I paid Anthropic US$1,892.38 over 13 months. That was Claude Code Max at $200 a month, plus another $20-100 for OpenAI's Codex. Combined, somewhere between $200 and $300 every month just on AI tooling.
I cancelled Claude Max in April. I wrote about it — multiple times. Each post was an honest attempt to figure out what the right setup looked like as the tools kept changing under my feet.
Two months ago, I stopped dual-wielding entirely. I switched to a single model — DeepSeek V4 Pro with High Thinking — through a single coding agent. Here is what actually happened.
June: $57.01. July: $33.09.
My June DeepSeek bill: 4.1 billion tokens processed, $57.01 total.
July: 2.1 billion tokens, $33.09. The drop was partly because I finished a massive localization project in June and partly because I got better at prompt engineering — smaller prompts, fewer retries, same output quality.
Two months combined: $90.10. That is less than half of what I was paying for a single month of Claude Max. Less than a third if you count the Codex subscription on top.
The immediate question everyone asks when you cut costs by 75% is: "Did you slow down?" Here is the answer.
786 Commits. 4 Products. Zero Slowdown.
I track my work through git. It is not a perfect productivity metric — line counts are meaningless, and a single commit can represent anything from a typo fix to a database migration. But directionally, it tells you something real.
Between June 1 and July 31, 2026, across the four products I maintain:
-
DIALOGUE (podcast platform): 382 commits. Launched a 6-language World Cup podcast network with automated publishing to YouTube, Spotify, and Apple Podcasts. Built a 41-post SEO content calendar across 8 weeks. Ran a full pipeline evaluation framework with 40 episodes and an LLM judge scoring 7 dimensions per episode.
-
chandlernguyen.com (this site + Ask Sydney): 168 commits. Fully localized my 7-module professional course into Japanese — every video, slide deck, template, and companion guide. Built a 30-question evaluation framework for Ask Sydney with checkpointing, LLM-as-judge scoring, and direct Supabase retrieval. Restructured Sydney's 8,000-char system prompt from scratch. Redesigned the homepage editorial flow.
-
Prova (AI builder coach): 228 commits. Shipped unified curriculum sequences across three personas. Built an iPad-native experience with master-detail patterns and drag-and-drop. Added cross-device draft sync. Built brief-driven reset paths with carry-forward logic.
-
STRAŦUM (marketing intelligence): 8 commits. Quiet period — doc syncs, dependency updates, an Apple SSO fix. Sometimes not shipping is the right call.
I was not slowing down. If anything, I was accelerating — 786 commits in two months is my most productive stretch since I started building publicly in 2023.
The Model: DeepSeek V4 Pro with High Thinking
The model I am using is not DeepSeek's cheapest option. V4 Flash costs $0.14 per million input tokens (cache miss) and $0.28 per million output. I used the Pro version — $0.435 per million input, $0.87 per million output — with High Thinking mode enabled. Not the cheapest config. But when the total bill is $33 in a month, I am too lazy to switch.
I will start using Flash more from now on. But that tells you something about the psychology of cheap tools: when the cost is low enough, you stop optimizing. You just build.
The Cache Multiplier Nobody Talks About
Here is the part that matters most, and the part that makes the headline numbers above possible.
OpenCode, the coding agent I use, maintains a DeepSeek prompt cache. My cache hit rate runs at 90-95%+. That transforms the economics completely:
| Token type | Price per 1M tokens |
|---|---|
| Input (cache miss) | $0.435 |
| Input (cache hit) | $0.003625 |
| Output | $0.87 |
Cache-hit input tokens cost 120 times less than cache-miss tokens. At 90%+ cache hit, your effective input cost is a rounding error. Most of what you pay is output — the actual generated code, the actual responses. That is the architecture that makes $33/month possible with heavy usage.
Without the cache, 4.1 billion tokens of input at $0.435/M would cost around $1,780 — just for June. With 95% cache hit, the effective input cost drops to roughly $7. The model is cheap. The cache makes it free.
What Quality Looks Like at This Price
The honest question: does DeepSeek V4 Pro produce the same quality as what I was getting from the $200/month Claude Max + Codex combo?
The answer is: it depends on what you measure, but for my workloads, the gap is small and closing.
On coding tasks: OpenCode + DeepSeek handles the same range of work I was doing with Claude Code and Codex — TypeScript, React, Next.js, FastAPI, Python, Swift, database migrations, YAML configs, shell scripts. The token-efficient difference I noticed with Codex (roughly 3x fewer tokens than Claude for equivalent work, as I mentioned in my March posts) holds here too in a different way — DeepSeek is more concise. You get less commentary, more code.
On writing tasks: I use the same DeepSeek model for writing — blog posts, course content, system prompts — directly in OpenCode. The editing cycle is different from what I was used to with Claude. Claude's default style is more polished; DeepSeek is more direct. Neither is better. You adjust your prompts.
On the Sydney eval: After switching Ask Sydney to DeepSeek, I ran a 30-question evaluation with an LLM judge. The results: retrieval quality was unchanged, citation accuracy averaged 4.2/5, and voice fidelity — how well Sydney sounds like me — scored 4.4/5. What did need work was source grounding on finance and healthcare topics, but that was a prompt engineering fix, not a model quality issue. I tightened the system prompt — added a rule requiring at least one blog post citation per major claim on sensitive topics — and redeployed.
The pattern I have seen across both DIALOGUE and Ask Sydney is the same one: at this price point, prompt engineering carries more weight. When you are paying $200 a month to Claude, you can afford the model to figure things out. When you are paying $33, you need to be more specific. The good news: the model follows instructions.
One Tool, One Model. The Mental Simplification Nobody Talks About.
The cost savings are what people notice. The shift that matters more: when you stop managing multiple subscriptions and tool chains, you reclaim mental bandwidth.
My April setup was Claude Code for architecture and frontend work, Codex for implementation, and Claude's API for DIALOGUE's podcast pipeline. Three different authentication flows. Three different rate limit behaviors. Three different subscription pages to check when something stopped working.
Now it is one model, one agent, one mental model for how the system behaves. I open OpenCode. I type. It works.
This is not a review of OpenCode versus Claude Code versus Cursor. They are different tools built for different philosophies. Claude Code's agent teams and project-level context are still best-in-class for large-scale architectural work. Codex's integration with OpenAI's ecosystem is still the smoothest path for GPT-native workflows. I am not saying DeepSeek is universally better — I am saying that for the specific combination of coding and writing I do, across four products, it is good enough at a price that makes me stop thinking about cost entirely.
That is the threshold that matters. Not "is it better." It is: "is it good enough that the price disappears as a variable."
The Pricing Reality — and What Happens Next
DeepSeek's current pricing includes this note:
We plan to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected.
Prices will go up. They always do. The $33 month will not stay $33 forever. But the pattern — cheap enough that you stop optimizing, cache hit rates that turn input costs into rounding errors, prompt engineering that closes the quality gap — is durable.
Two months in, $90.10 total, 786 commits. I am not going back to $300/month tool chains unless something forces me to. And I will start using Flash more — because at this point, why not?
Still coding, still learning, still shipping (for $45 a month).
Cheers, Chandler :P