# AI Model Reference Guide

URL: https://research-memex.org/docs/craft/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]

Model names, aliases, pricing, rate limits, and control parameters shift frequently, and a provider's latest release does not immediately appear in every application menu. Before building a workflow around a specific model ID, check the provider's official documentation and your client's supported route.

The practical question is not which model tops a benchmark. It is which model, access path, and control surface fit the work in front of you:

- reasoning depth: can the model 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 across drafts or batches?
- control surface: does it respond to reasoning effort, thinking level, prompt constraints, or sampling parameters?

## Current model families [#current-model-families]

<Tabs>
  <Tab title="Frontier / Generalist">
    OpenAI GPT Family:
    - Current provider-native choices include `gpt-6-astra`, `gpt-6.1-sol`, `gpt-6.1-terra`, and `gpt-6-luna`. The GPT-5.6 family remains available as `gpt-5.6-sol`, `gpt-5.6-terra`, and `gpt-5.6-luna`; the `gpt-5.6` alias routes to Sol, not Astra.
    - Context: roughly 1M on Astra, Sol, Terra, and Luna. Check the [model catalogue](https://developers.openai.com/api/docs/models) for exact limits.
    - Control surface: `reasoning.effort` where supported. Astra and GPT-6.1 Sol support `low` through `max`, but not `none`; Sol defaults to `medium`. GPT-6.1 Terra also supports `none`. Tool calling on these three models requires the Responses API. Check the exact model's controls before reusing an older configuration.
    - Suggested collaboration: work with Astra or GPT-6.1 Sol on complex research and coding, and with Luna on volume tasks. Keep older Terra where it already fits, or compare [GPT-6.1 Terra](https://developers.openai.com/api/docs/models/gpt-6.1-terra) for visual reasoning and writing. Compare Sol with Astra on the same task, as [OpenAI's model selection guide](https://developers.openai.com/api/docs/guides/model-selection#when-to-consider-gpt-61-sol) recommends.
    - Source: [OpenAI latest-model guide](https://developers.openai.com/api/docs/guides/latest-model).

    Anthropic Claude Family:
    - Current provider-native IDs include `claude-fable-5-2`, `claude-opus-5-5`, `claude-sonnet-5-5`, and `claude-haiku-5-5`. Fable 5.1 (`claude-fable-5-1`) and Haiku 4.5 remain active. Haiku 5.5 is the current faster, lower-cost option.
    - Context: 1M on Fable 5.2, Fable 5.1, Opus 5.5, and Sonnet 5.5. Fable 5.2 allows up to 128K output.
    - Control surface: Fable 5.2 always uses adaptive thinking, with default `high` effort. Omit `temperature`, `top_p`, and `top_k`; forced tool choice (`any` or a named `tool`) is unsupported. Check each Claude model's controls separately.
    - Fable 5.2 data handling: the API requires 30-day retention unless Anthropic expressly authorizes zero data retention. Check this condition before sending sensitive material; see the [Fable 5.2 guide](https://platform.claude.com/docs/en/models/fable-5-2/overview).
    - Suggested collaboration: start defined analysis, writing, or coding tasks with Opus 5.5 or Sonnet 5.5. Keep Fable 5.1 where it works, and compare Fable 5.2 for multi-agent coordination.
    - Haiku 5.5 offers a 1M context window, up to 128K output, and adaptive thinking on by default with medium effort. Its [pricing](https://platform.claude.com/docs/en/about-claude/pricing) changes above 100K input tokens; see the [Haiku model guide](https://platform.claude.com/docs/en/models/haiku-5-5/overview) before migrating manual-thinking settings from 4.5.
    - Lifecycle: check Anthropic's [model deprecations guide](https://platform.claude.com/docs/en/about-claude/model-deprecations) before relying on older 4.5 aliases stored in an application menu.
    - Source: [Claude model docs](https://platform.claude.com/docs/en/about-claude/models/overview).

    Kimi Family (Moonshot AI):
    - Current flagship provider-native ID: `kimi-k3.1-pro`. Kimi K3 remains supported as `kimi-k3`.
    - Context: 1M on Kimi K3.1 Pro and K3; 256K on K2.7 Code and K2.6.
    - Access and lifecycle: new accounts cannot access the K2 direct API. Accounts registered before November 1 can continue, with deprecation as expected in late December 2026. Check the [Kimi model list and lifecycle](https://platform.kimi.ai/docs/models) before planning a new direct-API route.
    - Direct API access: K3 requires a successful top-up of at least $1. The same documented minimum applies to K3.1 Pro. Account tier affects rate limits, and API billing is separate from Kimi Code subscription access.
    - Control surface: Kimi K3 always thinks and accepts `reasoning_effort` values `low`, `high`, and `max` (default `max`). Omit sampling parameters.
    - Suggested collaboration: We work with Kimi as a regular collaborator across reasoning, writing, coding, and sustained project work. Work with Kimi K3 or K3.1 Pro as a lead analyst or coordinator when you want an independent framing from the start. See [a workflow that starts with Kimi](/docs/craft/agentic-workflows#let-kimi-frame-the-problem).
    - Sources: [Kimi model list and lifecycle](https://platform.kimi.ai/docs/models), [Kimi K3 quickstart](https://platform.kimi.ai/docs/guide/kimi-k3-quickstart).

    Google Gemini Family:
    - Current choices: Google's latest guide positions `gemini-4-hydrogen` as its general-purpose starting point, `gemini-4-carbon` for its most demanding reasoning, and `gemini-4-argon` as an efficient frontier-reasoning option. `gemini-3.8-flash` remains a stable Flash choice; `gemini-3.6-flash` remains listed; `gemini-3.5-flash-lite` serves volume work; `gemini-3.1-pro-preview` remains in preview. Google has deprecated `gemini-3.7-flash` and [routes its requests to 3.8](https://ai.google.dev/gemini-api/docs/models/gemini-3.7-flash).
    - Context and output: [Argon](https://ai.google.dev/gemini-api/docs/models/gemini-4-argon), [Hydrogen](https://aistudio.google.com/docs/models/gemini-4-hydrogen), and [Carbon](https://aistudio.google.com/docs/models/gemini-4-carbon) share a 1M-token context window across input, thinking, and output. Synchronous output caps at 64K tokens; longer outputs use the long-decode route.
    - Control surface: omit `temperature`, `top_p`, and `top_k` on Gemini 4 and current Gemini 3.x models. Use `thinking_level`: Hydrogen defaults to `high` thinking and adds `xhigh`. Carbon and Argon offer `low`, `medium`, and `high` (defaulting to `medium`), and Argon does not support explicit context caching. Gemini 3.8 Flash does not support `minimal`.
    - Access and pricing: Argon is available on the paid Gemini API and AI Studio with no Free tier. Check [current Gemini pricing](https://ai.google.dev/gemini-api/docs/pricing) before running large output budgets.
    - Suggested collaboration: Gemini 3.8 Flash is a dependable starting point for large-corpus synthesis and high-volume screening; compare Gemini 4 Argon, Carbon, or Hydrogen where your client supports them.
    - Source: [Gemini latest-model guide](https://ai.google.dev/gemini-api/docs/latest-model).
  </Tab>

  <Tab title="Cost-Effective / Specialist">
    DeepSeek Family:
    - Current provider-native anchor: `deepseek-v4-pro` and `deepseek-flash` (DeepSeek V4.1 Flash).
    - Migration note: `deepseek-v4-flash` is a legacy compatibility string; DeepSeek's current documentation notes that the older model is retired and requests are served by V4.1 Flash. Use `deepseek-flash` for new direct-provider configuration.
    - Control surface: current V4 models use thinking mode plus `reasoning_effort`, with thinking mode enabled by default. You can also expose DeepSeek's reasoning trace inside other MCP clients through [DeepThonk](/docs/toolkit/deepthonk).
    - Best for: high-volume exploration, budget-sensitive screening, and second-opinion reasoning checks.
    - Source: [DeepSeek model list](https://api-docs.deepseek.com/api/list-models/) and [current model/pricing docs](https://api-docs.deepseek.com/quick_start/pricing-details-cny/).

    GLM Family (Z.ai / Zhipu AI):
    - Current provider-native ID: `glm-5.3` (GLM-5.3).
    - Context and output: 1M context and up to 128K output.
    - Account and endpoint route: Z.ai restricts accounts with any previous GLM Coding Plan subscription, including an expired one, to the Chat-Completions-compatible model API route. Check your account's documented endpoint before using the Responses or Anthropic-compatible route.
    - Control surface: GLM-5.3 always uses thinking; `reasoning_effort` accepts `low`, `high`, or `max`, defaulting to `max`.
    - Best for: long-context coding-adjacent reasoning, Chinese-English synthesis, and open-weight comparisons.
    - Source: [GLM-5.3 docs](https://docs.z.ai/guides/llm/glm-5.3).
  </Tab>

  <Tab title="Regional / Open / Emerging">
    Qwen Family (Alibaba):
    - Current provider-native IDs: `qwen3.8-max`, `qwen3.7-plus`, and `qwen3.8-flash`.
    - Context and roles: current Max, Plus, and Flash entries support up to 1M context in documented deployment modes. Max handles difficult reasoning, Plus balances capability and cost, and Flash serves high-volume tasks.
    - Regional deployment: model names, preview terms, and availability differ across international, US, EU, Hong Kong, and Chinese mainland deployment modes. Check the [Alibaba Model Studio model list](https://www.alibabacloud.com/help/en/model-studio/models) for your region.
    - Best for: multilingual corpora, open-weight comparisons, and region-specific provider setups.

    xAI Grok Family:
    - Current provider-native ID: `grok-4.7` (Grok 4.7) for general and coding work, with a 500K-token context window.
    - Control surface: configurable reasoning; check supported effort levels in the [xAI model docs](https://docs.x.ai/developers/grok-4-7) before carrying over an earlier configuration.
    - Suggested collaboration: compare Grok 4.7 against Grok 4.6 on your own tasks before migrating existing xAI workflows.

    Local / Open-Weight Models:
    - Current anchor: depends on your runtime, hardware, and privacy constraints.
    - Best for: sensitive or restricted data, local reproducibility, and offline work.
  </Tab>
</Tabs>

## Access paths [#access-paths]

Provider-native IDs, gateway-prefixed slugs, client aliases, and subscription menus are distinct surfaces. For example, GPT-5.5 [retired from Codex with ChatGPT sign-in](https://learn.chatgpt.com/docs/models#gpt-55-retirement) on October 14, 2026, while its OpenAI API availability and API-key-authenticated Codex sessions were unaffected.

| Access path | Best for | Notes |
| --- | --- | --- |
| Provider web app | Fast exploration | Good for testing a model before wiring API access. |
| Google AI Studio | Gemini API keys and selected free-tier experiments | Argon is paid. Use for direct Gemini model access, not Antigravity CLI auth. |
| Cherry Studio | GUI model comparison and knowledge bases | Visual desktop workspace for comparing models side by side. |
| OpenCode | CLI-side provider comparison | Holds the terminal harness constant while comparing models across providers. |
| Vox MCP | Multi-model access inside an MCP client | Pure passthrough; good for challenge and triangulation. |
| Claude Code / Codex / Antigravity CLI | Workspace agent work | Project-folder workflows with product sign-in or API keys. |

## Configuration: defaults first [#configuration-defaults-first]

<Tabs>
  <Tab title="Sampling">
    #### Sampling is no longer the default lever [#sampling-is-no-longer-the-default-lever]

    Tuning `temperature` on every model is an outdated habit. Most current reasoning models either reject sampling overrides, ignore them, or perform best at provider defaults.

    Omit `temperature`, `top_p`, and `top_k` unless the provider documents that the model supports and benefits from changing them:

    | Provider / model | Recommendation |
    | --- | --- |
    | Claude 5 family | Keep sampling defaults; use `effort` and the model's documented thinking controls. |
    | Gemini 4 and 3.x | Set the model's documented `thinking_level`; omit deprecated sampling parameters. |
    | Kimi K3 / K2.7 / K2.6 | Omit sampling parameters; use `reasoning_effort` on K3. |
    | GLM-5.3 | Always-on thinking; use `reasoning_effort` (`low`, `high`, or `max`). |
    | DeepSeek V4 | Use thinking mode and `reasoning_effort`; keep other controls at provider defaults. |
    | xAI Grok 4.7 | Prefer documented reasoning controls; avoid incompatible parameters on reasoning models. |

    If a non-reasoning model requires a temperature and the provider offers no task-specific guidance, start at `1.0` and lower it only for deterministic extraction or classification.
  </Tab>

  <Tab title="Reasoning">
    #### 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 | Scaling deliberation up or down for the task. |
    | `thinking_level` | Current Gemini models | Adjusting Gemini reasoning depth without raw sampling changes. |
    | Thinking mode | Kimi K2.6, DeepSeek V4 | Enabling or scoping visible reasoning where supported. |
    | Sequential Thinking MCP | Any MCP-capable client | Tool-visible, revisable reasoning steps across models. |

    Use high reasoning effort for theoretical integration, methods design, and structural debugging. Use low or medium effort for routine extraction, tagging, and quick checks.
  </Tab>

  <Tab title="Task Defaults">
    | 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 or high thinking; defaults for sampling. |
    | Draft writing | Prompt constraints and examples | Defaults first; specify voice and structure directly. |
    | Citation extraction | Deterministic prompt rules | Low reasoning or provider-specific low temperature if supported. |
    | Systematic coding | Rubric and validation pass | Low or medium reasoning; prioritize consistency checks. |
  </Tab>
</Tabs>

<span id="which-ai-model-fits-this-task" />
<span id="strategic-model-usage" />

## Working across models [#working-across-models]

Pick each model for the role it plays in a [differentiated team](/docs/craft/composing-agentic-research-systems#dfn-differentiated-team) or single-agent workflow. We work with Kimi as a regular collaborator across reasoning, writing, coding, and sustained project work because its perspective differs from Claude and GPT from the first framing onward. A model's role can shift with the stage of the project.

<Tabs>
  <Tab title="Task choices">
    | Work shape | Good starting models | Why |
    | --- | --- | --- |
    | One bounded theory, critique, or writing task | Claude Opus 5.5 or Claude Sonnet 5.5 | Strong nuance and close adherence to a scoped brief. |
    | Carry out an agreed project plan | GPT-6 Astra or GPT-6.1 Sol | Follows through reliably across multi-step plans once criteria and boundaries are explicit. |
    | Develop an independent framing from the outset | Kimi K3 or Kimi K3.1 Pro | Brings a distinct perspective across reasoning, writing, coding, and sustained project work. |
    | Multi-model coordination | Claude Fable 5.1, GPT-6 Astra, or Kimi K3; compare Fable 5.2 where its retention rules fit | Coordinates specialist passes or alternates between delegation and direct work. |
    | Large literature sets and long-context synthesis | Gemini 3.8 Flash; compare Gemini 4 Argon or GLM-5.3 | Handles large corpora cleanly before you narrow to close reading. |
    | Lower-cost volume screening | DeepSeek Flash, Gemini 3.5 Flash-Lite, GPT-6 Luna, or Qwen3.7-Plus / Qwen3.8-Flash | Keeps iteration affordable on screening, tagging, and extraction; compare error rates alongside cost. |
    | Quick Claude-family checks | Claude Haiku 5.5 | Fast, lower-cost option for bounded checks while Haiku 4.5 remains in legacy menus. |
    | Privacy-sensitive or offline corpora | Local / open-weight models | Keeps restricted material on your own hardware. |

    In our daily runs, Claude Opus 5.5 sometimes pauses mid-task to re-confirm actions even after permissions and boundaries are granted, whereas GPT-6 Astra and GPT-6.1 Sol tend to carry an agreed plan through.
  </Tab>

  <Tab title="Discovery">
    Test how each model handles your corpus before committing to a long run:

    - Start with Google AI Studio, Cherry Studio, OpenCode, or OpenRouter.
    - Run the same prompt and evidence pack across at least three model families.
    - Record failure modes, not just best answers.
    - Promote a model into your regular workflow only after it handles your real material.
  </Tab>

  <Tab title="Writing">
    - For one draft or revision, work with a single strong model and keep the task bounded.
    - For several independent readings or reviews, work with a coordinator to preserve the brief and consolidate disagreements.
    - Invite a model from another family to develop an alternative reading and examine where the accounts diverge.
    - Treat large-context synthesis as preparation for your own writing, not as the finished argument.
  </Tab>
</Tabs>

## Free and low-cost options [#free-and-low-cost-options]

Google AI Studio offers free access to selected Gemini models, though Argon is paid-only. Check [Gemini pricing](https://ai.google.dev/gemini-api/docs/pricing) and [rate limits](https://ai.google.dev/gemini-api/docs/rate-limits) before relying on a quota for batch work.

Other ways to manage cost:

- Use OpenCode or Cherry Studio to compare providers on a small sample before running a full corpus.
- Work with DeepSeek or Qwen on high-volume screening when their accuracy on your sample matches your rubric.
- Work with local models when data governance matters more than frontier capability.
- Reserve higher-cost reasoning models for the synthesis and critique stages where their judgment pays off.

## Next step [#next-step]

Test these candidates against your own materials with the [AI Model Discovery Protocol](/docs/craft/ai-model-discovery-protocol).