Ghostcli 

I ran a technical audit of GhostCLI on October 12, 2026: about 200 API responses across 10 test batteries. Here is what I found, backed by the raw files:

Invented tool results: on a stock-price test, the models never called the tool but still answered with prices for AAPL and MSFT. For a coding agent, this is dangerous.
Advertised 256k context: at 220k tokens, 380,935 characters were silently cut from the middle of my prompt.
Fake token counts: input_tokens follows a simple length ÷ 4 formula, not a real tokenizer.
Prefabricated replies: “hello” returns a fixed answer in 0.4 s with 0 tokens, even for a model whose backend was down.
Ignored parameters: max_tokens, stop_sequences, forced tool use and thinking are ignored, and streaming arrives in a single block.
Hidden instructions: requests sent with no system prompt receive injected instructions, and questions in French get answers in Portuguese.
On the models: the “Fable 5.1” and “Opus 5.5” endpoints use formatting typical of Gemini, not Claude. This is a strong indication, not proof. When I raised it, support did not name the upstream provider and only offered a screenshot.

Whatever the underlying model, the service does not deliver what it sells. I would not use it for professional work or send any confidential code through it.

Unfortunately I have similiar experience as other "1-starers"
The advertised context window is not true, at max you will be able to get around 100-120k ctx window usable, the gateway cuts usually the middle of the request.
Second thing - I do not believe that they actually serve real flagship models, when compared to real Astra/Opus/Fable their models give vastly different responses/much worse in quality.
GLM 5.3 doesn't work at all.
Support email is dead/unresponsive.
Discord support closes tickets without resolution.

After mentioning this on discord I got perm-banned.

Big red flag, stay away.

I’ve been using Ghost CLI for about two months now, and the experience has been really good. I was a little skeptical at first, but the token quality is great, the support is excellent, and the delivery was super fast. I’ve already run several benchmarks with the API, and so far everything has been working perfectly. I have an Enterprise plan, and it’s been solving my company’s needs really well. Nothing but praise and appreciation for the Ghost team.

What takes the real Claude 1 hour to do correctly it takes this one 3 hours and messes it all up. I dont believe this is a real Claude provider as the quality is not even close.

Something is wrong the claude 5.5 struggles to do simple coding which it should excel at, had to have a small local model correct a task it continued to fail at. Logs also stat it was switching to deepseek flash at times.

2026-10-03T22:51:49 sent custom claude-opus-5-5 75081 chars 29 redacted
2026-10-03T22:51:28 sent custom claude-opus-5-5 73067 chars 29 redacted
2026-10-03T22:50:03 sent custom deepseek-flash 73013 chars 29 redacted
2026-10-03T22:49:34 sent custom deepseek-flash 73013 chars 29 redacted
2026-10-03T22:44:49 sent deepseek deepseek-flash 0 chars 0 redacted
2026-10-03T22:41:30 sent deepseek deepseek-flash 0 chars 0 redacted
2026-10-03T22:40:15 sent deepseek deepseek-flash 0 chars 0 redacted
2026-10-03T22:08:51 sent custom claude-opus-5-5 92665 chars 67 redacted
2026-10-03T22:08:34 sent custom claude-opus-5-5 108491 chars 89 redacted
2026-10-03T22:08:03 sent custom claude-opus-5-5 88883 chars 71 redacted
2026-10-03T22:07:43 sent custom claude-opus-5-5 88310 chars 71 redacted
2026-10-03T22:07:10 sent custom claude-opus-5-5 87930 chars 71 redacted
2026-10-03T22:06:44 sent custom claude-opus-5-5 86607 chars 71 redacted
2026-10-03T22:06:14 sent custom claude-opus-5-5 85851 chars 71 redacted

As always if something sound really good its not.

Absolute scam, constant 3-5 token /s speed and models returning worse results than official api, also a very predatory refund policy which invalidates the second you use the service, absolute garbage, stay away.

Ghost CLI leaves a very mixed impression. In practice, it only works reasonably well for about 2–3 hours a day. Even during those relatively “good” periods, you still have to deal with constant tool errors, crashes, and freezes that seriously interfere with normal work.
The rest of the time the service is either completely unavailable or slows down to 2–5 tokens per second — which makes it practically unusable. Separately, the authenticity of the claimed models raises serious doubts: the behavior and quality of the responses often do not match what the developers promise.
The seven-day refund policy is also strange — it only applies if you haven’t actually used their services. In other words, the moment you try the product, you risk losing the right to a refund.
Overall, in its current state Ghost CLI looks unstable and unreliable. It is poorly suited for serious daily work.

Purchased GhostCLI Enterprise ($130/30 days) on 2026-09-13. Advertised as "GPT-6 Astra" + Claude Opus 5 / Sonnet 5 / Fable 5.1 with unlimited use.

Within the first hour I tested the API directly (no wrapper, just curl:

gpt-6-astra returned "（），由智谱AI（）开发。" — literally "developed by Zhipu AI" — twice, on two independent runs: msg_01Ezi5jvXcG1QdVCg5ADh2FO (2026-09-13T16:25Z) and msg_01uFtF0hEIaqmDWlUu5H05fa (17:31Z). This is Zhipu GLM's self-identification, not GPT-6.

Their own web chat labeled "GPT-6 ASTRA" admitted in Chinese it is trained by Zhipu AI, then said "GPT-6 Astra doesn't exist as far as it knows" and couldn't recognise its own advertised name.

claude-fable-5.1 answered "Qwen" when asked its name — a third, unrelated model.

Other Claude endpoints rotated "model provider did not complete the request" errors between runs.

$130/mo for "unlimited frontier models" is economically impossible without swapping — this proves it. Removed same day, filed chargeback + reports to Stripe/Hostinger/OpenAI/Anthropic.

Verify model identity before trusting any proxy: ask it who made it and check raw API payloads. This particular provider is injecting into your prompts to keep the Chinese model from revealing who/what it really is. One trick they use is telling the model a single Chinese character leak will break their live service.

I had a very disappointing experience with GhostCLI.

At first, the service worked surprisingly well. For a short period of time, the coding quality, reasoning, coherence, and general performance were actually quite good, and I thought I had found a genuinely useful alternative for working with Claude models.

Unfortunately, after only a few days of use, the quality dropped dramatically.

The coding performance became extremely inconsistent. I started getting poor responses even for very simple tasks, including basic code modifications where I provided the exact line that needed to be changed. The model would misunderstand straightforward instructions, introduce unnecessary errors, lose context, or fail to perform changes that had previously worked without any issue.

The reliability of the service has also been very poor. I have experienced frequent outages, errors, long maintenance periods lasting hours, and general instability. This makes it very difficult to rely on GhostCLI for any serious development workflow.

What makes the situation worse is the lack of communication from the developers. There is very little transparency when problems happen, very limited information about maintenance or service degradation, and essentially no meaningful communication with users when the platform is having issues.

The most frustrating part is the inconsistency. The service initially showed a lot of promise, which makes the drop in quality and reliability even more disappointing.

Based on my experience, I would not currently recommend GhostCLI. It doesnt provide what it promises.