Choosing Your Model - Factory Documentation

Balance accuracy, speed, and cost by picking the right model and reasoning level for each task. Every model we offer meets proprietary quality and cost-efficiency requirements. Model quality evolves quickly, and we tune the CLI defaults as the ecosystem shifts. Use this guide as a snapshot of how the major options compare today, and expect to revisit it as we publish updates. This guide was last updated on Wednesday, June 3rd 2026.

1 · Current stack rank (June 2026)

Rank	Model	Why we reach for it
1	Claude Fable 5^*†	First public release Mythos model; top pick for multi-step autonomous engineering work. 4× multiplier.
2	Claude Opus 4.8	Newest Anthropic flagship with Max reasoning; top pick for hardest work. 2× multiplier.
3	Claude Opus 4.8 Fast	Opus 4.8 tuned for faster response times; 4× multiplier.
4	Claude Opus 4.7	Anthropic flagship with Max reasoning; top pick for the hardest work. 2× multiplier.
5	Claude Opus 4.7 Fast	Opus 4.7 tuned for faster response times; 12× multiplier.
6	Claude Opus 4.6	Previous Anthropic flagship with Max reasoning; still excellent depth and safety for complex work.
7	Claude Opus 4.6 Fast	Opus 4.6 tuned for faster response times; 12× multiplier.
8	Claude Opus 4.5	Proven quality-and-safety balance; strong default for TUI and exec.
9	Claude Sonnet 4.6	Max reasoning at the Sonnet price point (1.2×); strong daily driver for planning and implementation.
10	GPT-5.6 Sol	New OpenAI flagship (GPT-5.6) with Max reasoning, 922K context, and verbosity support; top OpenAI pick for the hardest work. 2× multiplier.
11	GPT-5.6 Terra	Balanced GPT-5.6 with Max reasoning and 922K context at a 1× multiplier; strong general-purpose OpenAI option.
12	GPT-5.6 Luna	Cost-efficient GPT-5.6 at 0.4× with reasoning up to Max and 922K context; great for high-volume OpenAI work.
13	GPT-5.4	Latest OpenAI model with 922K context, 128K output, verbosity support, and Extra High reasoning; excellent for large-context tasks.
14	Claude Sonnet 4.5	Strong daily driver with balanced cost/quality; great general-purpose choice when you don’t need Opus-level depth.
15	GPT-5.3-Codex	Newest OpenAI coding model with Extra High reasoning and verbosity support; strong for implementation-heavy tasks.
16	GPT-5.2-Codex	Proven OpenAI coding model with Extra High reasoning; solid for implementation-heavy tasks.
17	GPT-5.2	OpenAI model with verbosity support and reasoning up to Extra High.
18	Claude Haiku 4.5	Fast, cost-efficient for routine tasks and high-volume automation.
19	Gemini 3.1 Pro	Newer Gemini Pro generation with strong structured outputs and mixed reasoning controls for research-heavy tasks.
20	Gemini 3 Flash	Fast, cheap (0.2× multiplier) with full reasoning support; great for high-volume tasks where speed matters.
21	Droid Core (NVIDIA Nemotron 3 Ultra)	Open-source, 0.4× multiplier, fast and cost-efficient model for high-frequency automations; no image support.
22	Droid Core (MiniMax M3)	Open-source, 0.12× multiplier with multimodal image support and strong agentic capabilities; dependable workhorse for everyday tasks with extended context.
23	Droid Core (MiniMax M2.7)	Open-source, 0.12× multiplier with reasoning support (Low/Medium/High) and image support; cheapest model available.
24	Droid Core (GLM-5.1)	Open-source, 0.55× multiplier, newer GLM option for bulk automation and air-gapped environments; no image support.
25	Droid Core (GLM-5)	Open-source, 0.4× multiplier, stable choice for bulk automation and air-gapped environments; no image support.
26	Droid Core (Kimi K2.7 Code)	Coding-specialized Kimi at 0.4× with optional High reasoning; strong agentic and tool use for cost-sensitive dev work.
27	Droid Core (Kimi K2.6)	Open-source, 0.4× multiplier with image support and optional High reasoning; good for cost-sensitive work when you still want a thinking toggle.

^* Anthropic requires all Mythos-class models comply with 30 day data retention for trust and safety, please see more information here. ^† As of June 12, 2026, Claude Fable 5 is not currently available—see Anthropic’s update on Mythos access here.

We ship model updates regularly. When a new release overtakes the list above, we update this page and the CLI defaults.

2 · Match the model to the job

Scenario	Recommended model
Deep planning, architecture reviews, ambiguous product specs	Start with Opus 4.8 for best depth and safety, or fall back to Opus 4.7 / Opus 4.6 for faster turnaround. Use Sonnet 4.6 or Sonnet 4.5 when you want balanced cost/quality, or GPT-5.4 for large-context reasoning.
Full-feature development, large refactors	Opus 4.8 or Opus 4.7 for depth and safety. GPT-5.4, GPT-5.3-Codex, or GPT-5.2-Codex when you need speed plus Extra High reasoning; Sonnet 4.6 or Sonnet 4.5 for balanced loops.
Repeatable edits, summarization, boilerplate generation	Haiku 4.5 or Droid Core (including MiniMax M2.7 at 0.12×) for speed and cost. GPT-5.2 when you need higher quality or structured outputs.
CI/CD or automation loops	Favor Haiku 4.5 or Droid Core for predictable, low-cost throughput. Use GPT-5.3-Codex or GPT-5.4 when automation needs stronger reasoning.
High-volume automation, frequent quick turns	Haiku 4.5 for speedy feedback. Droid Core (especially MiniMax M2.7 at 0.12× with reasoning) when cost is critical or you need air-gapped deployment.

Claude Opus 4.8 is the newest top-tier option for extremely complex architecture decisions or critical work where you need maximum reasoning capability. Claude Opus 4.7 and Claude Opus 4.6 remain excellent alternatives, and Opus 4.6 Fast is tuned for faster responses at a higher cost. Most tasks don’t require Opus-level power—start with Sonnet 4.6 or Sonnet 4.5 and escalate only if needed.

Tip: you can swap models mid-session with /model or by toggling in the settings panel (Shift+Tab → Settings).

3 · Switching models mid-session

Use /model (or Shift+Tab → Settings → Model) to swap without losing your chat history.
If you change providers (e.g. Anthropic to OpenAI), the CLI converts the session transcript between Anthropic and OpenAI formats. The translation is lossy—provider-specific metadata is dropped—but we have not seen accuracy regressions in practice.
For the best context continuity, switch models at natural milestones: after a commit, once a PR lands, or when you abandon a failed approach and reset the plan.
If you flip back and forth rapidly, expect the assistant to spend a turn re-grounding itself; consider summarizing recent progress when you switch.

4 · Reasoning effort settings

Opus 4.8: Off / Low / Medium / High / Max (default: High)
Opus 4.7: Off / Low / Medium / High / Max (default: High)
Opus 4.6 / Opus 4.6 Fast: Off / Low / Medium / High / Max (default: High)
Sonnet 4.6: Off / Low / Medium / High / Max (default: High)
Opus 4.5 / Sonnet 4.5 / Haiku 4.5: Off / Low / Medium / High (default: Off)
GPT-5.6 Sol / Terra / Luna: None / Low / Medium / High / Extra High / Max (default: Medium)
GPT-5.4: None / Low / Medium / High / Extra High (default: Medium)
GPT-5.2: Off / Low / Medium / High / Extra High (default: Low)
GPT-5.2-Codex: None / Low / Medium / High / Extra High (default: Medium)
GPT-5.3-Codex: None / Low / Medium / High / Extra High (default: Medium)
Gemini 3.1 Pro: Low / Medium / High (default: High)
Gemini 3 Flash: Minimal / Low / Medium / High (default: High)
Grok 4.5: Low / Medium / High (default: Medium)
Droid Core (GLM-5): None only (default: None; no image support)
Droid Core (GLM-5.1): None only (default: None; no image support)
Droid Core (Kimi K2.6): Off / High (default: High)
Droid Core (MiniMax M2.7): Low / Medium / High (default: High)

Reasoning effort increases latency and cost—start low for simple work and escalate as needed. Max is available on the GPT-5.6 family (Sol, Terra, Luna), Claude Opus 4.8, Claude Opus 4.7, the Opus 4.6 family (Opus 4.6 and Opus 4.6 Fast), and Sonnet 4.6. Extra High is available on the GPT-5.6 family (Sol, Terra, Luna), GPT-5.4, GPT-5.2, GPT-5.2-Codex, and GPT-5.3-Codex.

Change reasoning effort from /model → Reasoning effort, or via the settings menu.

5 · Bring Your Own Keys (BYOK)

Factory ships with managed Anthropic and OpenAI access. If you prefer to run against your own accounts, BYOK is opt-in—see Bring Your Own Keys for setup steps, supported providers, and billing notes.

Open-source models

Droid Core (GLM-5), Droid Core (GLM-5.1), Droid Core (Kimi K2.6), and Droid Core (MiniMax M2.7) are open-source alternatives available in the CLI. They’re useful for:

Air-gapped environments where external API calls aren’t allowed
Cost-sensitive projects needing unlimited local inference
Privacy requirements where code cannot leave your infrastructure
Experimentation with open-source model capabilities

Note: GLM-5 and GLM-5.1 do not support image attachments. Kimi K2.6 and MiniMax M2.7 do support images. Kimi K2.6 adds an Off/High reasoning toggle, while MiniMax M2.7 (the cheapest model available, with 0.12× multiplier) supports Low/Medium/High reasoning. For image-based workflows, use Claude, GPT, Kimi, or MiniMax M2.7. To use open-source models, you’ll need to configure them via BYOK with a local inference server (like Ollama) or a hosted provider. See BYOK documentation for setup instructions.

6 · Keep notes on what works

Track high-impact workflows (e.g., spec generation vs. quick edits) and which combinations of model + reasoning effort feel best.
Ping the community or your Factory contact when you notice a model regression so we can benchmark and update this guidance quickly.

​1 · Current stack rank (June 2026)

​2 · Match the model to the job

​3 · Switching models mid-session

​4 · Reasoning effort settings

​5 · Bring Your Own Keys (BYOK)

​Open-source models

​6 · Keep notes on what works

1 · Current stack rank (June 2026)

2 · Match the model to the job

3 · Switching models mid-session

4 · Reasoning effort settings

5 · Bring Your Own Keys (BYOK)

Open-source models

6 · Keep notes on what works