Models

code.foxl.ai is a separate managed SaaS. Its orchestrator and managed coding agents use Foxl Relay credits. Remote Kiro and Cursor use a vendor API key saved in Code settings and charge Foxl credits for compute only. Remote Code tasks opened from the desktop app use these same access paths. See Code model access.

Superseded models are hidden until you ask for them. Opus 4.7, Opus 4.6, Sonnet 4.6, Haiku 4.5, GLM 5, GPT-6 Sol, GPT-5.6 Sol / Terra / Luna, GPT-5.4 and Grok 4.6 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. The one exception is GPT-5.6 in Foxl Code's managed catalog (Foxl credits): GPT-6 Sol, GPT-6 Luna and GPT-6.1 Sol are not on credits, so GPT-5.6 Sol / Terra / Luna stay in the main list there.

GPT-6.1 Sol (gpt-6.1-sol, released September 29, 2026) is the upgrade to GPT-6 Sol and replaces it in the picker. It comes close to GPT-6 Astra on agentic coding, computer use and professional work, at GPT-6 Sol's price: $2 / $10 per MTok, with cached input down to $0.10 (from $0.20). 1.05M context, 128K output.

• Reasoning is always on. Like GPT-6 Astra, it accepts low through max (default medium) and refuses none, so Foxl offers no off switch for it. GPT-6 Sol, which can turn reasoning off, stays available with Show older model versions on.
• Reach it with your own AWS account (Settings > Model & provider > Amazon Bedrock) or an OpenAI API key. It is not in Foxl Code's managed catalog.
• It is also on a ChatGPT Plus/Pro sign-in, confirmed working on that connection - see Subscription (OAuth) Models below. There its window is 500K rather than 1.05M, and GPT-6 Sol moves behind Show older model versions as it does on Amazon Bedrock.

GPT-6.1 Terra (gpt-6.1-terra, released October 26, 2026) is the GPT-6 generation's fast, efficient middle tier and the successor to GPT-5.6 Terra: visual reasoning, precise computer use and clear writing at $1 / $5 per MTok, with cached input at $0.05. 1.05M context, 128K output.

• Reasoning can be turned off. It accepts none through max (default medium), like GPT-6 Sol and Luna and unlike GPT-6.1 Sol.
• Reach it with your own AWS account (Settings > Model & provider > Amazon Bedrock). It is not in Foxl Code's managed catalog.
• It is also on a ChatGPT Plus/Pro sign-in, confirmed working on that connection - see Subscription (OAuth) Models below. There its window is 500K rather than 1.05M, and reasoning can be turned off there too.

GPT-6 Sol and GPT-6 Luna (gpt-6-sol, gpt-6-luna, GA September 22, 2026) are the cheaper tiers of the Astra generation, at half their GPT-5.6 counterparts' prices: Sol $2 / $10 per MTok and Luna $0.10 / $0.50, with cached input at a tenth of input. 1.05M context, 128K output. Reasoning is adjustable and CAN be switched off - unlike Astra, both accept a none effort level - and the default is medium.

Reach them with your own AWS account (Settings > Model & provider > Amazon Bedrock) or an OpenAI API key. Two limits worth knowing, both measured rather than assumed:

• They are not in Foxl Code's managed catalog. Foxl's managed OpenAI path does not offer these two models yet, so they need your own account.
• They are also on a ChatGPT Plus/Pro sign-in, confirmed working on that connection - see Subscription (OAuth) Models below. There the window is 272K rather than 1.05M, which is the transport's limit for every GPT slug and not something specific to these two.

Claude Opus 5.5 (claude-opus-5-5, GA 2026-09-22) is the new default across Foxl Agent and Foxl Code's coding agent, and it backs the expert complexity tier in desktop routing. 1M context, 128K max output, vision, knowledge cutoff June 2026. It lists at $4 / MTok input and $20 / MTok output - LESS than the Opus 5 it replaces ($5 / $25). Cache reads are $0.20 / MTok (5% of input, like Sonnet 5.5 since October 7, 2026; most other Claude models charge 10%) and 5-minute cache writes are $5 / MTok.

Thinking cannot be turned off on this model. Adaptive thinking is always on, so the model selector shows the effort levels with no on/off switch, and effort is the only depth control. Its default effort is medium, not the high every other Opus tier uses - and at the same level it thinks more per turn than Opus 5, so re-check a setting you carry over rather than assuming the cost is unchanged.

Forced tool use is also gone: a request that pins tool_choice to a specific tool, or to "any tool", is refused. Nothing in Foxl sends one.

Foxl Notes keeps Sonnet 5 as its default. Opus 5 remains selectable. In Foxl Agent, reach Opus 5.5 with your own Anthropic or Bedrock credentials, or a Claude Pro/Max sign-in through Claude Code. Managed Code uses Foxl credits, charged at what Amazon Bedrock bills for the US cross-Region profile it runs on: $4.40 / $22 per MTok, 10% above list, like every Claude model on Foxl credits except Sonnet 5.5 and Sonnet 5 (Model rates used by Relay).

On Bedrock, Opus 5.5 is us.anthropic.claude-opus-5-5. Both the us. and global. cross-region inference profiles were verified live from us-west-2 and us-east-1, so there is no region pin, and no data-retention opt-in is required. Computer use is unaffected here: Bedrock still accepts the earlier computer_20251124 tool for this model, unlike the Claude API and Google Cloud.

Claude Opus 5 (claude-opus-5, GA 2026-07-24) is the previous default for Foxl Agent and Foxl Code's coding agent. It remains selectable alongside Opus 5.5. It has 1M context, 128K max output, vision, and a May 2026 knowledge cutoff. It lists at $5 / MTok input and $25 / MTok output. Cache reads are $0.50 / MTok and 5-minute cache writes are $6.25 / MTok. Reach it with your own Anthropic or Bedrock credentials, or a Claude Pro/Max subscription.

Claude Sonnet 5.5 (claude-sonnet-5-5, GA 2026-09-28) is the latest Sonnet. It is selectable with your own Anthropic key or Bedrock credentials, with a Claude Pro/Max sign-in, in Foxl Notes, and in Foxl Code on Foxl credits. Sonnet 5 stays the Foxl Notes default. 1M context, 128K max output, vision, knowledge cutoff June 2026. It lists at $2 / MTok input and $10 / MTok output, with 5-minute cache writes at $2.50 / MTok. Cache reads are $0.10 / MTok since October 7, 2026, half the earlier $0.20 - on the Claude API, on Amazon Bedrock (AWS lowered its price the same day) and so on Foxl credits, which bill what Bedrock charges.

Thinking cannot be turned off on this model, the same as Opus 5.5: the model selector shows its five effort levels with no on/off switch. Foxl defaults it to medium effort. Effort levels are recalibrated on this model, so a level does not produce the same amount of thinking as it did on Sonnet 5. Forced tool use is refused, as on Opus 5.5; nothing in Foxl sends it.

On Bedrock, Sonnet 5.5 is global.anthropic.claude-sonnet-5-5. That global cross-region profile is its only one: a us. profile does not exist for this model. It was verified live from us-west-2 and us-east-1. A global profile can route a request to any AWS commercial Region (AWS: cross-Region inference), so unlike the other Claude models on Foxl credits, a Sonnet 5.5 request is not kept inside the US. Claude Code (Pro/Max sign-in) mode lists it too, capped at 200K tokens there like Sonnet 5.

Claude Haiku 5.5 (claude-haiku-5-5, GA 2026-10-07) is the latest Haiku and replaces Haiku 4.5, which moves to the older models. It is Foxl's small fast model: the greeting suggestions, session summaries on Amazon Bedrock and Foxl Code's chat orchestrator run on it, it names chats on Amazon Bedrock with your own AWS account and on a Claude Pro/Max sign-in (and on OpenRouter when the free naming model cannot; see How chats are named), and so do Foxl Notes' title and live translation models when set to Automatic with an Anthropic key (on Amazon Bedrock, Automatic runs Gemma 4 31B first and Haiku 5.5 only if the account refuses Gemma 4 and GPT-6 Luna; a Claude Pro/Max sign-in lists no Haiku, so there Automatic runs Foxl Agent's own model). 1M context (Haiku 4.5 had 200K), 128K max output, vision, knowledge cutoff June 2026.

It is priced by prompt length. Up to 100,000 prompt tokens it lists at $0.10 / MTok input and $0.50 / MTok output, with cache reads at $0.01 and 5-minute cache writes at $0.125, a tenth of Haiku 4.5's prices. Above 100,000 tokens the whole request is billed at five times those rates. It also uses the newer tokenizer, so the same text counts as about 30% more tokens than on Haiku 4.5.

Thinking is adaptive and can be turned off. Unlike Haiku 4.5, it thinks when a request does not mention thinking, so Foxl sends thinking off explicitly whenever thinking is switched off and for every helper call - a title or a translation segment runs without reasoning, as it did on Haiku 4.5. With thinking on, it takes effort levels low through max; its default is medium. temperature, top_p and top_k are refused, and so is an assistant prefill; nothing in Foxl sends either.

On Bedrock it is us.anthropic.claude-haiku-5-5 on Foxl credits (the US cross-Region profile, $0.11 / $0.55 per MTok, 10% above list like the other Claude models) and global.anthropic.claude-haiku-5-5 on your own AWS account. Both profiles were verified live from us-west-2.

What moved to it. A Notes helper model left on its default moves to Haiku 5.5 by itself, and so does a Notes helper set to Haiku 4.5 - the two cannot be told apart, because Haiku 4.5 was the stored default. Every other saved selection (a chat model, a schedule, an agent definition) keeps running the model it names, Haiku 4.5 included. Context compaction is unchanged: it summarizes with the model the conversation runs on, which reads that model's prompt cache.

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. In Foxl Agent, use your own AWS account or an OpenAI API key. The supported ChatGPT subscription models are listed below; that path excludes GPT-5.4 and GPT-5.4 Mini. All three GPT-5.6 tiers support reasoning effort up to max.

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 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 exceptions are GPT-6.1 Sol and GPT-6.1 Terra at 500K. The figures are the Codex endpoint's own: every other slug it serves reports context_window: 272000 there, and the two GPT-6.1 models report 500000. 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 4.7 (grok-4.7, GA on Amazon Bedrock 2026-09-28) is xAI's frontier model for coding, agentic tasks and knowledge work, with a 500K-token context window, image and document input, and reasoning effort low, medium, high or xhigh. Its reasoning cannot be turned off, so the model selector shows the effort levels with no on/off switch; Foxl starts it at high, the platform default. It lists at $2.20 / MTok input and $6.60 / MTok output, with cache reads at $0.55 / MTok, on Foxl credits and on your own AWS account.

Grok 4.6 (grok-4.6) is the previous Grok and is now an older model version: hidden by default, still selectable with Show older model versions on. Its reasoning can no longer be turned off either (Bedrock now refuses effort none for it), so a thinking-off request runs at low. Same price as Grok 4.7.

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 5.3 (Z.AI) and Kimi K2.5 (Moonshot) are available in Foxl Agent 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.

Kimi K3 is your-own-AWS-account only - it is not on Foxl Relay credits, so it appears under Amazon Bedrock and nowhere else. Foxl calls it through Bedrock's OpenAI-compatible Responses API rather than Converse, which is what AWS recommends for this model and what lets it read an attached PDF and reuse a cached prompt; on Converse the same model refuses both. GLM 5 is superseded by GLM 5.3 and sits in the older-models group, hidden until you turn on the show-legacy-models switch in Settings; it still runs exactly as before, including for saved selections and scheduled tasks. GLM 5.3 runs only through Bedrock's US cross-Region inference profile, which Foxl uses automatically. AWS serves that profile from us-east-1 and us-west-2 (Foxl's default) today, so with your account set to another Region this model answers with an error. Amazon also gives GLM 5.3 only to AWS accounts it has approved.

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 the agent gets only Bash, Read, Grep and WebFetch; every other tool, including writing files, memory, subagents, schedules, the browser and your MCP and custom tools, is left out, and Settings > Model & provider lists them by name. Neither Haiku 4.5 nor Haiku 5.5 is 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. 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 (Foxl Code) 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.

Ultrafast, with your own AWS account. Amazon Bedrock also serves GPT-6 Astra on its Ultrafast tier (September 30, 2026): the same model, faster, at six times the Standard price - $66 / MTok input, $330 / MTok output and $6.60 cached input on the US profile. Turn it on with the Ultrafast switch in the model selector, which takes the place of Reasoning's fixed "Always on"; a lightning mark on the model button shows it is on. It is off by default, applies only to the chat messages you send (titles, suggestions and scheduled runs stay on Standard), and Foxl's usage figures count those messages at six times the Standard rate. OpenAI reports up to six times the speed; in Foxl's own measurement most Ultrafast replies streamed at 130-157 tokens per second against 25-63 on Standard, though one ran no faster. It is not on Foxl credits: the Bedrock endpoint Foxl credits use for this model does not offer the tier. The phone and iPad apps and app.foxl.ai show the same switch while you chat with GPT-6 Astra on your desktop's AWS account: there it applies to the messages sent from that device, and until you change it there it shows, and uses, the desktop's own setting. With a desktop older than Foxl 0.9.0 the switch there is greyed out, set the way the desktop has it, with a line saying to update Foxl on the computer to change it from that device, and messages from it run at the desktop's setting.

Adaptive thinking (type: "adaptive") is the recommended mode for Opus 5.5, Opus 5, Fable 5.1, Fable 5, Opus 4.8 / 4.7 / 4.6, Sonnet 5.5, Sonnet 5, Sonnet 4.6 and Haiku 5.5 (Opus 5.5, Fable 5.1, Fable 5, Sonnet 5.5, Sonnet 5 and Haiku 5.5 accept only adaptive when thinking is on). Haiku 4.5 uses type: "enabled" with a budget_tokens parameter instead.

Effort levels are per model, not one list. Opus 5.5, Opus 5, Fable 5.1, Fable 5, Opus 4.8, Opus 4.7, Sonnet 5.5, Sonnet 5 and Haiku 5.5 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, Opus 4.7 and Opus 4.6 default to high; Opus 5.5, Sonnet 5.5, Sonnet 5, Sonnet 4.6 and Haiku 5.5 default to medium.

On Opus 5.5 and Sonnet 5.5 thinking cannot be switched off at all - type: "disabled" and the older budget_tokens shape are both refused - so the model selector shows their five effort levels with no on/off switch.

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.