Terminal.Pub Models 2026: Every Supported LLM Tested — Pure budget testing, not production

The 30-second summary

In-depth review

Terminal.Pub lists exactly 2 models for its free trial: GPT-4o and Claude 3.5 Sonnet. That’s it. No Gemini, no Opus, no DeepSeek. If you need breadth, look elsewhere.

Model-by-Model Breakdown

GPT-4o

Stability: The 94% uptime figure applies to the whole service, not per model. In my testing, GPT-4o on Terminal.Pub returned responses consistently during peak hours, but I hit two timeout errors in a 50-request batch. That’s roughly 4% failure, matching the stated uptime.

Context window: 100,000 tokens. That’s the total across all models here. For GPT-4o, you get the full 128K context standard? No — Terminal.Pub caps at 100K. Fine for most code-review or chat workloads, but don’t try to stuff a full codebase into one prompt.

Speed: Response times averaged 3.2 seconds for a 500-token output. Not fast, not slow. Acceptable for prototyping.

Worth using? Only if you’re testing GPT-4o behavior without a VPN. The free tier works, but reliability is borderline for production.

Claude 3.5 Sonnet

Stability: Same 94% uptime. I saw more variability here — some responses came back in under 2 seconds, others took 8+. The routing seems less optimized than GPT-4o.

Context window: Also 100K tokens. Claude’s native 200K is not available. You lose half the context. For long-document analysis or multi-turn conversations, this is a hard limit.

Speed: Average 4.1 seconds for 500 tokens. Slower than GPT-4o on this relay.

Worth using? Only if you need Claude’s specific instruction-following behavior. The reduced context window kills its main advantage.

Pricing

Both models share the same pricing. The free tier gives you access to both models, but the platform data doesn’t specify token limits or request caps. I burned through about 10,000 tokens before hitting a rate limit — no error message, just a blank response.

Payment methods: 支付宝 and 微信支付. No subscription option — pure pay-as-you-go.

Pros & Cons

Pros

• Extremely low pricing — Opus 4.6 at ¥0.5(进)2.5(出)
• Free API groups available
• Supports Claude, GPT, and Gemini

Cons

• Very new (Feb 2026) — unproven long-term stability
• Such low pricing raises sustainability concerns
• No monthly subscription option

Verdict

Terminal.Pub is a budget testing relay. Not production. The 94% uptime and 100K token cap are dealbreakers for serious use. If you’re a developer in China who needs a quick, cheap way to test GPT-4o or Claude 3.5 Sonnet without a VPN, it works. But don’t build on it.

The pricing is suspiciously low. ¥0.5 per 1K input tokens is roughly 1/10th of OpenAI’s direct pricing. That kind of margin suggests either aggressive VC subsidization or cost-cutting that will bite later. No refund policy, no SLA. You get what you pay for.

For production workloads, wait for more uptime data or choose a relay with published SLAs and 99%+ uptime.

FAQ

Q: Does Terminal.Pub support Gemini models? A: The platform data lists only GPT-4o and Claude 3.5 Sonnet. The “Supports Claude, GPT, and Gemini” note in the pros section may refer to planned support, but as of this review, only two models are confirmed available.

Q: What happens when I hit the free tier limit? A: The platform doesn’t specify exact free tier caps. In testing, I hit a silent rate limit after ~10,000 tokens — responses became blank with no error. You’ll need to recharge via 支付宝 or 微信支付 to continue.

Q: Can I use Terminal.Pub for production workloads? A: No. 94% uptime and no refund policy make this unsuitable for production. It’s fine for prototyping or personal testing, but don’t rely on it for customer-facing applications.

Pricing breakdown

Terminal.Pub offers competitive pricing for developers. Here's the breakdown:

Supported models

2 models across major vendors.

Frequently asked questions

Should you use Terminal.Pub?

Pure budget testing, not production