Sync NEAR AI Cloud catalog with the live model list

[nearworkerbot]

The NEAR AI Cloud catalog has drifted from what the API serves.

Adds 17 models we now serve:

• Qwen/Qwen3.8-27B
• anthropic/claude-fable-5
• anthropic/claude-fable-5-1
• anthropic/claude-opus-4-8
• anthropic/claude-opus-5
• anthropic/claude-sonnet-5
• deepseek-ai/DeepSeek-V4-Flash
• deepseek/deepseek-v3.2
• moonshotai/kimi-k2.6
• moonshotai/kimi-k3
• openai/gpt-5.6-luna
• openai/gpt-5.6-sol
• openai/gpt-6-astra
• qwen/qwen3.5-397b-a17b
• qwen/qwen3.7-max
• z-ai/glm-5.2
• z-ai/glm-5.3-flash

Left out because they reason and we have not yet recorded how the reasoning is controlled, so reasoning_options would be a guess:

• google/gemini-3.8-flash
• qwen/qwen3-32b
• x-ai/grok-4.6

Generated from https://cloud-api.near.ai/v1/models.

[github-actions]

Action items

• [medium] [possible mistake] providers/nearai/models/deepseek-ai/DeepSeek-V4-Flash.toml:6 - Check: Reasoning options must match this host’s real controls, using the lab + same-surface peer baseline for the model. Why: Lab/openai-compat peers for DeepSeek V4 Flash (providers/deepseek/models/deepseek-flash.toml, providers/ofox/models/deepseek/deepseek-v4-flash-0731.toml, providers/alibaba/models/deepseek-v4-flash-0731.toml) use toggle + effort (high/max, sometimes low). This file only declares toggle and only documents that chat_template_kwargs.thinking works—unlike kimi-k3.toml, which explicitly states reasoning_effort is ignored. Catalog consumers will under-report graded effort if it passes through. Action: Live-test reasoning_effort (and any DeepSeek-style effort field this route accepts). If graded effort works, add it (e.g. high/max or the verified set). If it does not, keep toggle-only and add a leading comment that effort was tested and ignored/rejected.
• [low] [possible mistake] providers/nearai/models/anthropic/claude-fable-5.toml:14 - Check: Provider modality overrides must be real host deltas, not accidental truncation of inherited lab modalities. Why: These new Anthropic/OpenAI entries set input = ["text", "image"], dropping lab pdf (and, for Qwen3.8 / GLM-5.3-Flash, video/audio/pdf). Existing NEAR AI Claude/GPT files leave modalities unset and inherit full lab inputs, and only Kimi/Qwen3.5 entries document live modality tests. Unverified drops mis-advertise attachment/PDF support. Action: Confirm each override against this host (or cite /v1/models + a probe). Keep the narrower list only where verified; otherwise remove the [modalities] block so lab inputs inherit. Apply the same check to the other new Claude, GPT, Qwen3.8-27B, and glm-5.3-flash files that override input.
• [low] [possible mistake] .pr-review/pull-request.json (body) - Check: Material cost changes should cite direct pricing/docs evidence. Why: The PR cites https://cloud-api.near.ai/v1/models for the model list, but the 17 new [cost] blocks have no pricing source mapping (unlike several leading TOML comments for reasoning tests). Reviewers cannot confirm USD/MTok rates from the PR alone. Action: Add a PR-body (or top-of-file) pricing citation per model or shared NEAR price page, stating what each source supports.