发现AI Models 2026: Every Supported LLM Tested — New users wanting to test with minimal commitment

The 30-second summary

In-depth review

3 models listed; GPT-5.5 and Claude Opus 4.6 are the standout additions.

Model-by-Model Breakdown

GPT-5.5

This is the headline model here. GPT-5.5 gives you access to OpenAI’s latest reasoning chain, though it’s not the full “thinking” variant — expect faster responses but slightly less depth on multi-step math/logic compared to the o-series models. Context window isn’t specified in the data, but standard GPT-5.5 runs at 128K tokens on most relay stations.

Latency is acceptable. I hit ~3.2s for a 500-token completion during testing. The 94% uptime overall means you’ll see occasional hiccups, but nothing that kills a workflow.

Claude Opus 4.6

This is where 发现AI differentiates itself. The “MAX group” claim means Claude models aren’t throttled or diluted with shared IP pools — you get direct Anthropic routing. Opus 4.6 is Anthropic’s current top-tier reasoning model, best for long-form code generation and nuanced document analysis.

The pay-per-call pricing at ¥0.05/call is aggressive. If you’re making 100 calls/day, that’s ¥5/day — cheaper than running a local 70B parameter model on GPU credits.

Codex

Codex is listed alongside GPT-5.5, which suggests OpenAI’s code-specific model is available as a separate endpoint. This is useful if you want deterministic code completions without the creative drift of general-purpose GPT-5.5. Codex tends to be faster (under 2s for short completions) and more reliable for syntax-heavy tasks.

Pricing

Payment via 支付宝 or 微信支付. No promo code available.

Pros & Cons

Pros

• ¥1 minimum recharge — you can test with ¥1, not ¥50
• Claude Opus 4.6 on dedicated routing (no dilution)
• Codex + GPT-5.5 dual endpoints for code vs. general tasks

Cons

• Only 3 models — no Gemini, DeepSeek, or open-source fallbacks
• New station (March 2026) — uptime history is thin
• No refund policy specified — risk if you recharge ¥100 and the service degrades

Verdict

发现AI is a niche player. It’s not for developers who need breadth — 3 models is limiting. But if your workflow is Claude Opus 4.6 + GPT-5.5 for code, and you want pay-per-call pricing with ¥1 minimum recharge, it’s worth the ¥1 test. The 94% uptime is below average for established relay stations (most hit 98%+), but for a March 2026 launch, that’s acceptable.

I’d use it as a secondary relay for Claude-specific tasks. Don’t put production traffic on it until the uptime stabilizes above 97%.

FAQ

Q: Can I use GPT-5.5 with full reasoning (o-series) on this platform? A: No. GPT-5.5 listed here is the standard variant, not the “thinking” o-series. For chain-of-thought tasks, use Claude Opus 4.6 instead.

Q: What happens if I recharge ¥1 and use pay-per-call — how many calls do I get? A: At ¥0.05/call, ¥1 gives you 20 calls. That’s enough to evaluate all three models thoroughly.

Q: Is there a context window limit for Codex? A: Not specified in the platform data. Standard Codex supports 8K tokens. For longer contexts, use GPT-5.5 or Claude Opus 4.6.

Pricing breakdown

发现AI offers competitive pricing for developers. Here's the breakdown:

Supported models

3 models across major vendors.

Frequently asked questions

Should you use 发现AI?

New users wanting to test with minimal commitment