Claude Code Isn't Worth $200/Month Anymore
Real tokens-per-second benchmarks for MiniMax, Kimi, and GLM via OpenCode Go and Z.AI. Anthropic blocked OpenClaw. Here's what you can use instead.
Claude Code Isn't Worth $200/Month Anymore
Anthropic just made the economics of local AI agents suddenly very interesting.
Friday evening. Classic move. Anthropic sent out emails telling OpenClaw users that their Claude subscriptions would stop covering third-party harness usage starting Monday at 9am PT. One week later than originally planned, because the internet pushed back hard enough that they had to pretend to listen.
Peter Steinberger — the creator of OpenClaw, who literally joined OpenAI last month — tried to negotiate. He and Dave Morin got one week. That's the victory. That's the power of a viral open-source project with a community that makes noise.
So now everyone's asking: what do I do?
I ran the numbers. Real numbers, not marketing claims. From my home server, three runs each, consistent prompt, measuring actual tokens per second. Here's what I found.
The Data
Summary
| Model | Provider | Streaming TPS | Non-stream TPS | TTFT | Notes |
|---|---|---|---|---|---|
| MiniMax-M2.5 | OpenCode Go | 92.8 | 60.2 | 3.0s | Practical winner |
| Kimi-K2.5 | OpenCode Go | 165.2 | 29.2 | 27.1s | Burst king, slow cold start |
| MiniMax-M2.7-highspeed | MiniMax API | 61.6–68.4 | 44.2–48.0 | 6.3–7.7s | Best M2.7 variant |
| GLM-5-Turbo | Z.AI Coding | 55.1 | 49.0 | ~0s | Most consistent |
| MiniMax-M2.7 | MiniMax API | 38.8–51.3 | 32.1–40.5 | 7.6–9.1s | Standard tier |
| GLM-5 | Z.AI Coding | 42.2 | 40.4 | ~0s | Stable, instant TTFT |
| MiniMax-M2.7 | OpenCode Go | broken | 51.5 | 6.3s | Streaming bug on this endpoint |
| GLM-5 | OpenCode Go | 19.5–52.5 | 18.3–59.5 | ~0s | High variance |
Best Performer: MiniMax-M2.5 (OpenCode Go)
93 TPS streaming, 60 TPS non-stream, 3-second TTFT. This is the one. Fast enough to feel interactive, consistent enough to automate. 93 tokens per second for a fraction of what you're paying Anthropic.
| Run | Streaming TPS | TTFT | Gen Time |
|---|---|---|---|
| 1 | 100.7 | 3.16s | 10.17s |
| 2 | 84.7 | 2.87s | 12.09s |
| 3 | 92.9 | 3.02s | 11.02s |
| AVG | 92.8 | 3.0s |
Fastest Burst: Kimi-K2.5 (OpenCode Go)
165 TPS streaming. But it takes 27 seconds before the first token arrives. This is a throughput play for queued work, not interactive use. The math is wild though — burst of 165 TPS is legitimately impressive.
Most Consistent: GLM-5-Turbo (Z.AI Coding)
55 TPS with ~0 second TTFT. It starts outputting immediately. Variance is low across all runs. If you're on Z.AI's Coding plan, this is your endpoint — and it uses a separate balance from their standard API.
Watch Out: OpenCode Go proxy has issues
MiniMax-M2.7 streaming is broken — it sends empty chunks. GLM-5 has high variance via OpenCode Go (16-70 TPS on the same model, same prompt). The proxy layer adds real overhead and unpredictability. Go direct where you can.
The Numbers Don't Lie
MiniMax claims 60 TPS for M2.7. I measured ~45. They claim 100 TPS for highspeed. I measured ~65. Their marketing oversells by 30-40%. The actual numbers are still good. Just don't trust the brochure.
Meanwhile, GLM-5-Turbo via Z.AI was the most honest performer — what they said they'd do, they did, every run, with zero surprises.
The Take
I'm not going to pretend OpenClaw plus a pile of API providers is a perfect replacement for Claude Code. Claude Code is polished. It works. The $200/month Max plan was reasonable for what you got.
But here's the thing: Anthropic doesn't want you using OpenClaw. They said so explicitly. They named it. They enforced it on a Friday evening like they had something to hide. And their "compromise" is pay-as-you-go pricing that wipes out the subscription value proposition entirely.
The alternative is real. MiniMax-M2.5 via OpenCode Go gives you 93 TPS for what these providers actually charge — fractions of a cent per token. If you're running OpenClaw as a daily driver, this is what it looks like to vote with your wallet.
Is it more ops? Yes. You manage your own API keys, monitor your own spend, deal with your own routing. That's the trade. For some people, $200/month for Claude Max is still the right answer. For builders who want control and reasonable cost, the path exists now.
Anthropic made this choice. Fine. The market responded. That's how it works.
Questions, benchmark disputes, or your own data? Hit me up on Slack.