DeepSeek V4-Flash instead of Opus: where it holds

Three weeks on the preview inside Claude Code, and this morning the 0731 weights under MIT

The invoice, not the benchmark

One agent session, the kind where it reads half a repository and then rewrites twenty files, burns roughly two million input tokens and three hundred thousand output. On Claude Opus 4.7, at $5 per million in and $25 out, that is $17.50. For one session. Run three a day and the model costs more than the infrastructure serving what it wrote.

When the agent is deciding something hard, I do not mind the number. I mind it when I am handing it mechanical work, which is most of the time.

What shipped on July 31

This morning DeepSeek published the V4-Flash-0731 weights on Hugging Face, MIT licensed, 167 GB. I have been on the preview for three weeks. This checkpoint only reruns the post-training: same architecture, 284 billion total parameters, 13 billion active per token, hybrid CSA plus HCA attention, a 1M-token input window and 384K out.

The line that matters is the price list.

35x cheaper on input, 89x on output. And $0.0028 is not a discount, it is a different order of magnitude: 1785x under Opus. For an agent that resends the system prompt and the conversation on every tool call, that column is most of the bill.

On benchmarks DeepSeek reports 82.7 on Terminal Bench 2.1 and 54.4 on DeepSWE, above the V4-Pro preview on all nine agentic benchmarks it publishes. I am not putting that next to Opus 4.7's 69.4 on Terminal-Bench 2.0. Different version of the benchmark, so the comparison does not hold. What holds is what I watched it do.

Two environment variables, and Claude Code talks to Ollama Cloud

What got me to try it properly is that I did not have to open a new account. I already pay for Ollama Cloud, and the model is sitting in there.

ollama signin
ollama run deepseek-v4-flash:cloud

That works straight away, but a bare chat was not the point. I wanted the agent, with its tools and its permissions. Claude Code speaks the Anthropic Messages format. Ollama exposes an OpenAI-compatible endpoint. So something has to translate in between. I used LiteLLM, which serves /v1/messages and forwards to whatever provider you point it at.

model_list:
  - model_name: flash
    litellm_params:
      model: openai/deepseek-v4-flash:cloud
      api_base: https://ollama.com/v1
      api_key: os.environ/OLLAMA_API_KEY

litellm --config config.yaml --port 4000
 
export ANTHROPIC_BASE_URL=http://localhost:4000
export ANTHROPIC_AUTH_TOKEN=anything
export ANTHROPIC_MODEL=flash
claude

Four lines, and the CLI I open every day is talking to a different model with the same prompt and the same tools. It is the only honest way to compare them. Swap the harness too and you are no longer measuring the model.

One note on billing: Ollama Cloud charges GPU time, not tokens, so inside the subscription you never see the numbers in that table. Three weeks of daily use and I did not hit the ceiling on the $20 plan.

Where it held: refactors and whole features

I gave it the work I would have given the other model anyway. A refactor crossing twelve files, the kind where each edit is trivial and the difficulty is remembering at file nine what you decided at file two. Then a feature end to end: provider, component, translations in both locales, tests.

What I did not expect is that it followed the thread. It opened files in a sensible order, did not invent functions that are not in the repository, and got to the end with something that compiled. On that kind of work, building rather than deciding, the gap to Opus 4.7 felt small. Small enough that I am not paying 89x to close it.

The habit before a big rework did not change either: I have it draw the map of the system first, then let it work. It handled that the way I expected.

Where it sent me back

Then there is the place it breaks, and it is always the same one: long tasks.

I saw the pattern three times. I set a constraint up front, something like do not touch the translation files until I say so. It holds for ten, twelve tool calls. Around the twentieth it touches them. The code is not wrong. The instruction is gone. In an agent that is worse than a syntax error, because a syntax error shows up immediately and a dropped constraint shows up in the diff, if you look.

How the work is split now

I did not replace anything. I moved a line.

• On Flash: reading repositories, first-pass refactors, tests over code that already exists, translations, and anything with an output I can check in thirty seconds.
• On Opus: the long session, the architectural call, debugging something I do not understand yet.

The rule is not how hard the task is. It is how many steps it has, and how expensive it is to notice late that it went wrong.

Next time you open a session for mechanical work, the kind where you already know what has to come out and you just need someone to type it, check what it cost you. If you already pay for Ollama Cloud, trying the alternative is two commands.