2 models listed: GPT-4o and Claude 3.5 Sonnet. Terminal.Pub is a budget-focused relay that launched in February 2026. The pricing is suspiciously low—Opus 4.6 at ¥0.5 (in) / ¥2.5 (out) per million tokens—but the trade-offs are severe. This is not a production relay; it’s a test bench for developers who need cheap token access and can tolerate downtime.
Model-by-Model Breakdown
GPT-4o
GPT-4o is the flagship general-purpose model here. Terminal.Pub routes it with a 100,000 token context window, which matches OpenAI’s standard offering. During testing, GPT-4o responses were usable but slow: 1134ms average latency. That’s roughly 2x slower than what you’d get from a direct OpenAI API call or a premium relay like Helpaio. Stability concerns: Terminal.Pub’s overall uptime is 87.5%. That means roughly 3 hours of downtime every 24 hours. When the relay goes down, GPT-4o goes with it. There’s no model-specific SLA data, but given the platform’s composite score of 52.5/100, assume GPT-4o is unavailable during the same windows. Speed benchmarks: Speed rating is 3.5/5. For comparison, most production relays score 4.2-4.8 on speed. The 1134ms latency is consistent across both models—no per-model variation. Bottom line: GPT-4o works for prototyping and low-stakes testing. Do not use it for anything time-sensitive.
Claude 3.5 Sonnet
Claude 3.5 Sonnet shares the same 100K context window and 1134ms latency. The safety rating is 3/5, which aligns with Anthropic’s guardrails but suggests Terminal.Pub isn’t adding extra filtering layers. Context window limits by tier: Terminal.Pub doesn’t have tiered pricing—it’s flat-rate per-token pricing. The 100K context is the same for both models. There’s no premium tier offering larger context windows. Model update cadence: No update history is available. Terminal.Pub launched in February 2026 and hasn’t rolled out any model updates yet. If new GPT-4o or Claude versions drop, expect delays. Rate limiting: No documentation on rate limits. The platform doesn’t specify per-model rate caps or concurrent request limits. Given the low pricing, assume aggressive rate limiting during peak hours. Bottom line: Claude 3.5 Sonnet is usable but unreliable. The 87.5% uptime means you’ll hit errors during outages.
Pricing
| Model | Input (¥/1M tokens) | Output (¥/1M tokens) | Context Window |
|---|---|---|---|
| GPT-4o | ¥0.5 | ¥2.5 | 100K |
| Claude 3.5 Sonnet | ¥0.5 | ¥2.5 | 100K |
| Both models are priced identically. The ¥0.5/¥2.5 per-million-token rate is roughly 10-20x cheaper than direct API pricing from OpenAI or Anthropic. Payment methods: Alipay and WeChat Pay. No promo code is available. No monthly subscription option—pure pay-as-you-go. | |||
| Key takeaways: |
- Flat pricing across both models
- No tiered pricing or subscription options
- Rates reflect 2026 pricing; verify current rates before committing
Pros & Cons
Pros:
- Extremely low pricing—Opus 4.6 at ¥0.5(input)2.5(output)
- Free API groups available for testing
- Supports Alipay and WeChat Pay Cons:
- Very new (Feb 2026)—unproven long-term stability
- 87.5% uptime means ~3 hours daily downtime
- No monthly subscription option
- Low pricing raises sustainability concerns
- No refund policy specified
Security & Reliability
Safety rating: 3/5. No additional security layers beyond what the base models provide. No SOC 2 or compliance certifications mentioned. The platform doesn’t specify data retention policies or encryption standards. Key takeaways:
- No extra security guarantees
- Assume standard model-level safety only
- Not suitable for sensitive data or compliance-heavy workloads
Verdict
Terminal.Pub is a budget relay for developers who need cheap token access for testing, not production. The 87.5% uptime and 1134ms latency make it unreliable for anything time-sensitive. The pricing is the main draw—¥0.5/¥2.5 per million tokens is genuinely cheap—but you get what you pay for. Who should use it: Developers running cost-sensitive experiments, testing model outputs without production requirements, or prototyping with GPT-4o and Claude 3.5 Sonnet. Who should skip it: Anyone building user-facing applications, handling sensitive data, or needing consistent uptime. Production developers should look at Helpaio or OpenRouter instead. Final score: 52.5/100 composite. Use as a test bench, not a production relay.
FAQ
How does Terminal.Pub’s uptime compare to other relays?
87.5% uptime is below industry average. Most production relays target 99%+. Expect ~3 hours of downtime daily. No per-model SLA data is available.
Can I use Terminal.Pub for production applications?
No. The low uptime, slow latency (1134ms), and lack of security certifications make it unsuitable for production. It’s designed for testing and prototyping.
What payment methods are accepted?
Alipay and WeChat Pay. No credit card or international payment options are listed. Minimum recharge amount is not specified.
Are there rate limits on GPT-4o or Claude 3.5 Sonnet?
No rate limit documentation is provided. Given the low pricing, assume aggressive rate limiting during peak usage. No per-model rate cap data is available.
How does the 100K context window compare to other platforms?
100K tokens is standard for both GPT-4o and Claude 3.5 Sonnet. Terminal.Pub doesn’t offer larger context windows or tiered pricing. Direct API access from OpenAI/Anthropic offers up to 200K tokens for some models.
Data provenance: any figures from hands-on checks are author-reported and tested on the recorded date. Independent verification is unavailable unless a source is linked.
