Models

code.foxl.ai is a separate managed SaaS. Its cloud orchestrator and coding agents use Foxl Relay credits without a personal model key. Remote Code tasks opened from Desktop use that same SaaS path. See Code model access.

Superseded models are hidden until you ask for them. Opus 4.7, Opus 4.6, Sonnet 4.6 and GPT-5.4 stay fully runnable - same pricing, same tier eligibility, a saved selection keeps working - but they drop out of every picker unless Show older model versions is on in Settings > Model & provider.

Claude Opus 5 (claude-opus-5, GA 2026-07-24) is the new default across Foxl Desktop, Foxl Code's coding agent, and Foxl Notes, and it backs the expert complexity tier in desktop routing. 1M context, 128K max output, vision, knowledge cutoff May 2026, and the strongest Opus-tier model for computer use. It lists at $5 / MTok input and $25 / MTok output - the same price as Opus 4.8 and half of Fable 5. Cache reads are $0.50 / MTok and 5-minute cache writes are $6.25 / MTok. Opus 4.8 remains selectable. Reaching Opus 5 today means your own Anthropic or Bedrock key, or a Claude Pro/Max subscription.

On Bedrock, Opus 5 is us.anthropic.claude-opus-5 (both the us. and global. cross-region inference profiles work, no region pin) and runs with zero data retention by default - no data-retention opt-in is required, unlike Fable 5. Higher-risk prompts may be answered by Opus 4.8 instead and are billed at Opus rates.

Managed Code serves the GPT-5.6 family (Sol, Terra, Luna) plus GPT-5.5 / GPT-5.4 through Amazon Bedrock's native OpenAI Responses API. On Desktop you reach them with your own AWS account (Amazon Bedrock under Settings > Model & provider), an OpenAI API key, or a ChatGPT subscription (OAuth) below. All three GPT-5.6 tiers support reasoning effort up to the new max level.

As of August 3, 2026, GPT-5.6 Sol, Terra and Luna serve the full 1M-token window on Amazon Bedrock, so a whole repository, a long contract or a full multi-step agent history fits in one request. Prompt caching with explicit cache breakpoints applies to those long-context requests, and repeated context is billed at a 90% discount.

GPT-5.5 and GPT-5.4 are still capped at 272K tokens (278,528) on Bedrock - AWS's announcement covers the 5.6 generation only.

The ChatGPT (Codex) OAuth transport serves 272K for the whole GPT family, the 5.6 trio and GPT-6 Astra included - that 1M is Bedrock's, and it does not carry over to a subscription. The figure is the Codex CLI's own: every slug it serves reports context_window: 272000 in ~/.codex/models_cache.json. Foxl uses the limit of the active provider for its context meter and model configuration, so the same model reports a different window depending on how you are connected.

GPT-5.6 introduces billed prompt-cache writes (1.25x the input rate) alongside discounted cache reads (90% off). Foxl meters both accurately - cached GPT traffic is billed at the discounted read rate, not the full input price.

Grok 4.6 (grok-4.6, GA on Amazon Bedrock 2026-08-18) is xAI's frontier model for coding, agentic tasks and knowledge work, with a 500K-token context window and configurable reasoning effort (low, medium, high, xhigh). low is the platform default, which is what Foxl sends when you enable thinking without picking a level. Turning thinking off is honoured too: Foxl sends effort none, which the endpoint accepts and answers with zero reasoning tokens, even though the model card lists only the four levels above. It lists at $2.20 / MTok input and $6.60 / MTok output, with cache reads at $0.55 / MTok.

It is served from us-west-2, the one region Bedrock offers it in on that endpoint, independently of where the GPT models are served.

GLM 5 (Z.AI) and Kimi K2.5 (Moonshot) are available on Desktop with your own AWS account (Amazon Bedrock under Settings > Model & provider) and separately on Code SaaS through Foxl Relay credits. Both vendors also serve their newer models directly - add a Z.ai or Kimi key in the same pane.

Claude Fable 5 ships with built-in safeguards. Prompts in sensitive domains (cybersecurity, biology, chemistry, health) are automatically answered by Opus 4.8 instead and billed at Opus rates.

Claude Code (OAuth) runs in compatibility mode so requests route through your Claude Pro/Max subscription instead of pay-as-you-go "Extra usage". In this mode, Foxl-specific tools (memory, subagents, schedules, channel send, browser extension, view image) are disabled - only Bash, Read, Grep, and WebFetch are available. Haiku 4.5 is not exposed in this mode. For the full Foxl tool surface, use an Anthropic API key (BYOK) or Claude on Amazon Bedrock through your own AWS account.

GPT-6 Astra (gpt-6-astra, launched September 4, 2026; GA on Amazon Bedrock September 8, 2026) is OpenAI's most capable model and the newest one Foxl exposes. 1,050,000-token context, 128,000 max output, text and image input, knowledge cutoff April 30, 2026. It lists at $10 / MTok input and $50 / MTok output, with cached input at $1 and cache writes at $12.50 (1.25x the uncached input rate). The ChatGPT OAuth transport caps the context at 272K; every other route serves the full window.

Prompts over 272K input tokens cost more. OpenAI prices those at 2x the input and cache rates and 1.5x output, for the whole request. Foxl's usage meter records one rate per model, so a request past that line is under-reported rather than mis-charged - and on your own AWS account or your own API key the bill is the vendor's, not Foxl's.

It has no none reasoning level: low, medium, high, xhigh and max are accepted and none is refused, so Astra always reasons. With thinking switched off Foxl sends low rather than a level that would fail the turn. That holds on all three routes - Amazon Bedrock answers the same request Unsupported value: 'none' is not supported.

On Amazon Bedrock, Foxl reaches it two ways. With your own AWS account it calls the cross-region inference profile us.openai.gpt-6-astra, which works from us-east-1, us-east-2 and us-west-2 with no region pin. On Foxl credits it goes through Bedrock's OpenAI Responses endpoint as openai.gpt-6-astra, which serves this model in us-west-2 only - Foxl pins that region for you.

Adaptive thinking (type: "adaptive") is the recommended mode for Opus 5, Fable 5.1, Fable 5, Opus 4.8 / 4.7 / 4.6, Sonnet 5, and Sonnet 4.6 (Fable 5 and Sonnet 5 accept only adaptive). Haiku 4.5 uses type: "enabled" with a budget_tokens parameter instead.

Effort levels are per model, not one list. Opus 5, Fable 5.1, Fable 5, Opus 4.8, Opus 4.7 and Sonnet 5 expose all five (low / medium / high / xhigh / max); Opus 4.6 and Sonnet 4.6 expose four - xhigh arrived after them. Opus 5, Fable 5.1, Fable 5, Opus 4.8 and Opus 4.7 default to high; both Sonnets default to medium.

On Opus 5, adaptive thinking is on by default - that is the change from Opus 4.8, where you had to opt in. Opus 5 exposes all five effort levels (low / medium / high / xhigh / max) and defaults to high. It also supports reasoning summaries, and its max_tokens bounds thinking and response text together.

Opus 4.7 and 4.8 also reject non-default temperature / top_p / top_k with HTTP 400 - Foxl strips those automatically on those models, so the API uses its calibrated default.