feat(deepinfra): add Claude Fable 5

[jared-q15]

Adds the Deep Infra entry for anthropic/claude-fable-5, which Deep Infra serves as partner passthrough.

Host kind: multi-model relay, so the file is override-only and inherits models/anthropic/claude-fable-5.toml.

Data and sources

Field	Value	Source
cost.input / cost.output	10 / 50 USD per MTok	Model page, and the public catalog endpoint (cents_per_input_token: 0.001, cents_per_output_token: 0.005)
limit.context / limit.output	1,000,000 / 128,000, inherited	API max_tokens is 1000000, equal to the lab metadata, so no [limit] block
structured_output	true	Live structured-output catalog tag
reasoning_options	effort = low, medium, high, xhigh, max	Lab entry plus same-surface relay peers

rate_per_input_token_cached is null for this model, so no cache_read / cache_write is authored.

Reasoning

reasoning_effort (and reasoning.effort) controls it, with the values copied from the baseline for this model: the lab entry (providers/anthropic/) and the same-surface relay peers (OpenRouter and the other relays of claude-fable-5) all use low, medium, high, xhigh, max.

No toggle and no none in the effort list: Deep Infra's live catalog tags for this ID carry reasoning but not can-disable-reasoning, the tag its own sync module reads as "an on/off toggle exists" (27 of the 28 Deep Infra entries here that author a toggle carry it; the only exception is moonshotai/Kimi-K2.5). This model's API page carries no reasoning section either.

Verification

bun validate passes locally and the model resolves with the inherited metadata (canonical_model_id: anthropic/claude-fable-5).

Note on the Deep Infra sync

packages/core/src/sync/providers/deepinfra.ts filters ^anthropic/ out of this provider's catalog (EXCLUDED_PATTERNS), so this entry is hand-authored and deepinfra:sync will not manage it. With deleteMissing: false it is retained, and it will appear in that provider's missingNotice list on each sync.

If that exclusion is meant to cover curated entries too, say so and I will close this. If the family should instead be sync-managed, the filter is the line to change, and this entry can stand as the first one.

[github-actions]

Action items

• [medium] [possible mistake] providers/deepinfra/models/anthropic/claude-fable-5.toml:36 - Check: Relay effort baseline must be lab + same-surface peers; extra levels need host-specific proof. Why: Lab providers/anthropic/models/claude-fable-5.toml and same-surface peer providers/openrouter/models/anthropic/claude-fable-5.toml plus ~25 other relays use ["low", "medium", "high", "xhigh", "max"] with no none; DeepInfra generic header only documents low|medium|high|xhigh and GLM-5.2 precedent shows none is model-specific (toggle-only there), so generic docs plus outlier vercel peer do not prove reasoning_effort="none" works for anthropic/claude-fable-5 on DeepInfra. Action: Provide model-specific evidence (DeepInfra model schema URL / live test showing none disables reasoning for this ID) or drop none to match lab/OpenRouter baseline.

[jared-q15]

Addressed in e2c115317: none is gone and the effort set now mirrors the baseline this model shares with the lab entry and the relay peers (low, medium, high, xhigh, max).

On the toggle half, so the evidence is on the record: the live catalog tags for this ID carry reasoning but not can-disable-reasoning, the tag packages/core/src/sync/providers/deepinfra.ts reads as "an on/off toggle exists". 27 of the 28 Deep Infra entries in this tree that author a toggle carry that tag (the only exception is moonshotai/Kimi-K2.5), and this model's own API page has no reasoning section, so no off switch is claimed here.

For completeness, the generic reasoning_effort enum on the per-model schema endpoint cannot be used as model-specific evidence either: it returns the same seven values for a non-reasoning model.

[github-actions]

No actionable findings.