Mingles Router vs Claude API
On the like-for-like table DeepSeek publishes, DeepSeek-V4-Flash lands within a couple of points of Claude Opus 4.8 on the agentic benchmarks — and costs about 89.3× less per token. Every number on this page links to the source it came from.
Where it stands
Competitor scores as published on the DeepSeek-V4-Flash-0731 model card, evaluated under the same harness. Agentic rows first, repo-scale rows last — because that is the order in which the gap widens, and pretending otherwise would not survive your first eval.
| Benchmark | DeepSeek-V4-Flash | Claude Opus 4.8 | GLM-5.2 |
|---|---|---|---|
| Terminal-Bench 2.1 | 82.7 | 85 | 81 |
| NL2Repo | 54.2 | 69.7 | 48.9 |
| Cybergym | 76.7 | 83.1 | — |
| DeepSWE | 54.4 | 58 | 46.2 |
| Toolathlon-Verified | 70.3 | 76.2 | 59.9 |
| Agents' Last Exam | 25.2 | 25.7 | 23.8 |
| AutomationBench (public) | 25.1 | 27.2 | 12.9 |
| DSBench-FullStack | 68.7 | 71.6 | 61.8 |
| DSBench-Hard | 59.6 | 71.7 | 54.5 |
Source: DeepSeek — DeepSeek-V4-Flash-0731 model card. Dashes mark figures the card does not report.
MiniMax-M2.7, the depth option
MiniMax publishes its scores without a full side-by-side table, so these are its own numbers and its own stated comparisons — nothing inferred.
| Benchmark | Score | As published |
|---|---|---|
| SWE-Bench Pro | 56.2% | matches GPT-5.3-Codex |
| Terminal-Bench 2 | 57.0% | — |
| SWE Multilingual | 76.5 | — |
| VIBE-Pro | 55.6% | near Claude Opus 4.6 |
| MLE-Bench Lite | 66.6% | medal rate; behind Opus 4.6 and GPT-5.4 |
| Toolathon | 46.3% | — |
Source: MiniMax — MiniMax-M2.7 model card.
What a million tokens costs
Vendor list prices. The blended column is what a 4:1 input-to-output workload actually pays — the ratio our own traffic runs at. Comparing input prices alone flatters whichever model has the cheaper prompt tier, which is why we do not.
| Model | Input / 1M | Output / 1M | Blended (4:1) | vs DeepSeek-V4-Flash |
|---|---|---|---|---|
| DeepSeek-V4-Flashours | $0.084 | $0.168 | $0.101 | — |
| MiniMax-M2.7ours | $0.18 | $0.72 | $0.288 | — |
| Claude Haiku 4.5 | $1.00 | $5.00 | $1.80 | 17.9× |
| Claude Sonnet 5 | $2.00 | $10.00 | $3.60 | 35.7× |
| GPT-5.2 | $1.75 | $14.00 | $4.20 | 41.7× |
| GPT-5.4 | $2.50 | $15.00 | $5.00 | 49.6× |
| Claude Opus 5 | $5.00 | $25.00 | $9.00 | 89.3× |
| GPT-5.5 | $5.00 | $30.00 | $10.00 | 99.2× |
Sources: Anthropic — Claude pricing, OpenRouter — models & pricing. Our rates are the ones the gateway bills from — same numbers as /pricing. Prices verified October 13, 2026.
On subscriptions, and what we won't claim
Claude Pro and Max do not publish a token quota. Anthropic's own help centre says usage "is affected by several factors, including the length and complexity of your conversations, the features you use, which Claude model you're chatting with, and the effort level you've selected." That is a reasonable way to run a chat product — and it makes any "you get N× more tokens than Claude Pro" claim unfalsifiable. So we don't make one.
What we do instead: your plan holds a priced allowance, and your console shows exactly how much of the 5-hour, weekly and monthly windows you have spent, as a percentage. The per-token rates that convert allowance into tokens are published on /pricing. You can check our arithmetic; that is the point.
Sources: Anthropic — How do usage and length limits work?, Anthropic — Usage limit best practices.
Migrate in 60 seconds
# before — Anthropic SDK
client = Anthropic(api_key=ANTHROPIC_API_KEY)
client.messages.create(model="claude-sonnet-5", ...)
# after — OpenAI SDK against the router
client = OpenAI(
base_url="https://router.mingles.ai/v1",
api_key=ROUTER_API_KEY,
)
client.chat.completions.create(model="deepseek-ai/DeepSeek-V4-Flash-0731", ...)
Tool-calling, streaming and tool_choice behave the
same. Need more depth: swap the model id to
MiniMaxAI/MiniMax-M2.7.
FAQ
▸Is DeepSeek-V4-Flash actually as good as Claude for agent work?
On the like-for-like table DeepSeek publishes, it lands within a couple of points of Claude Opus 4.8 on the agentic rows: 82.7 vs 85.0 on Terminal-Bench 2.1, 25.2 vs 25.7 on Agents' Last Exam, 25.1 vs 27.2 on AutomationBench. It is further behind on the hardest repo-scale rows — 54.2 vs 69.7 on NL2Repo, 59.6 vs 71.7 on DSBench-Hard. Close enough that the price gap decides most agent workloads, not close enough that we would claim parity where a task spans a whole codebase.
▸How does this compare to a Claude Pro or Max subscription?
Different products, and Anthropic does not publish a fixed message or token quota for those plans — their own help centre says usage depends on conversation length, features, model and effort level. That makes a token-for-token comparison impossible to make honestly, so we do not make one. What we can compare is the API: our plans and theirs both bill usage, and the per-token rates below are public on both sides.
▸Can I keep using the Anthropic SDK?
The router speaks the OpenAI Chat Completions format. Point the OpenAI SDK at https://router.mingles.ai/v1 — most agent frameworks (Cursor, Cline, Continue, OpenClaw) let you swap the base URL and model id without touching anything else.
▸What do I give up?
Vision, long-document handling and Anthropic-specific features like computer use are not on offer here. If your workload depends on those, Claude is the right tool. If it is code generation, tool-calling and agent loops at volume, the economics here are hard to argue with.
Also worth reading: vs OpenAI API · vs OpenRouter