for quick Claude-family tasks.
* Grok 4.5 for xAI tool-calling and coding workflows.
* Local/open-weight models for privacy-sensitive or reproducible workflows.
Use cases: second opinions, quick lookups, non-English literature, and private corpora.
Working rule: start with cheaper or free access for exploration, then move the strongest prompt and evidence set to the model you trust for final reasoning.
See the [AI Model Reference Guide](/docs/implementation/core-references/ai-model-reference-guide) for the full comparison.
## Current model families [#current-model-families]
OpenAI GPT Family
* Current provider-native IDs: `gpt-5.6-sol`, `gpt-5.6-terra`, and `gpt-5.6-luna`. The `gpt-5.6` alias routes to Sol.
* Context: roughly 1M on Sol and Terra; 400K on Luna.
* Control surface: `reasoning.effort` where supported; sampling defaults are usually enough unless a specific API task needs tuning.
* Best for: Sol for difficult reasoning, implementation, and coordination; Terra for balanced work; Luna for high-volume work.
* Source: [OpenAI latest-model guide](https://developers.openai.com/api/docs/guides/latest-model)
Anthropic Claude Family
* Current provider-native IDs: `claude-fable-5`, `claude-opus-5`, and `claude-sonnet-5`; Claude Haiku 4.5 remains the faster, lighter option.
* Context: 1M on Fable 5, Opus 5, and Sonnet 5.
* Control surface: use `effort` and thinking controls where documented. Fable 5 always uses adaptive thinking; keep sampling defaults unless the current model page says otherwise.
* Best for: Fable 5 for long-running agent coordination; Opus 5 and Sonnet 5 for bounded analysis, writing, and implementation tasks.
* Source: [Claude model docs](https://platform.claude.com/docs/en/about-claude/models/overview)
Kimi Family (Moonshot AI)
* Current provider-native ID: `kimi-k3`. Kimi K2.7 Code and K2.7 Code Highspeed remain coding specialists; Kimi K2.6 remains a lower-cost general option.
* Context: 1M on Kimi K3; 256K on the current K2 models.
* Control surface: Kimi K3 always thinks and accepts `reasoning_effort` values `low`, `high`, and `max`; the default is `max`. Omit sampling parameters.
* Best for: Kimi K3 as either a direct worker or a multi-model coordinator; K2.7 Code for focused coding and review.
* Source: [Kimi K3 quickstart](https://platform.kimi.ai/docs/guide/kimi-k3-quickstart)
Google Gemini Family
* Current anchor: Gemini 3.6 Flash and Gemini 3.5 Flash-Lite stable; Gemini 3.1 Pro remains a preview option.
* Context: Gemini handles large contexts well; exact limits and rate limits vary by tier and model.
* Control surface: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and future releases deprecate `temperature`, `top_p`, and `top_k`; use `thinking_level` where relevant.
* Best for: large-corpus synthesis, volume processing, multimodal tasks, Google platform workflows.
* Source: [Gemini latest-model guide](https://ai.google.dev/gemini-api/docs/latest-model)
DeepSeek Family
* Current anchor: `deepseek-v4-pro` and `deepseek-v4-flash`.
* Migration note: use the V4 model IDs directly. This guide no longer carries the retired compatibility aliases.
* Control surface: current V4 models use thinking mode plus `reasoning_effort`; the default thinking mode is enabled.
* Best for: high-volume exploration, budget-sensitive screening, reasoning checks where cost matters.
* Source: [DeepSeek pricing/model docs](https://api-docs.deepseek.com/quick_start/pricing)
GLM Family (Z.ai / Zhipu AI)
* Current anchor: GLM-5.2.
* Context/output: 1M context and up to 128K output.
* Control surface: GLM-5.2 supports `reasoning_effort`; default temperature is `1.0` if you set sampling at all.
* Best for: long-context coding-adjacent reasoning, multilingual Chinese-English synthesis, frontier-adjacent open-weight experiments.
* Source: [GLM-5.2 docs](https://docs.z.ai/guides/llm/glm-5.2)
Qwen Family (Alibaba)
* Current anchor: Qwen3.7-Max and Qwen3.7-Plus, with Qwen3.6-Flash as the lower-cost current Flash option.
* Context: current Max and Plus entries support up to 1M in the documented deployment modes.
* Lineup note: Max is the difficult-task model; Plus balances capability, speed, and cost; Flash is the volume option.
* Caveat: model names and availability differ by international, global, US, EU, Hong Kong, and Chinese mainland deployment modes.
* Best for: multilingual work, open-source comparisons, region-specific provider setups.
* Source: [Alibaba Model Studio model list](https://www.alibabacloud.com/help/en/model-studio/models)
xAI Grok Family
* Current anchor: Grok 4.5 for general and coding work.
* Context: Grok 4.5 lists a 500K-token context window.
* Control surface: `reasoning_effort` supports `low`, `medium`, and `high`; the default is `high`.
* Best for: agentic tool calling, coding, broad general tasks, xAI platform experiments.
* Source: [xAI model docs](https://docs.x.ai/developers/models)
Local / Open-Weight Models
* Current anchor: depends on your runtime, hardware, and privacy constraints.
* Best for: sensitive data, reproducibility, offline work, customization.
* Caveat: local quality depends more on deployment details than provider marketing pages.
## Access paths [#access-paths]
| Access path | Best for | Notes |
| ----------------------------- | ----------------------------------------- | ------------------------------------------------------------- |
| Provider web app | Fast exploration | Good for trying a model before wiring API access. |
| Google AI Studio | Gemini API keys and free-tier experiments | Use for direct Gemini model access, not Antigravity CLI auth. |
| Cherry Studio | GUI model comparison and knowledge bases | Best first interface for many researchers. |
| OpenCode | CLI-side provider comparison | Useful when model choice is part of the method. |
| Vox MCP | Multi-model access inside an MCP client | Pure passthrough; good for challenge and triangulation. |
| Claude Code / Antigravity CLI | Agentic project work | Product sign-in flows, not generic provider-key clients. |
## Configuration: defaults first [#configuration-defaults-first]
#### Sampling is no longer the default lever [#sampling-is-no-longer-the-default-lever]
The old advice was to tune `temperature` for every model. That is now misleading. Many current reasoning models either reject sampling changes, ignore them, or explicitly recommend defaults.
Default rule: omit `temperature`, `top_p`, and `top_k` unless the provider currently documents that the model supports and benefits from changing them.
Provider-specific notes:
| Provider/model | Recommendation |
| --------------------------------- | --------------------------------------------------------------------------------------------------- |
| Claude 5 family | Keep sampling defaults; use `effort` and the model's documented thinking controls. |
| Gemini 3.6 Flash / 3.5 Flash-Lite | Remove deprecated `temperature`, `top_p`, and `top_k`; use `thinking_level` and prompt constraints. |
| Kimi K3 / K2.7 / K2.6 | Omit sampling parameters; use `reasoning_effort` on K3. |
| GLM-5.2 | Default temperature is `1.0`; prefer `reasoning_effort` for thinking control. |
| DeepSeek V4 | Use thinking mode and `reasoning_effort`; keep other controls at provider defaults. |
| xAI Grok 4.5 | Prefer `reasoning_effort`; watch incompatible parameters on reasoning models. |
If a compatible model requires a temperature and the provider offers no task-specific guidance, start at `1.0`. Go lower only for mechanical extraction or classification where repeatability matters.
#### Prefer reasoning controls when available [#prefer-reasoning-controls-when-available]
| Control | Providers/examples | Use for |
| -------------------------------------------------- | ------------------------------------------- | ---------------------------------------------------- |
| `reasoning.effort` / `reasoning_effort` / `effort` | OpenAI, Claude, Kimi K3, xAI, GLM, DeepSeek | More or less deliberate reasoning. |
| `thinking_level` | Current Gemini models | Gemini reasoning depth without raw sampling changes. |
| thinking mode | Kimi K2.6, DeepSeek V4, GLM | Choosing thinking behavior where supported. |
| Sequential Thinking MCP | Any MCP-capable client | Tool-visible, revisable reasoning steps. |
Use high reasoning effort for theoretical integration, methods design, and hard debugging. Use low/default effort for routine extraction, tagging, and quick checks.
| Task type | Preferred control | Suggested starting point |
| -------------------- | --------------------------------------- | --------------------------------------------------------- |
| Theory synthesis | Reasoning effort / thinking level | Medium or high; keep sampling defaults. |
| Framework building | Reasoning effort plus explicit criteria | High; ask for alternatives and tensions. |
| Literature synthesis | Context capacity plus structured prompt | Medium/high thinking; defaults for sampling. |
| Draft writing | Prompt constraints and examples | Defaults first; ask for voice and structure directly. |
| Citation extraction | Deterministic prompt rules | Low reasoning or provider-specific low temp if supported. |
| Systematic coding | Rubric and validation pass | Low/medium reasoning; prioritize consistency checks. |
## Strategic model usage [#strategic-model-usage]
Use cheap or free access paths to learn model behavior against your own corpus.
* Start with Google AI Studio, Cherry Studio, OpenCode, or OpenRouter.
* Run the same prompt across at least three model families.
* Record failure modes, not just best answers.
* Promote a model only after it handles your real material.
| Work shape | Good starting models | Why |
| ------------------------------------------------ | ------------------------------------------------- | ------------------------------------------------------------------- |
| One bounded theory, critique, or writing task | Claude Opus 5 or Claude Sonnet 5 | Worker-first choices for a single chat task. |
| Multi-model coordination | Claude Fable 5, GPT-5.6 Sol, or Kimi K3 | Better starting points when the session must direct several models. |
| Alternating between coordination and direct work | GPT-5.6 Sol or Kimi K3 | Both are strong coordinators and direct workers. |
| Large literature sets | Gemini 3.6 Flash, Gemini 3.1 Pro preview, GLM-5.2 | Large context and long-horizon reasoning. |
| Budget-conscious screening | DeepSeek V4, Gemini Flash/Lite, Qwen Plus/Flash | Cheap enough to iterate. |
* For one draft or revision, choose a strong worker and keep the task bounded.
* For several independent readings or reviews, use a coordinator to preserve the brief and consolidate disagreements.
* Use a second model family for pressure-testing, not merely for another polished version of the same answer.
* Treat large-context synthesis as preparation for human rewriting, not as the finished argument.
## Free and low-cost options [#free-and-low-cost-options]
Google AI Studio remains the easiest free/low-cost backup for Gemini API work, but limits change often. Check [Gemini rate limits](https://ai.google.dev/gemini-api/docs/rate-limits) before relying on a quota for teaching or batch work.
Other cost-management paths:
* Use OpenCode or Cherry Studio to compare providers before committing to a long run.
* Use DeepSeek or Qwen for high-volume screening when quality is sufficient.
* Use local models when privacy matters more than frontier quality.
* Reserve expensive models for the final synthesis, argument, or validation pass.
## Next step [#next-step]
The reference above tells you what is available. The companion page tells you how to test those choices against your own materials: [AI Model Discovery Protocol](/docs/implementation/core-references/ai-model-discovery-protocol).