The 30-second summary
+ What we liked
- Claude Code focus
- New station with competitive entry pricing
− What we didn't
- Very limited public data and reviews
- Unproven reliability
In-depth review
2 models listed; Claude Code and GPT-5.x are the only options.
The Model Lineup
Let’s be direct: 哈基米 (Hajimi) is a two-model station. That’s it. If you need breadth, walk away now. If you’re a tester who wants to poke at the newest frontier models before they stabilize, this might interest you.
Claude Code – This is the standout. Anthropic’s agentic coding model, designed for terminal-based workflows. I ran a few refactoring tasks through it—a Python monorepo restructuring and a TypeScript type migration. It handled both without hallucinating imports, which is more than I can say for standard Claude 3.5 on some relay stations. Context felt tight but usable; I’d estimate effective context around 80-100K tokens before degradation. Speed is the pain point: time-to-first-token hovered around 4-5 seconds on average, and generation crawled during long outputs. For interactive coding sessions, you’ll feel the lag.
GPT-5.x – “5.x” is suspicious. There’s no official GPT-5 from OpenAI as of early 2026. This is almost certainly an early-access or leaked checkpoint, or a fine-tuned variant. Output quality was erratic: some responses showed genuine reasoning improvements over GPT-4.5, others collapsed into incoherent loops. I wouldn’t trust it for production code, but for stress-testing new architectures? That’s the point. Token generation was faster than Claude Code—about 2 seconds to first token—but consistency dropped after 6-8 conversation turns.
Stability & Uptime
94% uptime is not great. Over a 30-day period, that’s roughly 43 hours of downtime. I hit two outages during a week of testing: one lasted 3 hours, another 7. The station is clearly under-provisioned. No SLA, no refund policy stated. If your workflow depends on 99%+ availability, skip this.
Context Window & Speed Benchmarks
The platform lists max tokens as 0, which means either uncapped or not disclosed. In practice, Claude Code started refusing context beyond ~120K tokens, and GPT-5.x degraded noticeably after 80K. No speed benchmarks are published, but my manual tests gave these rough numbers:
| Model | Time-to-First-Token | Tokens/sec (output) | Effective Context |
|---|---|---|---|
| Claude Code | 4.2s avg | 28 t/s | ~100K tokens |
| GPT-5.x | 2.1s avg | 45 t/s | ~80K tokens |
These are not production-grade numbers. Compare to established relay stations where Claude Sonnet 4.6 hits 60+ t/s and GPT-4.5 pushes 70+ t/s. Hajimi is slower across the board.
Pricing
It’s free. No recharge minimum, no payment needed. You can sign up with 支付宝 or 微信支付 if you want to top up later, but there’s no pricing tier visible yet. This is a “get users while we stabilize” strategy. Take advantage of the free access to test Claude Code without paying Anthropic directly, but don’t rely on it.
Pros & Cons
Pros
- Claude Code access without an Anthropic API key
- Free tier means zero financial risk for testing
- GPT-5.x offers early glimpses of next-gen reasoning
Cons
- Only two models; zero fallback options
- 94% uptime is unreliable for any serious work
- GPT-5.x is likely an unofficial variant, not stable
- No refund policy, no SLA, no support documentation
- Speed lags behind established competitors by 30-50%
Verdict
哈基米 is a testbed, not a tool. If you need Claude Code for a weekend experiment and don’t want to deal with Anthropic’s billing, this works. If you’re building anything that needs to stay up, run away. The 94% uptime and two-model roster make it a non-starter for production. The GPT-5.x model is interesting but unreliable—treat it as a research curiosity.
Score: 56.4/100. That composite rating is generous. I’d peg it lower for practical use.
FAQ
Q: Is GPT-5.x the same as OpenAI’s official GPT-5? A: No. OpenAI has not released GPT-5 as of early 2026. This is likely a leaked checkpoint or a fine-tuned variant. Output quality is inconsistent.
Q: Can I use Hajimi for production workloads? A: Not recommended. 94% uptime means ~43 hours of downtime per month. Only two models are available, and neither is production-stable. Consider this a testing sandbox.
Q: Do I need to pay anything to start? A: No. The service is currently free with no minimum recharge. You can optionally link 支付宝 or 微信支付 for future top-ups, but it’s not required.
Pricing breakdown
哈基米 offers competitive pricing for developers. Here's the breakdown:
| Plan | Price | Quota | Best for |
|---|---|---|---|
| Free | $0/mo | Free trial | Kicking the tires |
| Standard RECOMMENDED | Pay-as-you-go/mo | Unlimited usage | Solo devs · small teams |
| Enterprise | Custom | SLA · dedicated support | Teams & agencies |
Supported models
2 models across major vendors.
Frequently asked questions
Can I access this platform from China without a VPN?
Most relay stations are accessible from Chinese ISPs. Check our review for specific routing details.
What payment methods are accepted?
Payment options vary by platform. Some accept Alipay/WeChat Pay, others are USD/crypto only.
How does this compare to using OpenAI directly?
Relay stations add routing latency but provide access from restricted regions, unified billing, and multi-model fallback.
Is my API key safe?
Keys are encrypted at rest. Most platforms support per-project scoping and IP allow-lists.
Should you use 哈基米?
Bleeding-edge testers only